https://burbagelaw.com/updates.html Developments in United States, European and Asian patent law affecting artificial intelligence and software inventions, each with what it changes for a company holding or seeking patents. en-us Mon, 19 Oct 2026 12:00:00 +0000 https://burbagelaw.com/og-card.pnghttps://burbagelaw.com/updates.html https://burbagelaw.com/data/2026-10-15-top-25-ai-patent-applicants.html https://burbagelaw.com/data/2026-10-15-top-25-ai-patent-applicants.html Thu, 15 Oct 2026 12:00:00 +0000 Data AI-related published U.S. applications by first-named applicant, the 13 weekly publication dates from July 23 to October 15, 2026, compared with the same weeks of 2025, and where SoftBank Group's applications came from. SoftBank Group Corp. was the first-named applicant on 566 AI-related U.S. patent applications published in the 13 weeks ending October 15, 2026, more than any other applicant. Alphabet was second with 347. In the same weeks of 2025, SoftBank Group published one U.S. application, and it was not AI-related. The jump is real, and it has a plain cause. Most of the 566 are U.S. versions of applications SoftBank Group filed in Japan in February and March 2025. SoftBank Group filed them in the United States a year later, and they published together about 18 months after the Japanese filings. Only four of the eight companies tracked in edition 1 made the top 25: Alphabet, Nvidia, Microsoft and Apple. The other 21 places went to companies such as Samsung Electronics, NEC, Dell, Qualcomm and IBM, along with banks and carmakers. The top 25 together account for 3,492 of the 12,030 AI-related publications in the window, or 29%. Five applicants are in this year’s top 25 that were not in last year’s: SoftBank Group, Oracle (42 to 88), Ericsson (19 to 84), Kia (36 to 65) and Sony Group Corporation (25 to 61). Oracle and Sony Group more than doubled. So did Beijing Zitiao Network Technology, a ByteDance company, which rose from 47 to 120 and into the top 10. NEC nearly doubled, from 113 to 222. None of the other large rises looks like SoftBank Group’s. For Oracle, Ericsson, Sony Group and Beijing Zitiao, no more than 40% of their AI-related publications in the window came in their three busiest weeks, compared with 94% for SoftBank Group. Ericsson’s rise starts from a small base of 19. Five of the top 25 published fewer than a year earlier: IBM (134 to 121), Bank of America (93 to 90), Canon (79 to 68), Bosch (80 to 67) and Capital One (73 to 66). Toyota, Intel, Fujitsu, Maplebear and Digital Global Systems dropped out of the top 25. They arrived in bursts: 364 on September 24, 107 on August 27 and 60 on September 17, or 531 of the 566 in three weeks. 518 carry the one-word title “System.” G06N classifications appear on 340 of them. The timing comes from the Japanese filings. Under the Paris Convention, an applicant that files first in Japan has 12 months to file in the United States and keep the Japanese filing date as its priority date (35 U.S.C. 119). Of the 566, 481 claim priority to a Japanese application, and 453 of those Japanese applications were filed in February or March 2025. SoftBank Group used nearly the whole year: the median gap between the Japanese and U.S. filings was 359 days, and 534 of the 566 were filed in the United States in February or March 2026. All but 17 were filed directly in the United States rather than through the international (PCT) route. Of the 84 with no foreign priority claim, 79 claim the benefit of a U.S. provisional application filed in March 2025. The USPTO publishes an application about 18 months after its earliest priority date (35 U.S.C. 122(b)). The median here was 18.2 months, so every one of the 454 applications with a February or March 2025 priority date published between August 27 and September 24, 2026. Xsensus LLP, a patent firm based in Vienna, Virginia, is the attorney of record on 506 of the 566: its address is the correspondence address and the address in the power of attorney. Other firms appear as the correspondence address on the rest, among them Morgan, Lewis & Bockius on 23, SOEI USA on 19 and Banner & Witcoff on 11. The U.S. filings are a small part of a larger program. In Japan, SoftBank Group was the first-named applicant on 10,400 published applications in 2025 and 22,848 so far in 2026, more than any other applicant, according to the IP Force patent rankings. A report on WIPO’s generative AI patent data ranks SoftBank Group first in the world in that field and says the company has run a generative AI idea competition open to all group employees since 2023, filing patent applications on the best ideas. The U.S. records fit that account: the 566 name 463 different inventors, and Masayoshi Son is named on 25. Without SoftBank Group, AI-related publications rose 31%, from 8,740 to 11,464. With it, they rose 38%, to 12,030. Watch lists built on last year’s leaders are out of date. If they work in your space, add SoftBank Group, Oracle, Ericsson and Beijing Zitiao to competitor monitoring, and search affiliates’ names as well as the parent’s. Title searches will not find most of SoftBank Group’s work, because 518 of these applications are titled “System.” Screen by classification and read the claims. Date these references correctly. A published application is prior art as of its effective filing date (35 U.S.C. 102(a)(2)), and for subject matter described in a foreign priority application that is the foreign filing date (35 U.S.C. 102(d)). For most of these, that is February or March 2025, not 2026. If you filed an AI application after early 2025, several hundred newly published references from one applicant may bear on it. Update the novelty search before the first Office action, and disclose anything material. Expect more. SoftBank Group publishes far more applications in Japan than it has brought to the United States, and any Japanese application it carries over within 12 months of filing will appear here about 18 months after the Japanese date. https://burbagelaw.com/data/2026-10-13-who-is-filing-ai-patent-applications.html https://burbagelaw.com/data/2026-10-13-who-is-filing-ai-patent-applications.html Tue, 13 Oct 2026 12:00:00 +0000 Data Published U.S. applications from eight of the largest AI companies, July 16 to October 8, 2026, compared with the same weeks of 2025. Alphabet published 350 AI-related patent applications in the 13 weekly publication dates ending October 8, 2026, up from 221 in the same weeks of 2025. Nvidia went from 128 to 226. Microsoft rose more slowly, from 193 to 213. OpenAI went from 10 to 24, from a small base. Publications classified in G06N, the classification most people mean by “AI patents,” rose from 3,868 to 6,017 over the same weeks, an increase of 56%. All published utility applications rose 28%, from 90,195 to 115,085, so G06N’s share of everything published rose from 4.3% to 5.2%. Across the eight companies, AI-related applications whose titles name an agent or an agentic system rose from 14 to 47. Alphabet went from 3 to 17, Microsoft from 2 to 16 and Nvidia from 1 to 10. The median gap between effective filing date and publication for these 47 was 18 months, so most were filed in early 2025. Many of these filings are not in G06N at all. They are classified under natural language processing (G06F 40), computer vision (G06V) or speech (G10L) instead. Across the seven companies with at least 20 AI-related publications, G06N missed 419 of 949, or 44%. The gap runs from 37% for Alphabet to 71% for Meta and 77% for Apple. A landscape, freedom-to-operate or competitor-monitoring search that relies on G06N alone will miss a large share of the relevant art, and for some companies most of it. Add G06F 40, G06V and G10L, and search titles and claims for model and agent terms. https://burbagelaw.com/entries/2026-10-19-board-closed-fiscal-year-2026-smallest-appeal-invent.html https://burbagelaw.com/entries/2026-10-19-board-closed-fiscal-year-2026-smallest-appeal-invent.html Mon, 19 Oct 2026 12:00:00 +0000 Costs The Board decided 4,829 ex parte appeals in fiscal year 2026 against 3,684 filed, and its inventory of pending appeals fell from 3,125 at the start of the year to 1,977 at the end, a drop of 1,148, or 36.7 percent; the Office's chart of year-end inventories since fiscal year 2012, which starts at 26,570, has no lower value. Across the Board, 2,886 appeals were affirmed, or 59.8 percent, 387 affirmed in part, or 8.0 percent, and 1,521 reversed, or 31.5 percent. Technology Center 2100, which examines computer architecture, software and artificial intelligence subject matter, decided 526 appeals and reversed 202, a reversal rate of 38.4 percent, and affirmed in part 36 more, so 45.2 percent of its appellants had at least one rejection reversed. Technology Center 2400 reversed 36.0 percent and Technology Center 2600 reversed 35.9 percent, while Technology Center 3600, which holds the business method art units, decided 996 and reversed 224, or 22.5 percent, with 71.6 percent affirmed outright. Single months swing: in September 2026 alone Technology Center 2100 reversed 6 of 29. The source is the Office's fiscal-year-to-date table; each row was checked so that affirmed, affirmed in part, reversed, remanded and dismissed sum to the appeals decided, every row passed, and the percentages are those counts divided by that row's decided total. The same document reports that appeals decided in July through September 2026 took longer from receipt to decision than those decided a year earlier in nearly every technology center, which is consistent with a Board working through its oldest appeals. The Office does not report outcomes by type of rejection, so there is still no official reversal rate for eligibility rejections. What it changes: The queue an appeal enters today is a third shorter than the one it entered a year ago, and the Board is deciding 131 appeals for every 100 filed; the pendency figures in the same document run the other way, which is what deciding old appeals does to the measure, so the right reading is that an appeal filed this quarter joins a shorter line, not that the Board has become fast. For an artificial intelligence application in Technology Center 2100 carrying a rejection the examiner would not drop after an interview, the numbers favor the appeal over a second request for continued examination: a 45 percent chance that at least one rejection is reversed, against a continued examination fee that buys another round with the same examiner. In Technology Center 3600 the same rejection is a 28 percent chance, which is why the classification an application receives at filing is worth arguing about, and why a claim drafted to the computing improvement rather than to the transaction it serves is a budget decision as much as an eligibility one. Anyone quoting a reversal rate for eligibility rejections in particular is quoting a private analysis of Office data, and that caveat belongs next to the number when it is used to justify an appeal budget. United States Patent and Trademark Office, Patent Trial and Appeal Board, Appeal Statistics as of September 30, 2026, ex parte appeal receipts and dispositions, fiscal year 2026 (October 1, 2025 through September 30, 2026); percentages computed by Burbage Law, PC from the Office's table https://burbagelaw.com/entries/2026-10-19-invention-record-survives-priority-fight-native-file.html https://burbagelaw.com/entries/2026-10-19-invention-record-survives-priority-fight-native-file.html Mon, 19 Oct 2026 12:00:00 +0000 In-House Google petitioned for inter partes review of United States Patent No. 6,728,704, which claims merging the result lists returned by several search engines, on grounds built around a reference called Bushee that was filed five weeks before the patent's August 27, 2001 application. Valtrus antedated Bushee under the pre-America Invents Act version of Section 102(e) of Title 35 of the United States Code with a declaration from co-inventor Tsaparas, the original invention disclosure documents, and the Microsoft Word metadata of those documents. Google argued on appeal that documents and metadata under the inventor's own control cannot corroborate his testimony. The Federal Circuit disagreed: assuming without deciding that the metadata was authentic, metadata that Word generates automatically was not authored by the inventor, so it is independent corroboration and the Board's finding was supported by substantial evidence, following Global Health Solutions LLC v. Selner, 148 F.4th 1363 (Fed. Cir. 2025), and unlike the unwitnessed laboratory notebooks and unwitnessed emails that cannot corroborate on their own; Google's counsel acknowledged at argument that a human is not a required element of corroboration. The court also upheld the finding of reasonably continuous diligence over the roughly eleven business days between August 8 and August 27, 2001, which were spent getting a signed declaration back by mail from an inventor in Ontario and finishing the application at the law firm. What it changes: Antedating a reference ended for applications filed on or after March 16, 2013, but the question this case decides, what proves when an inventor conceived, still decides derivation proceedings, inventorship disputes and the question of who contributed what when a machine learning tool took part in the work, so an in-house program should keep its invention records in the form that proved itself here. Keep the disclosure in the system that created it, with the authorship and timestamp metadata that system writes without anyone asking, and treat the step where someone exports the document to a portable format or retypes it into an intake form as the step that turns independent evidence back into the inventor's own say-so. The same rule covers the prompts, model outputs and session logs of an engineer who used a code assistant or a model to get to the invention: the system's own timestamps are the metadata, and they are the first thing an opponent will ask for when inventorship is contested. The diligence holding is the smaller lesson, and it cuts the other way: Valtrus won because it could account for the entire period with timestamped evidence, so the program should docket the signature chase and the drafting hand-offs the way it dockets a statutory deadline, because the record of a routine three weeks is what the Board was asked to read. Practice note, Burbage Law, PC, October 19, 2026, on Google LLC v. Valtrus Innovations Ltd., No. 2024-1945 (Fed. Cir. Oct. 19, 2026) (precedential) (Cunningham, J., joined by Hughes, J., and Burroughs, D.J., sitting by designation), affirming the Patent Trial and Appeal Board in IPR2022-01545 https://burbagelaw.com/entries/2026-10-15-openai-claims-control-prompt-box-sets-how-many-token.html https://burbagelaw.com/entries/2026-10-15-openai-claims-control-prompt-box-sets-how-many-token.html Thu, 15 Oct 2026 12:00:00 +0000 Filings Claim 1 is a method in which the front end to a generative response engine shows, in the prompt composer, an indicator that opens a compute allocation interface with a selector object manipulated by a pointer device, offering a default setting and at least one greater or lesser setting; the setting controls the relative amount of tokens the engine may generate for the response, the engine receives the prompt and the setting together, and the greater setting makes it use more tokens than the default while the lesser setting makes it use fewer. Dependent claims add a trained model that predicts how long the prompt will take at the default setting and shows that estimate in the interface (claims 2 and 3), the same model choosing which version of the engine to use and preferring the more efficient one until the prompt calls for a more capable one (claims 4 through 6), keeping the setting for the rest of the conversation thread (claim 12), sending the setting to a data center load balancer (claim 15), and, in claims 23, 24 and 26, defining the setting as a maximum token budget, as an amount of time the engine may spend reasoning, and as counting the tokens generated during reasoning. The published set is an amended one: twenty-two claims numbered 1 through 27, with claims 9 through 11, 20 and 21 absent and claim 25 marked canceled. The Office's Open Data Portal file history shows a Track One request granted May 23, 2025, non-final rejections mailed July 9, 2025 and February 6, 2026, a final rejection mailed October 23, 2025, three applicant-initiated examiner interviews, a request for continued examination on January 15, 2026, a notice of allowance mailed June 5, 2026, a second request for continued examination filed July 9, 2026 with an information disclosure statement, a second notice of allowance mailed August 25, 2026, and an amendment after allowance filed October 1, 2026; as of October 11, 2026 the issue fee had not been paid. The application published at eighteen months from the provisional, on the first publication Thursday after the October 10, 2026 date, after it had already been allowed twice. What it changes: The prior art that matters is not OpenAI's own: o3-mini shipped on January 31, 2025 with low, medium and high reasoning effort options in the application programming interface and a higher-effort model in the ChatGPT picker, which the one-year grace period covers if the disclosure came from these inventors, but Anthropic announced Claude 3.7 Sonnet on February 24, 2025 with a developer-set thinking budget of up to N tokens and a user's choice in the applications between answering normally and thinking longer, forty-five days before the provisional, and the exception for a third party's disclosure that follows the inventor's own reaches only the subject matter the inventors had themselves disclosed first. That is why the claim lives in the composer: the selector object moved by a pointer device, the greater and lesser settings around a default, and, in claims 2 and 7, the duration estimate shown beside them, which is also why a developer who sets a token budget through an application programming interface is outside claim 1 as published and a consumer product with a thinking-effort control in its own prompt box is not. Compare the shipped interface to the published claim now rather than at grant, because the allowance has already happened and the next event is a patent number. The file is the second lesson: Track One and three interviews bought a first allowance in thirteen months, and the request for continued examination filed after that allowance, with an information disclosure statement, is the price of art that surfaced late; a program that finishes its own search before paying for Track One does not buy that second cycle. United States Patent Application Publication No. 2026/0319862 A1, OpenAI OpCo, LLC, "Adaptive Compute Allocation Display for Generative Response" (the title as printed on the publication; the Office's application record reads "Adaptive Compute Allocation Display for Generative Response Engines"), Application No. 19/194,991, filed April 30, 2025, claiming priority to Provisional Application No. 63/786,668 filed April 10, 2025, published October 15, 2026 https://burbagelaw.com/entries/2026-10-15-google-claims-procedure-sizing-model-s-depth-width-c.html https://burbagelaw.com/entries/2026-10-15-google-claims-procedure-sizing-model-s-depth-width-c.html Thu, 15 Oct 2026 12:00:00 +0000 Filings Claim 1 is a method for choosing a neural network architecture for a given compute constraint in three steps. For each architecture dimension to be adjusted, it fits a power law relating a performance metric to that dimension and to compute, from evaluations of networks identical except in that dimension, and takes a scaling exponent from the fit. It then picks a base architecture from candidate networks of different shapes, chosen by grid search in claim 13 and trained under a smaller compute budget, taking the best of them on performance against compute in claim 5. Finally it modifies the base architecture's dimensions using the exponents, the base architecture's compute use and the target constraint, in claim 6 by raising the ratio of target to base compute to a power set by the exponent. The dependent claims name depth and width (claim 17), the width of the multi-layer perceptron in a Transformer block (claim 18), image processing as the task (claims 21 and 23), and training a network of the output shape (claim 22); claims 25 and 26 repeat the method as a system and a storage medium. The four inventors posted "Getting ViT in Shape: Scaling Laws for Compute-Optimal Model Design" to arXiv on May 22, 2023, five days after the provisional, and reported the SoViT-400m/14 model that the method produced. The file history shows a request for immediate examination under Section 371(f) and docketing to an examiner on July 16, 2026, with no office action yet. What it changes: Five days is the whole margin: the United States grace period would have forgiven the paper, but the European Patent Office and China's office have no grace period for an inventor's own paper, so without the May 17 provisional the paper would have been prior art against the counterparts in the two largest markets outside the United States, and a research group that files the provisional on the way to the arXiv submission button, as this one did, keeps those rights for the cost of a form. On the United States claim as published I would expect a subject matter eligibility rejection: claim 1 fits curves, takes exponents and ends in an architecture, which is a set of numbers describing a model no one has yet trained, and the eligibility decisions that have favored machine learning claims, Ex parte Desjardins among them, involved claims that train or run a model, not a calculation whose output is a design. The repair is in the application already, which is to move claim 22's training step, or claim 23's processing of image data with the resulting network, into claim 1. Detection is the other problem with a design-time method: it is practiced inside a lab, and the only evidence that a competitor used it is the competitor's own paper, so a claim like this earns its budget as the thing that stops a published recipe from being free for everyone, and the family should add a claim to the model shape that the recipe produced. United States Patent Application Publication No. 2026/0320392 A1, Google LLC, "Scaling Neural Network Architectures to Computational Resource Constraints," Application No. 19/485,257, the national stage, entered November 17, 2025, of PCT/US2023/029718 filed August 8, 2023, claiming priority to Provisional Application No. 63/502,790 filed May 17, 2023, published October 15, 2026 https://burbagelaw.com/entries/2026-10-08-openai-claims-loading-narrative-memory-user-written.html https://burbagelaw.com/entries/2026-10-08-openai-claims-loading-narrative-memory-user-written.html Thu, 08 Oct 2026 12:00:00 +0000 Filings Claim 1 is a method in which a system server receives a prompt starting a conversation thread with a user account, inserts a personalized memory state into the context window before inference begins, and sends the prompt and the memory state together to a language model generative response engine. The memory state must contain two things: curated facts in natural language about the user account, and synthesized memories, which the claim defines as natural-language narrative descriptions of the user account generated by a language model that periodically analyzes past conversation threads in an asynchronous offline process before the prompt arrives, each summarizing conceptual connections across multiple threads and stored in a persistent user profile. The dependent claims add reasoning over which parts of the memory are relevant (claim 2), searching an archive of past interactions (claim 3), a short-term memory of recent threads (claim 4), a specialized model instance trained to decide what to synthesize (claim 6), demoting a stale summary to the archive (claim 7), keeping separate memory sets for two distinct usage profiles of one account (claim 8), trimming or summarizing the memory to free room in the context window (claim 10), and a personalization notepad the model writes to (claim 11). The published claim set shows claim 21 canceled and claims 22 and 23 added, so at least one amendment preceded publication; claim 1 as published also recites an "interference operation" where an inference operation is plainly meant. The provisional application was filed April 7, 2025, and OpenAI announced on April 10, 2025 that ChatGPT memory would draw on chat history as well as saved memories. What it changes: The February 2024 version of ChatGPT memory, which stored facts the user asked it to keep, is more than one year older than the April 7, 2025 provisional, so it is OpenAI's own prior art under Section 102(a)(1) of Title 35 of the United States Code with no grace period, and Nvidia's Publication No. 2026/0288790 A1 on this page carries a 2022 priority date for storing what the user said and applying it later. That is why the claim puts its weight on the synthesized memories: narrative descriptions written offline by a model, summarizing connections across threads, and loaded before the first prompt. Anyone building assistant memory should compare their design to those limitations, because a product that stores only extracted facts, or that writes its summaries at inference time rather than in a background job, is outside claim 1 as published. The three-day gap between the provisional and the launch is the cadence an in-house program should copy: the provisional goes in when the launch date is set, not after the press coverage, because the launch starts the one-year clock in the United States and ends novelty everywhere else. United States Patent Application Publication No. 2026/0310729 A1, OpenAI OpCo, LLC, "Memory Platform for Generative Response Engines," Application No. 19/195,022, filed April 30, 2025, claiming priority to Provisional Application No. 63/784,739 filed April 7, 2025, published October 8, 2026 https://burbagelaw.com/entries/2026-10-08-google-claims-attention-layer-compresses-whole-conte.html https://burbagelaw.com/entries/2026-10-08-google-claims-attention-layer-compresses-whole-conte.html Thu, 08 Oct 2026 12:00:00 +0000 Filings Claim 1 is a method of generating an output sequence token by token with a generative neural network that contains a compressive attention layer. For that layer, at each output position, the method identifies the layer's memory state, derives a query vector, a key vector and a value vector from the layer input token, performs one or more optimization steps on a compression objective using the key and value vectors to update the memory state, and then applies the updated memory state to the query vector to produce the layer output. The specification describes the memory as a fixed number of slots that replaces the key-value cache, whose size grows with every token in a standard transformer and which the specification identifies as the limit on context length and batch size in high-bandwidth memory. The dependent claims narrow to a single gradient descent step per token (claims 2 and 3), a compression objective that measures the difference between the memory applied to the key and the value, by L2 norm or cosine similarity (claims 4 through 6), a column-normalized memory (claim 7), and projecting each memory column onto the unit sphere after the update (claim 11). The two named inventors, with Razvan Pascanu, posted the paper "Lattice: Learning to Efficiently Compress the Memory" to arXiv on April 8, 2025, one day after the provisional was filed. What it changes: Claim 1 is written at the level of the mechanism, a learned update of a fixed-size state under an explicit objective, rather than at the level of the result, a bounded cache, and that makes it both enforceable and exposed: it reads on any architecture that updates a memory matrix by optimizing a reconstruction loss at inference time, which now describes a family of published test-time-training and linear-attention designs, including Google's own Titans paper (arXiv 2501.00663) and the test-time-training layers of July 2024 (arXiv 2407.04620), so the examiner will have to sort out which of those disclose the specific key-to-value objective. For anyone shipping long-context or on-device models, the dependent claims are the ones to check against the implementation, since a single gradient step with an L2 objective on a column-normalized state is a design choice an engineer either made or did not. The one-day gap between the provisional and the arXiv posting is the standard the research groups at the large companies hold themselves to, and an in-house program that lets its researchers post first has given away every jurisdiction without a grace period. United States Patent Application Publication No. 2026/0310945 A1, Google LLC, "Generative Neural Networks With Compressive Attention Layers," Application No. 19/641,299, filed April 7, 2026, claiming priority to Provisional Application No. 63/784,960 filed April 7, 2025, published October 8, 2026 https://burbagelaw.com/entries/2026-10-08-google-claims-training-agent-second-model-grades-eac.html https://burbagelaw.com/entries/2026-10-08-google-claims-training-agent-second-model-grades-eac.html Thu, 08 Oct 2026 12:00:00 +0000 Filings Claim 1 is a method of training a machine-learned agent model by reinforcement learning to improve tool use. The system obtains an output of the agent that contains an action feature, generated from a sequence of preceding states, specifying a function call to a tool external to the agent; a machine-learned reward model processes that output together with the preceding states to produce a quality indicator of whether the action advances the goal of the sequence; and the quality indicator goes to a model trainer running a reinforcement learning algorithm that updates the agent's parameters. The dependent claims let the reward model make its own tool call to fetch reference data before grading (claim 2), define the action as a token sequence with start-of-action and end-of-action tokens around a function identifier and parameters (claim 3), execute the call and append the action and its result to the state sequence for the next iteration until a finish token is generated (claims 4 through 6), name search, database lookup, calculation and internal reasoning as the tools (claim 7), and add a cost term that penalizes the number of tool calls (claims 8 and 9). The specification describes training the reward model from a larger supervisor model that has access to reference sources, so that the reward model can be small and run beside the agent. What it changes: A reward computed per tool call, rather than per final answer, is the step-level supervision that agent training converged on in 2024 and 2025, and this claim puts a July 7, 2023 priority date on it, which is earlier than most of that work. ReAct (arXiv 2210.03629) and Toolformer (arXiv 2302.04761) were public before July 2023, but neither trained a separate learned reward model to grade each call against the goal, so the obviousness question is whether putting a reward model on the intermediate action was a predictable combination, and anyone training agents with step-level rewards on tool use should get an opinion on that question rather than assume the answer. The continuation filed in June 2026 means Google is keeping the family open and can draft further claims to whatever the agent market ships next. The claim 9 penalty on the number of tool calls is the kind of narrow, engineering-specific limitation that tends to survive both eligibility and prior art, and a defensive filing in this area should claim its own reward design at that level of detail. United States Patent Application Publication No. 2026/0311065 A1, Google LLC, "Reinforcement Learning With Information Retrieval Feedback," Application No. 19/711,492, filed June 17, 2026 as a continuation of Application No. 18/348,687 filed July 7, 2023, published October 8, 2026 https://burbagelaw.com/entries/2026-10-06-patent-budget-defended-abandonment-review-not-filing.html https://burbagelaw.com/entries/2026-10-06-patent-budget-defended-abandonment-review-not-filing.html Tue, 06 Oct 2026 12:00:00 +0000 In-House Most in-house programs have a filing committee and almost none have an abandonment committee, which is backwards, because the filing decision commits a small fraction of what a patent family will eventually cost. The rest is committed later, one decision at a time: the foreign filing decision at twelve months, the national phase decision at thirty months, the decision to answer each office action rather than let the application go, the issue fee and the continuation that goes with it, and the maintenance fees and foreign annuities that run for the life of the patent. In most programs those later decisions are made by default, by a docketing system that generates a reminder and a paralegal who pays the fee, so the budget fills with payments nobody chose to make on applications nobody has reread since they were filed. The programs that hold their budgets treat each of those dates as a decision with an owner, and the question asked at each one is the same: do the claims as they now stand read on something we ship, something a competitor ships, or something still on the roadmap. In two decades of prosecuting for companies from startups to Google and Toyota, the single change that did the most for a budget was reversing the default, so that at each date the application is abandoned unless someone states on the record why it should be paid for. What it changes: For an artificial intelligence program the review matters more than elsewhere because the technology turns over faster than the patent process runs: an application drafted to a 2022 training method reaches its first office action after the team has moved on, and its foreign filing deadline arrives before anyone knows whether the feature survived. Put the three expensive dates, foreign filing, national phase and the response to the first office action, on the roadmap owner's calendar rather than the attorney's alone, and require a one-paragraph written reason to proceed, naming the product or competitor the claims read on. Keep continuations only in families still mapped to the roadmap, and let the rest issue or lapse. When the budget is challenged, the list of what the program chose not to pay for, with the reasons, is a better defense than the count of what it filed, because it shows the spending is being managed rather than accumulated. Practice note, Burbage Law, PC, October 6, 2026 https://burbagelaw.com/entries/2026-10-01-system-prompt-trade-secret-only-long-model-will-not.html https://burbagelaw.com/entries/2026-10-01-system-prompt-trade-secret-only-long-model-will-not.html Thu, 01 Oct 2026 12:00:00 +0000 Trade Secrets System prompts, routing prompts and evaluation rubrics are the most copied assets in an artificial intelligence product and the hardest to protect. Patent protection is rarely available for the text of a prompt, because the words are not a process, and a claim to prompting a model with instructions to achieve a result is the result-oriented claim the Federal Circuit has held abstract since Recentive Analytics, Inc. v. Fox Corp. Copyright protects the exact wording against literal copying and nothing about the technique. That leaves trade secret law, which has one condition that matters here: the owner must have taken reasonable measures to keep the information secret. A system prompt shipped to a model that can be talked into repeating it, or sent to an outside inference provider under terms that permit logging or training, stops meeting that condition the first time someone extracts it, and extraction attempts against any public product are routine. What it changes: Run the prompt library as a trade secret program rather than a code folder: an access-controlled repository, prompt text kept out of client-side code and out of logs, confidentiality and no-training terms in every inference vendor contract, and a documented defense against extraction, because the first question in any misappropriation case is what measures the owner actually took. Where the value sits in the machinery around the prompt rather than in its words, how prompts are selected, assembled, cached or routed, that machinery can be patented, and the application should claim the mechanism and leave the prompt text in the specification as an example at most. A team that cannot say which of its prompts are secrets and which are examples has not made the decision yet, and the decision gets more expensive after the product ships. Practice note, Burbage Law, PC, October 1, 2026 https://burbagelaw.com/entries/2026-10-01-claim-covers-something-cannot-work-definite-invalid.html https://burbagelaw.com/entries/2026-10-01-claim-covers-something-cannot-work-definite-invalid.html Thu, 01 Oct 2026 12:00:00 +0000 Fed. Cir. United States Patent No. 6,711,385 claims a communications apparatus "for transmitting electric or electromagnetic signals over air," with a transmitter and a coupler whose non-magnetic core transformer matches the transmitter's output impedance to the characteristic impedance of the air. Both sides agreed that an electric signal cannot be transmitted over air, and the district court held the claims indefinite under Section 112(b) of Title 35 of the United States Code on that basis. The Federal Circuit disagreed on indefiniteness, holding that indefiniteness is not the same as impossibility and that claims which encompass inoperable embodiments can still be clear about what they cover, but affirmed the judgment of invalidity on a ground the district court had declined to reach: the specification must enable the full scope of the claims, and a claim that explicitly covers something a skilled artisan cannot make or use is not enabled, citing Amgen Inc. v. Sanofi and EMI Group North America, Inc. v. Cypress Semiconductor Corp. The court refused Satius's proposed construction, under which the electric signal would be converted to an electromagnetic one before reaching the air, because that step is not in the claim and a court may not redraft claims to make them operable. The following day, in Pioneer Hi-Bred, a different panel affirmed a post-grant review decision holding claims to a genus of dual-function herbicide-degrading enzymes not enabled by two working examples, and held that experimental data generated after the priority date may be used to show that the genus did not behave as the specification predicted. What it changes: The exposure in machine learning claims is the list of alternatives: a claim that recites a model that is "a neural network, a decision tree, a support vector machine or a rule-based classifier," or training "on labeled or unlabeled data," covers every alternative, and under this decision every alternative has to work for the claimed purpose as described, because the court will not read the failing one out of the claim later. Review each alternative in the independent claims against the specification and ask whether it is shown to work; move the ones included only for breadth into dependent claims or leave them in the specification. Pioneer Hi-Bred is the second edge of the same rule for claims that recite a model trained to produce a result: two worked examples do not enable the genus, and the opponent may run the untested configurations after the fact and put the failures in evidence, so an application on a training method should report results across the range of architectures, data regimes and tasks the claims cover, and the claims should stop where the results do. Satius Holding, LLC v. Samsung Electronics Co., Ltd., No. 2025-1446 (Fed. Cir. Oct. 1, 2026) (precedential) (Moore, C.J., joined by Lourie and Hughes, JJ.), affirming on other grounds a judgment of the United States District Court for the District of Delaware; see also Pioneer Hi-Bred International, Inc. v. Inari Agriculture, Inc., No. 2025-1287 (Fed. Cir. Oct. 2, 2026) (nonprecedential) https://burbagelaw.com/entries/2026-09-30-court-refuses-fees-after-eligibility-loss-must-expla.html https://burbagelaw.com/entries/2026-09-30-court-refuses-fees-after-eligibility-loss-must-expla.html Wed, 30 Sep 2026 12:00:00 +0000 Fed. Cir. Epic Tech, LLC owned United States Patent No. 8,545,317, which claims an electronic sweepstakes system connecting gaming terminals over a server network. Before it sued in 2020, the Supreme Court had decided Alice Corp. v. CLS Bank International, the United States Patent and Trademark Office had reopened prosecution of two related applications and rejected them under Section 101 of Title 35 of the United States Code, Epic Tech had abandoned three related applications, and a district court in litigation over related United States Patent No. 8,545,315 had held similar claims ineligible. The United States District Court for the Northern District of Georgia held the asserted claims ineligible on summary judgment and then denied the defendant's motions for sanctions under Rule 11 and for attorney fees under Section 285. The Federal Circuit, in an opinion by Chief Judge Moore joined by Judge Cunningham and District Judge Subramanian sitting by designation, vacated the denial, holding that those facts taken together created a compelling concern over the validity of the claims before the litigation was brought, and that the district court needed to address why the claim nevertheless had a reasonable chance of success. The panel added that facts showing a reasonable pre-suit infringement investigation do not address the distinct question of whether the patentee should have investigated validity. What it changes: An eligibility rejection in a continuation or a sibling application now follows the issued parent into court, and so does any ruling on a related patent, so the prosecution history of the whole family is part of the fee exposure when a software or machine learning patent is asserted. Before suing on such a patent, the file should contain a pre-suit validity memorandum on eligibility, separate from the infringement analysis, that confronts every family-member rejection and every decision on related claims, and where that memorandum is weak the better course is to prosecute a continuation with claims that recite the technical mechanism and assert that instead. For a company on the receiving end of an assertion, this decision makes the fee motion after an eligibility win worth filing, and the first exhibit should be the examiner's rejections in the patentee's own related applications. Epic Tech, LLC v. Pen-Tech Associates, Inc., No. 2025-1624 (Fed. Cir. Sept. 30, 2026) (precedential), vacating and remanding Epic Tech, LLC v. Pen-Tech Associates, Inc., No. 1:20-cv-02428-VMC (N.D. Ga.) https://burbagelaw.com/entries/2026-09-28-invention-disclosures-come-out-roadmap-review-not-ou.html https://burbagelaw.com/entries/2026-09-28-invention-disclosures-come-out-roadmap-review-not-ou.html Mon, 28 Sep 2026 12:00:00 +0000 In-House Every in-house patent program has an invention disclosure form, and in most of them the form sits unused, because an engineer has to decide alone that something is an invention before filling it in. Engineers make that judgment badly in a predictable direction: they submit the clever workaround they are proud of and never mention the architectural decision that took the team three months, because to them the second one is simply the design. The programs that harvest well put the patent attorney in the room where the product roadmap is reviewed, have the attorney read the design documents and the largest proposed code changes as they come up for review, and then send the attorney to the engineer with a draft description of what the attorney believes the invention is, asking whether it is right. The form is filled in afterward, by the attorney, and the engineer's share of the work is a forty-minute conversation. Across two decades of drafting for companies from four-person startups to Google and Toyota, this is the difference between programs that filed on what mattered and programs that filed on what happened to arrive. What it changes: For a company building artificial intelligence products the roadmap review matters more than it does elsewhere, because the inventions sit in the training pipeline, the evaluation harness and the data handling, none of which an engineer thinks of as a product feature, and because the interval between a design decision and its appearance in a published model card or a conference paper can be a few weeks. Budget the attorney's time as a standing hour in the engineering planning cycle rather than as a response to disclosures, and measure the program by how many of the roadmap's major items have a filing decision recorded against them, counting a recorded decision not to file. A quarter in which every disclosure arrived through the form is a sign that harvesting has stopped, not a sign that the engineers have run out of inventions. Practice note, Burbage Law, PC, September 28, 2026 https://burbagelaw.com/entries/2026-09-24-waymo-claims-prompting-generative-model-sensor-data.html https://burbagelaw.com/entries/2026-09-24-waymo-claims-prompting-generative-model-sensor-data.html Thu, 24 Sep 2026 12:00:00 +0000 Filings Claim 1 recites an autonomous vehicle with a sensing system, a data processing system that generates a prompt for a generative artificial intelligence model, the prompt comprising spatial data characterizing the driving environment, being either the sensing data itself or the spatial arrangement of objects identified from it, together with a request to determine whether the vehicle is obstructing traffic, causes the model to process the prompt and generate a response indicating that the vehicle is obstructing traffic, and a control system that repositions the vehicle in response. Claim 14 places the model on a computing device outside the vehicle, and claim 15 is the method. The dependent claims add that the vehicle is stationary when the assessment is made, that the request is natural language text, that the model is a vision language model or a multimodal language model, that heuristic metrics screen a location before the model is asked, and that the model was trained on datasets carrying ground truth blocking determinations. The specification describes the model generically as a language model, a large language model, a vision language model or a multimodal language model, and no claim is limited to a particular architecture or training method. What it changes: This is the claim form for a language model used as a judgment component inside a physical system: build the prompt from sensor data, ask a yes-or-no question, act on the answer with a physical operation. The repositioning step is what separates it from the claims the Federal Circuit held abstract in Recentive Analytics, Inc. v. Fox Corp. and the decisions that followed it, where a generic model was applied to a new category of data and the output was more data, so a company putting a language model in the loop of a robot, a vehicle, a production line or a medical device should claim the action the answer triggers and not only the answer. Because it was published, the application is prior art under Section 102(a)(2) of Title 35 of the United States Code as of March 18, 2025 against anyone who filed later on using a vision language model to evaluate a stopping location, and it belongs in the file of any company running a similar model-in-the-loop design. It also shows an Alphabet company treating a general purpose language model as a component of the driving stack rather than a research tool, and filing on the integration rather than on the model. United States Patent Application Publication No. 2026/0285357 A1, Waymo LLC, "Prevention of Traffic Obstructions Using Artificial Intelligence Systems in Automotive Vehicle Applications," Application No. 19/082,835, filed March 18, 2025, published September 24, 2026 https://burbagelaw.com/entries/2026-09-24-nvidia-claims-assistant-stores-what-user-told-applie.html https://burbagelaw.com/entries/2026-09-24-nvidia-claims-assistant-stores-what-user-told-applie.html Thu, 24 Sep 2026 12:00:00 +0000 Filings Claim 1 recites a processor comprising circuits that use one or more natural language processing models to determine an intent expressed in a first query to a conversational artificial intelligence platform, the intent indicating a request to store personalization data from that query in a data store associated with the user; use the models to determine the relevance of the stored personalization data to a second query; and generate a response to the second query using the personalization data on the basis of that relevance. Claims 11 and 16 are the method and the system. The dependent claims add inserting the personalization data into the second query before it is processed, searching a structured data store first and an unstructured store only when the structured store lacks the data, distinguishing a personalization query from other query types, invoking domain-specific agents according to the classification of the query, keeping interaction logs as unstructured text documents, and confirming to the user that the data was stored. No claim is limited to a model architecture, and the specification describes the models generically. The verbatim claim text could not be retrieved from the publication; the elements are paraphrased as reported by Justia and by the trade site Patentlyze. What it changes: This is, in general terms, the memory feature that OpenAI added to ChatGPT in 2024 and that most assistant products now offer, and the application carries a February 17, 2022 priority date, which is before any of those products existed. As a published application it is prior art under Section 102(a)(2) of Title 35 of the United States Code as of that date against anyone who filed later on storing user preferences when the user expresses an intent to be remembered and retrieving them by relevance to a later query, so any company that filed on assistant memory after early 2022 should read it against its own claims. As paraphrased, claim 1 reads on a generic pipeline of storing and retrieving information, which is the kind of claim the Federal Circuit has held abstract since Recentive Analytics, Inc. v. Fox Corp., so expect the examiner to push Nvidia toward the structured-then-unstructured fallback of claim 4 and watch where the allowed claim lands before treating it as a threat. The filing also shows the continuation practice an in-house program should copy: Nvidia kept a 2022 family pending and in 2026 filed claims drafted to a feature the market had since shipped, which is only possible if at least one continuation is kept open in every family that describes a platform capability. United States Patent Application Publication No. 2026/0288790 A1, Nvidia Corporation, "Query Response Generation," Application No. 19/689,724, a continuation of Application No. 17/674,704 filed February 17, 2022, published September 24, 2026 https://burbagelaw.com/entries/2026-09-21-appeal-outcomes-software-technology-center-run-sixte.html https://burbagelaw.com/entries/2026-09-21-appeal-outcomes-software-technology-center-run-sixte.html Mon, 21 Sep 2026 12:00:00 +0000 Costs Across the whole Patent Trial and Appeal Board, 4,554 ex parte appeals were decided in the fiscal year through August 2026: 2,711 affirmed, or 59.5 percent, 357 affirmed in part, or 7.8 percent, and 1,454 reversed, or 31.9 percent. The technology centers diverge sharply. Technology Center 2100, which examines computer architecture, software and artificial intelligence subject matter, decided 497 appeals and reversed 196 of them, a reversal rate of 39.4 percent. Technology Center 3600, which holds the business method art units, decided 943 appeals and reversed 216, a reversal rate of 22.9 percent, while Technology Center 2600 decided 287 and reversed 103, at 35.9 percent. The Office does not break these outcomes down by type of rejection, so no official figure exists for reversal rates on subject matter eligibility rejections specifically. What it changes: Where an application is classified changes the expected value of an appeal by roughly sixteen percentage points, and that belongs in the decision about whether to appeal or file a continuation. An application in Technology Center 2100 carrying a rejection the examiner will not withdraw is worth appealing on these numbers; the same rejection in a business method art unit in Technology Center 3600 is a harder bet and usually argues for a claim amendment and a continuation instead. Anyone quoting a reversal rate for eligibility rejections in particular is quoting a private analysis of Office data rather than an Office statistic, and that distinction is worth making when the number is being used to justify a budget. United States Patent and Trademark Office, Monthly Appeal Statistics, August 2026, ex parte appeal receipts and dispositions, fiscal year to date, October 2025 through August 2026 https://burbagelaw.com/entries/2026-09-21-inventive-concept-claim-does-not-require-cannot-carr.html https://burbagelaw.com/entries/2026-09-21-inventive-concept-claim-does-not-require-cannot-carr.html Mon, 21 Sep 2026 12:00:00 +0000 Fed. Cir. United States Patent Nos. 7,596,521, 7,882,137 and 8,612,316 claim a computer-implemented process for determining object-level profitability with a relational database management system: preparing information, establishing processing rules in the database, using the database system to independently calculate at least one marginal value of profit for each object being measured, calculating a fully absorbed profit adjustment value for each object, and combining the two. The district court granted the patentee summary judgment of eligibility, agreeing that the claims were directed to the abstract idea of calculating profitability but finding an inventive concept in the independent calculation step, which it read as an unconventional combination that improves the functioning of the computer itself by performing parallel computations in less time. The Federal Circuit reversed in an opinion by Judge Reyna for a panel that included Judge Dyk and Chief District Judge Bissoon sitting by designation, holding at step one that performing calculations based on collecting and analyzing information is an abstract idea and that performing them faster on a computer does not change that. At step two the panel held that if a claim is so broad that it can be practiced without the alleged inventive concept, so that the concept is not required by the claim, the claim cannot be significantly more than the abstract idea. The parties agreed that parallel processing was not a requirement of the claim, and that agreement decided the appeal. What it changes: The technical improvement that wins an eligibility argument has to be a limitation of the claim rather than a property of the preferred embodiment. Machine learning applications routinely describe a faster training loop, a smaller model, a lower memory footprint or a parallelized inference path in the specification while claiming only the data flow, and this decision says the argument built on that improvement fails the moment the accused infringer shows the claim can be practiced without it. For anyone drafting now, the independent claim should recite the mechanism that produces the improvement in words specific enough that the claim cannot be practiced without it, and for anyone holding issued claims, an eligibility review of the portfolio should be run against what the claims require rather than against what the patents describe. The decision is nonprecedential, and it reversed a district court that had found the claims eligible on summary judgment. Berkeley*IEOR d/b/a B*IEOR v. W.W. Grainger Inc., Nos. 2024-2353, 2025-1009 (Fed. Cir. Sept. 21, 2026) (nonprecedential), reversing a grant of summary judgment of eligibility by the United States District Court for the Northern District of Illinois, No. 17 C 7472 https://burbagelaw.com/entries/2026-09-18-california-moves-toward-placing-independent-auditors.html https://burbagelaw.com/entries/2026-09-18-california-moves-toward-placing-independent-auditors.html Fri, 18 Sep 2026 12:00:00 +0000 US Policy Senate Bill 813 creates a framework under which independent verification organizations assess artificial intelligence systems and models for compliance with California law, and Assembly Bill 1405 creates a state registry of artificial intelligence auditors with standards for independence, transparency and integrity. Both were signed on September 9, 2026. Executive Order N-9-26 then directed the Government Operations Agency to accelerate the implementation timelines for both statutes and to convene national experts within two months to recommend strengthening state law. Among the recommendations the order names are requiring frontier artificial intelligence companies to embed a designated independent verification organization onsite in their laboratories to conduct regular audits, requiring that safety frameworks and risk assessments be verified by independent organizations, advancing an emergency shutdown mechanism for frontier models, and expanding the definition of a critical safety incident to cover loss-of-control scenarios. What it changes: Trade secret protection rests on reasonable measures to keep the information secret, and a statutory regime that seats an outside auditor inside the building with access to model internals changes what those measures have to look like and enlarges the set of people who have seen the method. Anything a company is currently holding as a trade secret rather than filing on, including training pipelines, evaluation harnesses and safety classifier architectures, should be re-examined now against the question of whether it will still be secret once an audit regime reaches it. Disclosure to an auditor under legal compulsion is not a printed publication and does not by itself start a statutory bar, but it multiplies the number of places from which the secret can escape, and the decision to file is far cheaper made before the auditor arrives than after. California Executive Order N-9-26 (September 18, 2026); Senate Bill 813 (McNerney) and Assembly Bill 1405 (Bauer-Kahan), signed September 9, 2026 https://burbagelaw.com/entries/2026-09-17-openai-s-prompt-caching-application-claims-hash-rout.html https://burbagelaw.com/entries/2026-09-17-openai-s-prompt-caching-application-claims-hash-rout.html Thu, 17 Sep 2026 12:00:00 +0000 Filings As published, claim 1 is a method in which a cloud computing service receives a request containing a prompt with a natural language task and an access key, generates a hash from a portion of the prompt, identifies the generative response engine that will handle the request on the basis of that hash, transmits the prompt to that engine, and receives a response reporting how many of the input tokens were already cached, from which a credit is determined. The specification describes computing the hash from a prefix of the prompt combined with a user identifier, so that repeated requests with the same opening are sent to the machine that already holds the computed attention state for those tokens, and describes a caching window with a minimum duration of 300 seconds. Other independent claims extend the scheme to multimodal inputs, to encoder tokens, and to a centralized cache within a data center. Both OpenAI OpCo, LLC and Anthropic, PBC announced prompt caching as a product feature before the earliest filing date in this family, Anthropic on August 14, 2024 and OpenAI on October 1, 2024, as reported by SiliconANGLE and in coverage of OpenAI's 2024 developer day. The trade site Patentlyze lists five OpenAI publications in September 2026, against a portfolio that stood at four documents when Justia Patents was checked on September 21, 2026. What it changes: What is claimed is not the idea of reusing a computed prefix, which was public from both companies months before the earliest filing date, but the routing mechanism around it: hashing the prefix to choose the server, and reporting cached token counts back for billing. That is a lesson in where to put the claim when the feature is already public, because the two 2024 announcements are prior art against the family under Section 102(a)(1) of Title 35 of the United States Code for everything they disclosed, and the hash-based routing is what they did not disclose. Any company running an inference service with prefix caching across more than one server should compare its request routing against these claims now rather than after grant, and companies buying inference from OpenAI should note that the billing credit for cached tokens is itself a claimed step. The wider signal is that OpenAI has started filing in volume on the serving infrastructure rather than on the models. United States Patent Application Publication No. 2026/0278026 A1, OpenAI OpCo, LLC, "Prompt Caching in Generative Response Engines," filed February 18, 2026, a continuation of Application No. 19/240,708, filed June 17, 2025, in a family that also identifies Application No. 19/077,920, filed March 12, 2025, published September 17, 2026 https://burbagelaw.com/entries/2026-09-14-private-sale-not-public-disclosure-so-does-not-switc.html https://burbagelaw.com/entries/2026-09-14-private-sale-not-public-disclosure-so-does-not-switc.html Mon, 14 Sep 2026 12:00:00 +0000 Fed. Cir. Section 102(b)(1)(B) of Title 35 of the United States Code removes a third party disclosure from the prior art if the inventor publicly disclosed the same subject matter first. The patentee sold its first commercial unit to a customer in July 2012 with no confidentiality restriction, and argued that the sale was a public disclosure that disqualified a competing third party sale in August 2012. The Federal Circuit disagreed, holding that a private sale does not constitute a public disclosure even where the goods are actually delivered to the buyer without any confidentiality restriction. The panel extended to the grace period the reasoning it had applied to the derivation exception in Sanho Corp. v. Kaijet Technology International. What it changes: A quiet commercial sale before filing buys an inventor nothing against a competitor who publishes or sells the same thing a month later. This matters most to companies that place a model, a tool or an instrumented system with design partners under a paid pilot before anything is on file, because those transactions feel like disclosure to the business and count as nothing to the statute. The only reliable protection against an intervening third party disclosure is a filed application, so the pilot agreement and the provisional application should be executed in the same week rather than in that order. NCS Multistage Inc. v. Nine Energy Service, Inc., No. 2025-1000 (Fed. Cir. Sept. 14, 2026) (precedential) https://burbagelaw.com/entries/2026-09-10-nvidia-seeks-two-element-coverage-step-tells-agent-w.html https://burbagelaw.com/entries/2026-09-10-nvidia-seeks-two-element-coverage-step-tells-agent-w.html Thu, 10 Sep 2026 12:00:00 +0000 Filings Claim 1 recites a system with processing circuitry to query a data collection with a zero-vector to retrieve a subset of data from the data collection, and to generate, using a machine learning model, a summary of the data collection based at least on the subset of data. In plain terms, the system probes a vector database with an all-zeros query embedding, which returns an arbitrary rather than a semantically targeted sample, and then has a model write a description of what the whole collection holds. That description is what an agent later reads when deciding which data source a task should be routed to. The claim has two elements and is not limited to large language models, to any particular agent framework, to retrieval augmented generation, or to text. It was one of six Nvidia applications published the same day, another of which, Publication No. 2026/0267835 A1, claims using agreement among multiple retriever models as a substitute for human labeling when selecting a filter for synthetic training data. What it changes: This is the automatic data source description step that sits in most agentic data stacks, claimed at a breadth that will read on a great deal of ordinary practice if it issues near its published scope. Any company building agent orchestration over its own data should put this publication in front of whoever runs freedom to operate, and should check whether its own equivalent step is described anywhere in its own filings. The wider signal is that the company selling the hardware underneath everyone's agents is filing on the orchestration software above it, while the two laboratories with the most valuable models hold almost nothing: as indexed by Justia Patents on September 21, 2026, the entire United States portfolio of Anthropic, PBC runs to fourteen documents and that of OpenAI OpCo LLC to four. United States Patent Application Publication No. 2026/0267897 A1, Nvidia Corporation, "Automated Generation of Data Source Descriptions for Agentic Workflows," published September 10, 2026, filed August 20, 2025, claiming priority to United States Provisional Application No. 63/768,550, filed March 7, 2025 https://burbagelaw.com/entries/2026-09-09-randomized-trial-finds-artificial-intelligence-draft.html https://burbagelaw.com/entries/2026-09-09-randomized-trial-finds-artificial-intelligence-draft.html Wed, 09 Sep 2026 12:00:00 +0000 Research One hundred thirty-three patent lawyers at eleven United States intellectual property firms were randomly assigned access to an artificial intelligence drafting tool for ninety days. Blinded patent attorneys scored the resulting work on enforceability, technical accuracy, strategic ambiguity, completeness and clarity. Access raised the quality of benchmark drafting work by 0.34 standard deviations at ten days and 0.38 standard deviations at ninety days, with the larger gains going to junior lawyers. When the tool was withdrawn, the lawyers who had used it still outperformed the control group by 0.32 standard deviations, but that advantage sat entirely with the senior lawyers, at 0.45 standard deviations. Junior lawyers showed no average improvement on unaided judgment tasks, and their scores spread out rather than rising: fewer middling results, and more that were either poor or excellent. What it changes: The measured gain is real, and it is largest exactly where the supervision burden is heaviest. Two things follow for anyone running a patent program. A quality figure produced under this design is a figure for supervised output, so the attorney review that produced it belongs in the budget rather than being treated as the cost the tool removes. And the people whose drafts improve most are the people least able to work without the tool afterward, which makes tool access a staffing and training decision rather than a procurement decision. David H. Autor, Tanya Rodchenko, Josh Martin, Zanna Iscenko, Scott Strand, David Pearl and Melissa Ferere, “Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting,” National Bureau of Economic Research Working Paper No. 35720 (September 2026) https://burbagelaw.com/entries/2026-09-02-congress-carries-united-states-patent-trademark-offi.html https://burbagelaw.com/entries/2026-09-02-congress-carries-united-states-patent-trademark-offi.html Wed, 02 Sep 2026 12:00:00 +0000 Costs The authority Congress granted the United States Patent and Trademark Office to set and adjust its own fees under the Leahy-Smith America Invents Act was scheduled to lapse on September 16, 2026, at the end of the fifteen-year period running from enactment. Section 2015 of Division B of Senate Resolution 830 amended Section 10(i)(2) of that Act to carry the authority to December 11, 2026 instead. The Subcommittee on Courts, Intellectual Property, Artificial Intelligence, and the Internet held an oversight hearing on renewal on September 2, 2026, taking testimony from Teresa Rea of Rea Consulting, Morgan Reed of the Association for Competitive Technology, Professor Melissa Wasserman of the University of Texas School of Law, and David Kappos of the Council for Innovation Promotion. The witnesses supported renewal but divided on whether Congress should attach conditions, among them a genuine sunset, limits on the Office's discretionary denial of inter partes review institution, and full reconstitution of the Patent Public Advisory Committee, which is operating with four of its nine voting members. Professor Wasserman urged that the Office recover more of its examination cost before allowance while keeping the small entity and micro entity discounts in place. What it changes: A patent budget built more than a year out should not assume the current fee schedule. The restructuring discussed at the hearing moves money forward, toward filing and examination and away from issuance and maintenance, which raises the cash requirement in the first eighteen months of every new matter. The program most exposed is the one with a large pending docket and a small issued base, which describes most companies filing heavily on artificial intelligence right now. Model the front-loaded case against the current docket before the December deadline, not after a notice of proposed rulemaking appears. Hearing, “Renewal of USPTO Fee Setting Authority: Giving Full Effect to the America Invents Act,” Subcommittee on Courts, Intellectual Property, Artificial Intelligence, and the Internet, Committee on the Judiciary, United States House of Representatives, September 2, 2026 https://burbagelaw.com/entries/2026-07-23-physically-separable-interface-limitation-carried-on.html https://burbagelaw.com/entries/2026-07-23-physically-separable-interface-limitation-carried-on.html Thu, 23 Jul 2026 12:00:00 +0000 Fed. Cir. The district court held the asserted claims of four digital picture frame and digital display patents ineligible under 35 United States Code section 101 on summary judgment. The Federal Circuit affirmed as to three of the patents and reversed as to claim 19 of United States Patent No. 6,442,573. The panel held that claim 19 was not directed to an abstract idea because it required a user interface physically separable from the picture frame, a requirement the specification tied to a concrete technical advance: operating and loading the device remotely, where the prior art required the user to be in front of it. What carried the claim was the structural limitation recited in the claim itself rather than any advantage asserted in the specification. Amazon petitioned for rehearing on August 24, 2026, arguing that the limitation cannot bear the weight the panel placed on it. What it changes: The claim survived on what it recited, read as a whole. The panel treated the physically separable interface as a critical concrete limitation within the claimed combination, and read the claim in light of a specification that explained the prior-art problem it solved, so no single limitation carries eligibility on its own. Where an artificial intelligence system has an architectural constraint the prior art lacked, such as where inference runs, what a component is separated from, or what a model is denied access to, put that constraint in an independent claim and describe the problem it solves in the specification, because the court will look at both. The decision is nonprecedential and binds no one, so use it as an illustration in an examiner interview rather than as authority in a brief, and watch the combined petition for rehearing and rehearing en banc that Amazon filed on August 24, 2026. Ceiva Opco, LLC v. Amazon.com, Inc., No. 2024-1721 (Fed. Cir. July 23, 2026) (nonprecedential) (Reyna, J., joined by Moore, C.J., and Lourie, J.); United States Patent Nos. 6,442,573, 9,203,930, 9,654,562 and 9,124,656 https://burbagelaw.com/entries/2026-07-14-patent-eligibility-restoration-act-gets-first-full-c.html https://burbagelaw.com/entries/2026-07-14-patent-eligibility-restoration-act-gets-first-full-c.html Tue, 14 Jul 2026 12:00:00 +0000 Legislation The Senate Judiciary Committee held its first full-committee hearing on the Patent Eligibility Restoration Act, the Tillis–Coons bill that would abolish the judge-made exceptions to 35 U.S.C. § 101 and replace them with five statutory exclusions, one of which covers processes that are substantially economic, financial, business, social, cultural, or artistic. The bill did not advance. Chairman Grassley described the hearing as fact-finding and said the committee had not reached consensus. What it changes: Nothing to plan around yet. This is the fourth Congress to consider a § 101 rewrite without passing one, so applications filed today should still be drafted to survive Alice and Mayo rather than a hoped-for statute. S. 1546 (119th Cong.); H.R. 3152 — Senate Judiciary Committee hearing, July 14, 2026 https://burbagelaw.com/entries/2026-07-14-uk-examiners-now-search-examine-software-claims-feat.html https://burbagelaw.com/entries/2026-07-14-uk-examiners-now-search-examine-software-claims-feat.html Tue, 14 Jul 2026 12:00:00 +0000 UK The United Kingdom Intellectual Property Office published a practice notice changing how examiners search and examine computer-implemented inventions after the Supreme Court's judgment of 11 February 2026 in Emotional Perception, which replaced the Aerotel approach to the exclusions in section 1(2) of the Patents Act 1977. Part 1 summarizes the judgment, including that an artificial neural network is a program for a computer whatever hardware implements it, and that programs generated by machine learning are computer programs. Part 2 tells examiners to stop applying Aerotel and sections 1.07 to 1.51.7 of the Manual of Patent Practice. Under the first hurdle, the "any hardware" test, a claim using technical means such as computer hardware is an invention, and a section 1(2) objection is raised only where the claim involves no technical means. Examiners then identify, feature by feature, which features contribute to the technical character of the claim as a whole, a feature contributing if it interacts with the technical subject matter to solve a technical problem, and may consult the European Patent Office Guidelines without being bound by them. Inventive step is still assessed under Pozzoli, with an objection where the only differences from the prior art make no technical contribution. Searches must cover the contributing features, and where an applicant answers an earlier Aerotel objection, examiners reconsider patentability afresh. What it changes: The judgment set the test; the notice decides what examiners do with pending files, and three points change United Kingdom prosecution practice. First, the notice ties the fresh assessment of an outstanding Aerotel objection to the applicant's response, so every pending application carrying one should get a response now, ordered by compliance date, rather than waiting for the Manual of Patent Practice to be rewritten. Second, the search examiner searches the features that contribute to technical character, so an application that explains its contribution only in commercial or user terms risks a search aimed at the wrong features; the specification should name the technical problem and tie each claimed feature of the model or its training to the effect it produces. Third, because examiners may consult the European Patent Office Guidelines, technical-effect arguments prepared for European prosecution can be used before the United Kingdom Office with little change, and one specification can serve both. Intellectual Property Office, Practice Notice, "Search and Examination of UK Patent Applications under the Patents Act 1977 (as amended)" (14 July 2026), implementing Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks [2026] UKSC 3 (11 February 2026) https://burbagelaw.com/entries/2026-07-14-wipo-counts-37-808-generative-artificial-intelligenc.html https://burbagelaw.com/entries/2026-07-14-wipo-counts-37-808-generative-artificial-intelligenc.html Tue, 14 Jul 2026 12:00:00 +0000 Research The World Intellectual Property Organization updated its 2024 report on generative artificial intelligence patents through 2025. Published generative artificial intelligence patent families rose from about 14,000 in 2023 to 18,862 in 2024 and 37,808 in 2025, so the more than 56,000 families published in those two years exceed the total for 2014 through 2023. Generative families were 8.7 percent of all published artificial intelligence families in 2025, up from 6.1 percent in 2023. Families directed to large language models rose from 881 in 2023 to more than 14,100 in 2025, overtaking generative adversarial networks, with diffusion models third; code-generation families rose from 339 to 1,616. China-based inventors published more than 43,000 families in 2024 and 2025. Families with United States inventors grew from about 1,185 in 2023 to more than 4,380 in 2025, and Japan's from 398 to 3,835, a compound annual growth rate of 210 percent driven largely by SoftBank, the largest holder with 2,985 families, nearly all published in 2025. Alphabet has 1,083 families, Microsoft 865, IBM 821 and Nvidia 497, and the report counts 35 filings by OpenAI. Only about 9 percent of 2025 families, 3,297, were international families. The counts are of published families, which lag filing by about eighteen months. What it changes: The prior art against a generative artificial intelligence claim is now dense, recent and largely in Chinese and Japanese, and most of it was filed in one country only: with about one family in eleven filed internationally, a search confined to United States and European publications will miss most of what an examiner in Beijing or Tokyo, or a well-funded challenger, can cite. Budget machine-translated full-text searching of Chinese and Japanese publications into every freedom-to-operate and validity study on language models, and draft claims to large language model systems on the assumption that the general architecture, the retrieval pattern and the agent loop are already in someone's publication. The same numbers show where assertion risk sits: the largest holders are SoftBank, research institutions, utilities and the large platform companies, not the model laboratories, whose own filings remain small. Because published counts lag filings by eighteen months, the 2025 figures describe what was filed in 2023 and 2024. World Intellectual Property Organization, "SPARK: Patent Trends Update in GenAI," WIPO Technology SPARK series (July 2026), with press release PR/2026/957, "GenAI Innovation Soaring, With Patent Activity Nearly Tripling in Two Years" (Geneva, July 14, 2026) https://burbagelaw.com/entries/2026-07-07-deep-learning-claims-held-abstract-where-network-its.html https://burbagelaw.com/entries/2026-07-07-deep-learning-claims-held-abstract-where-network-its.html Tue, 07 Jul 2026 12:00:00 +0000 Fed. Cir. The claims covered methods of using a deep learning device to analyze images of dental arches, to evaluate how an orthodontic aligner fit and to measure separation between teeth. Applying the two-step Alice test, the Federal Circuit held the claims were directed to collecting and analyzing information and supplied no inventive concept. The specification described the deep learning device as selectable from a list of well-known and available neural networks, and the court declined to treat the training methodology or the claimed accuracy gains as a technical improvement. What it changes: Accuracy and speed figures do not rescue a claim. What survives is a claimed technical mechanism — a specific architecture, a specific training-data construction, a specific way the model is integrated with the rest of the system — written into the specification before filing, not argued for the first time in a response. Dental Monitoring SAS v. Align Technology, Inc., No. 2024-2270 (Fed. Cir. July 7, 2026) (nonprecedential); U.S. Pat. Nos. 11,049,248 and 10,755,409 https://burbagelaw.com/entries/2026-06-04-jury-deciding-step-two-must-told-abstract-idea-claim.html https://burbagelaw.com/entries/2026-06-04-jury-deciding-step-two-must-told-abstract-idea-claim.html Thu, 04 Jun 2026 12:00:00 +0000 Fed. Cir. Ollnova asserted four patents on wireless building automation against ecobee's smart thermostats, and a jury found infringement of at least one, without saying which, and awarded a lump sum of $11.5 million. The Federal Circuit, in an opinion by Judge Chen joined by Judges Cunningham and Stark, ordered a new trial because a single infringement question covering all four patents was an abuse of discretion. Under Section 101 of Title 35 of the United States Code, it affirmed that claim 1 of United States Patent No. 7,746,887, whose controller polls a sensor only within a polling interval and transmits the most recent reading only during a transmission interval and only when the sensed condition moves outside a predetermined range, is "directed to a specific technology-based improvement in the operation of a network," because the timing constraints and conditional trigger are "a defined technique" rather than a result. It reached the same conclusion for United States Patent No. 8,264,371, which aggregates change-of-value messages and repeats transmission until acknowledged. For United States Patent No. 7,860,495, the district court had held the claims abstract but sent step two to the jury without identifying the abstract idea or instructing that it cannot supply the inventive concept. That was error and not harmless, and on remand the jury "should be instructed as to the abstract idea." What it changes: Two lessons for software and artificial intelligence portfolios. First, the eligible claims here are rules about when a device samples and when it transmits, and they survived because each claim states the rule rather than the goal. That is the drafting pattern for edge inference, model update scheduling and sensor-triggered transmission: recite the interval, the condition and the action, and explain in the specification which technical constraint, whether bandwidth, power or reliability, each one addresses. Second, a patentee who takes step two to a jury must expect the instruction to name the abstract idea and tell jurors it cannot be the inventive concept, so the expert's account of the inventive concept has to point at claim limitations beyond that idea. Ollnova's expert described the inventive concept in terms that tracked the abstract idea, and the verdict did not hold. Ask the court to fix the abstract idea before trial, which the opinion notes can avoid a retrial. Ollnova Technologies Ltd. v. ecobee Technologies ULC, Nos. 2025-1045, 2025-1046 (Fed. Cir. June 4, 2026) (precedential) (Chen, J., joined by Cunningham, J., and Stark, J.), affirming in part, dismissing in part, and vacating and remanding a judgment of the United States District Court for the Eastern District of Texas, No. 2:22-cv-00072-JRG (Gilstrap, J.) https://burbagelaw.com/entries/2026-05-18-supreme-court-declines-usaa-s-challenge-rule-compute.html https://burbagelaw.com/entries/2026-05-18-supreme-court-declines-usaa-s-challenge-rule-compute.html Mon, 18 May 2026 12:00:00 +0000 Supreme Court The Supreme Court denied United Services Automobile Association's certiorari petition, which asked whether the Federal Circuit "has wrongly extended the prohibition on patenting an 'abstract idea'" to "concrete technological processes," and whether it "has wrongly held that, as a matter of law, a computer-implemented technological invention is patent-eligible only if it claims improvements to computer functionality itself." The patents, including United States Patent Nos. 10,402,638 and 10,013,605, claim remote check deposit using a handheld device's camera, an application that guides the user in capturing images of the front and back of a check, and optical character recognition to confirm that the deposit can proceed. Juries in the Eastern District of Texas had awarded $218,450,000 for willful infringement in one case and $4,300,000 in another. On June 12, 2025 the Federal Circuit held claim 20 of the '638 patent ineligible under Section 101 of Title 35 of the United States Code, as directed to the abstract idea of depositing a check using a handheld mobile device with no inventive concept, in a precedential opinion, and applied the same analysis to three related patents in a nonprecedential one, eliminating both verdicts. PNC waived its response, the Court requested one, and the American Intellectual Property Law Association filed an amicus brief. The petition was denied after the May 14 conference, with Justice Alito taking no part. What it changes: The Federal Circuit's line between claims that improve computer functionality and claims that use a computer to improve a user's task stands without Supreme Court review, and the Court declined even after asking for a response. That line decides most claims to applied artificial intelligence, because a model that makes a business process faster or more accurate improves the task, not the computer, unless the claim recites what changes in the computing system. Draft accordingly: put the change to the system, such as the memory footprint, the data path, the training procedure or the hardware interaction, in the independent claim, and keep the user benefit in the specification. For enforcement, eligibility can erase a verdict of more than $200 million after trial, so a plaintiff's budget should assume a Section 101 appeal, and the claims chosen for trial should recite the technical mechanism. Any broader relief will have to come from Congress, where the eligibility bills remain pending. United Services Automobile Association v. PNC Bank N.A., No. 25-853 (U.S. May 18, 2026) (petition for a writ of certiorari denied; Alito, J., took no part), leaving in place United Services Automobile Association v. PNC Bank N.A., 139 F.4th 1332 (Fed. Cir. June 12, 2025), rehearing denied September 16, 2025 https://burbagelaw.com/entries/2026-05-13-european-union-delays-high-risk-artificial-intellige.html https://burbagelaw.com/entries/2026-05-13-european-union-delays-high-risk-artificial-intellige.html Wed, 13 May 2026 12:00:00 +0000 EU Policy Obligations for stand-alone high-risk artificial intelligence systems listed in Annex III move from 2 August 2026 to 2 December 2027. Obligations for artificial intelligence embedded in regulated products under Annex I move from 2 August 2027 to 2 August 2028. The core transparency obligations under Article 50 still take effect on 2 August 2026, with a watermarking grace period for existing systems running to 2 December 2026. The Article 5 prohibitions were expanded, and the deadline for Member States to establish regulatory sandboxes moved to 2 August 2027. What it changes: This is a compliance timetable rather than a patent rule, and it still moves filing dates. A client who was holding a European launch until the high-risk obligations landed now has sixteen more months of roadmap. Filings have to be on file covering what ships, so the launch date moving moves the filing plan with it. Digital AI Omnibus — provisional political agreement 6 May 2026, confirmed by Member State representatives 13 May 2026, amending Regulation (EU) 2024/1689 (the Artificial Intelligence Act) https://burbagelaw.com/entries/2026-04-16-uspto-extends-artificial-intelligence-prior-art-pilo.html https://burbagelaw.com/entries/2026-04-16-uspto-extends-artificial-intelligence-prior-art-pilo.html Thu, 16 Apr 2026 12:00:00 +0000 Costs The Office extended the Artificial Intelligence Search Automated Pilot Program, known as ASAP!, which opened in October 2025 for original, noncontinuing utility applications filed under Section 111(a) of Title 35 of the United States Code. Petitions, which must be filed on Form PTO/SB/470 in Patent Center on the application's filing date, with the specification, claims and abstract in DOCX format and enrollment in the e-Office Action Program, will be accepted until June 1, 2026 or until each utility technology center has 400 accepted applications, whichever comes first, instead of closing on April 20. On March 23 the Office had waived the petition fee under 37 C.F.R. 1.17(f) for petitions filed on or after that date and doubled its target to at least 3,200 applications. A participating applicant receives, after pre-examination processing and before examination, an Automated Search Results Notice listing up to ten documents from United States patents, published applications and foreign image and text collections, ranked by an internal artificial intelligence tool. No reply is required. The examiner considers the listed documents as part of the Office search file and lists them on a PTO-892 only if relied on in a rejection; they appear on the face of the patent only if the examiner lists them or the applicant cites them in an information disclosure statement. Continuing, national stage, design, plant and reissue applications are excluded. What it changes: For an applicant filing original utility applications on artificial intelligence or software before June 1, the pilot now costs nothing but one form on the filing date, and it returns the Office's own machine search months before a first action. Use it as a preliminary amendment opportunity: if the notice lists a reference that reads on the broadest claim, amend before the examiner picks up the case and save a round of prosecution. Cite in an information disclosure statement any listed reference the claims will be argued around, because the Office prints only documents the examiner lists or the applicant cites, and a continuation will not receive its own automated search. Record how the listed documents compare with the examiner's eventual search, because the Office is collecting exactly that data to decide whether automated search becomes a standard part of examination. United States Patent and Trademark Office, "USPTO extends the Artificial Intelligence Search Automated Pilot Program (ASAP!)," uspto.gov (Apr. 16, 2026); "USPTO waives the petition fee for the Artificial Intelligence Search Automated Pilot Program (ASAP!)," uspto.gov (Mar. 23, 2026); Automated Search Pilot Program, 90 Fed. Reg. 48161 (Oct. 8, 2025) https://burbagelaw.com/entries/2026-04-14-selling-device-runs-secret-code-bars-patent-what-cod.html https://burbagelaw.com/entries/2026-04-14-selling-device-runs-secret-code-bars-patent-what-cod.html Tue, 14 Apr 2026 12:00:00 +0000 Fed. Cir. United States Patent No. 8,458,689, with a priority date of March 30, 2001, claims methods and apparatuses for upgrading software in an engine controller. Definitive Holdings asserted eight claims against Powerteq, and the District of Utah granted summary judgment that all were invalid under the on-sale bar of pre-America Invents Act Section 102(b) of Title 35 of the United States Code. The prior sale was by non-party Hypertech Inc. of its Power Programmer III, a device that saved and restored original engine software and reprogrammed vehicles, before the critical date of March 30, 2000. The evidence was Hypertech's sales records, a 1996 installation guide, dated source code and the testimony of Hypertech's chief executive and owner. The Federal Circuit affirmed in an opinion by Judge Cunningham, joined by Chief Judge Moore and Judge Dyk. It held that the bar does not "require that the sale make the details of the invention available to the public," the question being "whether the sale relates to a device that embodies the invention," and noted that Definitive had not disputed that the device embodied every claim limitation. It held that "the actual source code commands themselves are not offered for their truth and thus are not hearsay," and that the executive's testimony was enough to authenticate the sales records and tie the analyzed code to the 1996 device. What it changes: Selling a product puts its inventions on sale even when no customer can see how they work, and this decision applies that rule to software running inside a sold device. A company that ships hardware with an embedded model, a firmware update or an on-device inference engine starts the one-year clock on every method that product performs, whether or not the code is encrypted or obfuscated. File before the first commercial shipment, including paid pilots, and treat the device release calendar as a filing calendar. The evidentiary holding matters to anyone defending against an old software patent: archived source code and dated builds of products sold before the critical date are admissible to show what the product did, so an engineering archive, the defendant's or a third party's, is a prior art source worth searching before paying a demand. The patent was governed by pre-America Invents Act law, but the court relied on Helsinn, which carries the same on-sale rule into the current statute. Definitive Holdings, LLC v. Powerteq LLC, No. 2024-1761 (Fed. Cir. Apr. 14, 2026) (precedential) (Cunningham, J., joined by Moore, C.J., and Dyk, J.), affirming summary judgment of the United States District Court for the District of Utah, No. 2:18-cv-00844-DBB (Barlow, J.), reported at 731 F. Supp. 3d 1243 (D. Utah 2024) https://burbagelaw.com/entries/2026-03-17-2026-epo-guidelines-add-dedicated-treatment-artifici.html https://burbagelaw.com/entries/2026-03-17-2026-epo-guidelines-add-dedicated-treatment-artifici.html Tue, 17 Mar 2026 12:00:00 +0000 EPO The 2026 edition of the EPO Guidelines for Examination adds a new general section on the use of artificial intelligence and tightens the sufficiency-of-disclosure requirements for artificial intelligence inventions at F-III, 3. Applicants must disclose the mathematical methods and the training data in enough detail that a skilled person can reproduce the invention without undue burden, and must make the asserted technical effect credible on the face of the application. What it changes: The European disclosure standard is now stricter than the United States written-description practice most applicants file to. For a family headed to both offices, the training-data description and the technical-effect showing have to be in the priority document, because neither can be added later without a new matter objection. Guidelines for Examination in the EPO (2026 edition), General Part 5 “The use of artificial intelligence”; E-III, 10.1; F-III, 3 https://burbagelaw.com/entries/2026-03-11-comparing-program-behavior-against-combined-model-ab.html https://burbagelaw.com/entries/2026-03-11-comparing-program-behavior-against-combined-model-ab.html Wed, 11 Mar 2026 12:00:00 +0000 Fed. Cir. Columbia University's United States Patent Nos. 8,601,322 and 8,074,115 claim detecting anomalous program executions by running part of a program in an emulator, "comparing a function call made in the emulator to a model of function calls," where "the model is a combined model created from at least two models created using different computers," and flagging the call as anomalous. A jury found that Norton's SONAR/BASH feature willfully infringed claims 2, 11 and 27 of the '322 patent and claim 2 of the '115 patent and awarded $185,112,727, including $94,037,265 for foreign sales; the district court, which had refused to find the claims abstract under Section 101 of Title 35 of the United States Code, enhanced damages 2.6 times. The Federal Circuit, by Judge Dyk, held the claims directed to "the abstract idea of comparing data against a model (created using different computers) to determine if it is anomalous." Distributed model building was a conceded abstract idea, and selective emulation and diversified models appeared only in the specification: "nothing in the claims requires the steps necessary to make the improvement." Whether the claimed model of function calls was conventional is a fact question, remanded for step two. The court also held that copies foreign customers downloaded and installed abroad were made outside the United States under Microsoft v. AT&T, eliminating the foreign royalty. What it changes: The court applied to a model-based claim the rule it stated a month earlier in GoTV Streaming v. Netflix: the improvement has to be in the claim. A claim that compares new data against a model to flag anomalies, which describes most monitoring, fraud and security products built on machine learning, is abstract at step one even when the model is built in a distributed way, unless the claim recites how the model is built or what it contains in terms that make the improvement unavoidable. The remand helps drafters. The court treated the conventionality of the claimed model as a question of fact, so a claim that defines the model concretely, backed by a specification explaining why that model was not routine, can survive the pleadings. For enforcement budgets, the foreign-sales holding means a royalty on software that customers download and install abroad needs a causation theory put to the jury, not a made-in-the-United-States theory. The Trustees of Columbia University in the City of New York v. Gen Digital Inc., Nos. 2024-1243, 2024-1244 (Fed. Cir. Mar. 11, 2026) (precedential) (Dyk, J., joined by Prost, J., and Reyna, J.), reversing in part, vacating in part and remanding a judgment of the United States District Court for the Eastern District of Virginia, No. 3:13-cv-00808-MHL (Lauck, C.J.) https://burbagelaw.com/entries/2026-03-11-director-adds-american-manufacturing-small-business.html https://burbagelaw.com/entries/2026-03-11-director-adds-american-manufacturing-small-business.html Wed, 11 Mar 2026 12:00:00 +0000 PTAB Director John A. Squires told users of the Patent Trial and Appeal Board that, in deciding whether to institute inter partes review or post-grant review, he will consider three further factors: the extent to which products accused of infringement in a parallel proceeding are manufactured in the United States or tied to investment in American manufacturing; the extent to which the patent owner's own competing products, made, sold or licensed, are manufactured in the United States; and whether the petitioner is a small business that has been sued for infringement of the challenged patent. The memorandum rests on the direction in Sections 316(b) and 326(b) of Title 35 of the United States Code that the Director consider the effect of institution standards on the economy and the integrity of the patent system. It cites the movement of electronics and computing manufacturing overseas and notes that many frequent petitioners report no significant domestic manufacturing. Manufacturing includes where components are made and whether domestic products are finished abroad. For method claims, the relevant product is the device used to perform the method, such as the computer performing a method of operating a computer. Small-business status turns on Small Business Administration size standards, eligibility for reduced fees and other relevant facts. The factors apply to every proceeding whose discretionary brief is not yet due. What it changes: Institution of an inter partes review now depends partly on the parties' factories, and for software and artificial intelligence patents the memorandum answers the threshold question of what the product is: for a method of operating a computer, it is the computer. A company sued on a software patent should establish, before discretionary briefing opens, where the servers, accelerators and devices that run the accused method are made and assembled, because that record now goes to the Director alongside the merits. A startup sued on a patent should identify itself as a small business in its petition with the evidence the memorandum names, since the third factor exists for petitioners in exactly that position. A patent owner whose technology is practiced by a domestic manufacturer, directly or under license, should put that fact in the record early. None of this changes the merits standard, but it changes what a defendant needs in hand in the first weeks after a complaint is served. Memorandum, "Additional Discretionary Institution Considerations—U.S. Manufacturing and Small Business Use of AIA Proceedings" (Mar. 11, 2026), from John A. Squires, Under Secretary of Commerce for Intellectual Property and Director of the United States Patent and Trademark Office, to all PTAB users https://burbagelaw.com/entries/2026-03-04-japan-s-supreme-court-refuses-dabus-appeal-leaving-n.html https://burbagelaw.com/entries/2026-03-04-japan-s-supreme-court-refuses-dabus-appeal-leaving-n.html Wed, 04 Mar 2026 12:00:00 +0000 Japan Japan's Supreme Court, Second Petty Bench, with Justice Kazumi Okamura presiding, decided on March 4, 2026 not to accept Stephen Thaler's appeal in the DABUS case, without giving reasons. Japanese Patent Application No. 2020-543051 entered the national phase from International Application PCT/IB2019/057809, which named the artificial intelligence system DABUS as inventor. When the applicant did not amend the national document to name a natural person, as Article 184-5 of the Patent Act requires, the Commissioner of the Japan Patent Office dismissed the application, and the Tokyo District Court upheld the dismissal on May 16, 2024. The Intellectual Property High Court affirmed on January 30, 2025. Reading the grant provision of Article 29(1) with the employee-invention provision of Article 35 and the requirements of Articles 36 and 66 that the inventor's name be stated and published, it held that the Patent Act "presupposes that the inventor is a natural person," so that no patent may be granted under the current statute for an invention made by artificial intelligence. It held that the inventor's name "is the mandatory information to be stated in a national document," that the Agreement on Trade-Related Aspects of Intellectual Property Rights imposes no obligation to protect such inventions, and that whether to grant them patents "is an issue that requires a discussion for legislation." What it changes: Japan now stands with the United States, the United Kingdom, Germany and Australia, whose courts of last resort have each left in place or imposed a natural-person requirement for inventors, and the Japanese courts have said that any change belongs to the legislature. The practical point in Japan is procedural. The inventor's name is a required entry in the domestic document filed at national phase entry, a defective entry draws a correction order, and an applicant who will not name a human loses the application without examination. For a portfolio that enters Japan from international applications, confirm before the thirty-month national phase deadline that every inventor designation names the people who made the claimed invention, and settle any internal dispute about whether a model rather than an engineer produced a claimed feature before entry, not after a correction order arrives. The Supreme Court gave no reasons, so the High Court's judgment is the authority to cite. Supreme Court of Japan, Second Petty Bench (Okamura, J., presiding), decision of March 4, 2026 declining to accept the appeal in the DABUS case, leaving in place Intellectual Property High Court (Second Division), judgment of January 30, 2025, 2024 (Gyo-Ko) 10006 (Shimizu, presiding), affirming Tokyo District Court, 2023 (Gyo-U) 5001 (May 16, 2024) https://burbagelaw.com/entries/2026-03-02-supreme-court-leaves-place-rule-work-generated-machi.html https://burbagelaw.com/entries/2026-03-02-supreme-court-leaves-place-rule-work-generated-machi.html Mon, 02 Mar 2026 12:00:00 +0000 Supreme Court The Supreme Court denied Stephen Thaler's petition for a writ of certiorari, which asked whether works "outputted by" an artificial intelligence system "without a direct, traditional authorial contribution by a natural person can be copyrighted." The work was a two-dimensional image, "A Recent Entrance to Paradise," which Thaler's registration application attributed to his generative system, the Creativity Machine, and described as "[c]reated autonomously by machine." The Copyright Office refused registration because no human created the work, and the District of Columbia Circuit affirmed on March 18, 2025, holding that the Copyright Act requires a human author. The petition, filed October 9, 2025 with Ryan Abbott as counsel of record, also argued that Thaler owned the output as the system's owner and employer. The government waived a response, the Court requested one, and the Solicitor General's brief in opposition, filed January 23, 2026, argued that the case "does not implicate questions about the nature or extent of the human contribution" needed to register works made with artificial intelligence, because Thaler had disavowed any creative control, and that no court of appeals had reached his result. Amicus briefs were filed by Professors Shlomit Yanisky-Ravid, Lawrence Lessig and others and by Phyllis Schlafly Eagles. The petition was distributed for the conference of February 27, 2026 and denied on March 2, 2026. What it changes: Human authorship is now as settled for copyright as the natural-person rule has been for inventorship since the Court declined the DABUS patent case in 2023, and the two regimes run in parallel: a machine can be neither author nor inventor, and the open question in both is how much a human must contribute. For a company whose code, documentation, designs or model outputs are increasingly generated by models, copyright will not protect the portions no human authored, so functionality worth protecting has to be claimed in a patent or held as a trade secret. Build one contribution record that serves both regimes: for each asset, which person framed the problem, chose and edited the outputs and made the decisions the claims or the registration rest on. The government's brief expressly set aside works made with the assistance of artificial intelligence, so registrations that disclose and disclaim machine-generated material remain available. Thaler v. Perlmutter, No. 25-449 (U.S. Mar. 2, 2026) (petition for a writ of certiorari denied), leaving in place Thaler v. Perlmutter, No. 23-5233 (D.C. Cir. Mar. 18, 2025) (Millett, J., joined by Wilkins, J., and Rogers, S.J.), rehearing denied May 12, 2025 https://burbagelaw.com/entries/2026-02-24-applying-known-machine-learning-technique-new-field.html https://burbagelaw.com/entries/2026-02-24-applying-known-machine-learning-technique-new-field.html Tue, 24 Feb 2026 12:00:00 +0000 Fed. Cir. The patent claimed a method of processing natural language input using case-based reasoning, a machine learning technique that resolves ambiguity by drawing on stored historical examples in a metadata database. The Federal Circuit affirmed ineligibility, holding that generic use of artificial intelligence without other parameters is abstract, and that a conventional application of case-based reasoning, even to a novel environment, does not become eligible because the environment is new. The specification's own description of the database elements as well known in the art was used against the patentee at step two. What it changes: Background sections that concede prior-art status get quoted back at step two. Describe what is conventional at the level of generality the claims do not depend on, and reserve the detailed description for the parts of the system that are actually new. Rensselaer Polytechnic Institute v. Amazon.com, Inc., No. 24-1725 (Fed. Cir. Feb. 24, 2026) (Dyk, J.); U.S. Pat. No. 7,177,798 https://burbagelaw.com/entries/2026-02-11-uk-supreme-court-drops-aerotel-moves-neural-network.html https://burbagelaw.com/entries/2026-02-11-uk-supreme-court-drops-aerotel-moves-neural-network.html Wed, 11 Feb 2026 12:00:00 +0000 UK Emotional Perception's application claims a system that recommends media files using an artificial neural network trained on pairs of files compared by two measures, so that distances between file vectors track semantic distances. The Hearing Officer refused the application, the High Court allowed the applicant's appeal, and the Court of Appeal restored the refusal. The Supreme Court agreed that a network is "a set of instructions to manipulate data in a particular way" and so "a program for a computer" (para 87), however its weights were set. It held there is "a compelling case, in the light of G1/19, for rejecting the Aerotel approach" (para 64) and adopted the approach of the EPO Boards of Appeal. The claims involve technical means "because the ANN can only be implemented on some form of computer hardware" and refer to a database, a communications network and a user device (para 98), so their subject matter "constitutes an invention within the meaning of article 52" (para 100). Excluded matter is dealt with at inventive step through the intermediate step of G 1/19, under which "the sole criterion is whether the feature contributes to the technical character of the invention as a whole" (para 109). The court allowed the appeal, set aside the refusal and left inventive step to the Hearing Officer (para 118). What it changes: The United Kingdom now examines computer-implemented and machine learning inventions the way the EPO does: any hardware clears the exclusion, and the contest moves to inventive step, where only features that contribute to technical character count. A trained network gets no special route, since the court held it is a program like any other. For United Kingdom filings and validations, draft as for the EPO: identify the technical effect each model feature produces, such as reduced computation, memory or latency or a measurement of a physical signal, state it in the specification, and tie it to the claimed features, because a recommendation that is merely more semantically apt is the kind of feature the intermediate step can filter out. Applications refused or narrowed under Aerotel deserve a fresh look, and pending appeals should be argued on the new approach. The court decided only the exclusion; the Hearing Officer still decides inventive step. Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks [2026] UKSC 3 (11 February 2026) (Lord Briggs and Lord Leggatt, with whom Lord Hamblen, Lord Stephens and Lord Kitchin agree), allowing the appeal from [2024] EWCA Civ 825 https://burbagelaw.com/entries/2026-01-01-china-raises-inventiveness-bar-artificial-intelligen.html https://burbagelaw.com/entries/2026-01-01-china-raises-inventiveness-bar-artificial-intelligen.html Thu, 01 Jan 2026 12:00:00 +0000 CNIPA The amended guidelines state that changing only the application scenario or the target object of an existing artificial intelligence algorithm does not establish an inventive step. An applicant must show meaningful changes to model structure, parameters, training methods or technical effects. Disclosure requirements were tightened at the same time: a specification must describe the model architecture including modules, layers and connections, the training steps and key parameters, and the technical relationship between the input and the output data. Examiners may examine the description alongside the claims. The guidelines also permit rejection on public policy grounds where an invention involves unlawful collection or use of personal data, or algorithmic decision-making on discriminatory criteria. What it changes: China has arrived at the Federal Circuit's central point from the opposite direction. Applying a known model to a new field is not enough in either country. For a family filed in the United States, Europe and China, the model architecture and the training description have to sit in the priority document, because none of the three offices will let that material be added afterwards. Amended Patent Examination Guidelines, China National Intellectual Property Administration Order No. 84, issued 10 November 2025, effective 1 January 2026 https://burbagelaw.com/entries/2025-12-05-examination-manual-now-lists-improving-how-machine-l.html https://burbagelaw.com/entries/2025-12-05-examination-manual-now-lists-improving-how-machine-l.html Fri, 05 Dec 2025 12:00:00 +0000 USPTO The United States Patent and Trademark Office revised Sections 2106.04(d), 2106.05(a) and 2106.05(f) of the Manual of Patent Examining Procedure to incorporate Ex parte Desjardins, the Appeals Review Panel decision of September 26, 2025 designated precedential on November 4, 2025. Section 2106.05(a) adds two examples of technological improvements: "An improved way of training a machine learning model that protected the model's knowledge about previous tasks" and "Improvements to computer component or system performance based upon adjustments to parameters of a machine learning model." Section 2106.04(d) now summarizes Desjardins, where a claim to training a model on a series of tasks recited a mathematical concept but integrated it into a practical application because the specification described improvements to how the model operates, including overcoming catastrophic forgetting. Examiners are told not to evaluate claims "at such a high level of generality" that technical limitations are dismissed, nor to dismiss elements as generic computer components without considering whether they provide an improvement, and that "the claim must include the components or steps of the invention that provide the improvement described in the specification," though it need not recite the improvement. The Director's December 4 memorandum on Subject Matter Eligibility Declarations requires examiners to weigh on its merits an optional declaration under 37 C.F.R. 1.132 offering expert testimony or objective data on a technological improvement. What it changes: Every response to an eligibility rejection of a machine learning claim under Section 101 of Title 35 of the United States Code should now cite Section 2106.05(a) examples xiii and xiv and the Desjardins paragraph in Section 2106.04(d), because the manual binds the examiner in a way an argument from a single Appeals Review Panel decision did not. The drafting rule is stated in the manual's own words: the specification has to describe the improvement to how the model trains or runs, and the claim has to contain the steps that produce it. That is the line between Desjardins and Recentive Analytics v. Fox, where applying a generic model to a new data environment was abstract. The declaration route rewards keeping evidence in the invention record, such as ablation results, memory and latency measurements and benchmark comparisons, because a declaration can prove an improvement the specification describes but cannot add one it omits. The manual governs examiners only; courts apply Recentive. Memorandum, "Advance notice of change to the MPEP in light of Ex Parte Desjardins" (Dec. 5, 2025), from Charles Kim, Deputy Commissioner for Patents, to the Patent Examining Corps; Memorandum, "Subject Matter Eligibility Declarations" (Dec. 4, 2025), from John A. Squires, Under Secretary of Commerce for Intellectual Property and Director of the United States Patent and Trademark Office, to the Patent Examining Corps https://burbagelaw.com/entries/2025-11-28-uspto-rescinds-2024-inventorship-guidance-treats-art.html https://burbagelaw.com/entries/2025-11-28-uspto-rescinds-2024-inventorship-guidance-treats-art.html Fri, 28 Nov 2025 12:00:00 +0000 USPTO The USPTO withdrew its February 2024 guidance on artificial-intelligence-assisted inventions and replaced it with guidance treating artificial intelligence systems as tools rather than as candidate joint inventors. The Office stated that the Pannu joint-inventorship factors apply only among natural persons and cannot be extended to an artificial intelligence system, and directed examiners to apply the ordinary inventorship standard without a separate pathway for applications involving artificial intelligence. What it changes: Thaler v. Vidal still controls: every named inventor must be a natural person. The practical change is that inventorship questionnaires no longer need a separate significant-contribution analysis for the model itself; they need a record of which human made the conception decisions the claims rest on. Revised Inventorship Guidance for AI-Assisted Inventions, 90 Fed. Reg. 54636 (Nov. 28, 2025), rescinding 89 Fed. Reg. 10043 (Feb. 13, 2024) https://burbagelaw.com/entries/2025-11-04-english-court-holds-stable-diffusion-s-weights-not-i.html https://burbagelaw.com/entries/2025-11-04-english-court-holds-stable-diffusion-s-weights-not-i.html Tue, 04 Nov 2025 12:00:00 +0000 UK Getty Images sued Stability AI in the High Court over the Stable Diffusion image models. By the end of trial Getty had abandoned its primary copyright claim over training, there being no evidence of training in the United Kingdom, its claim over outputs, because prompt blocking had achieved the relief sought, and its database right claim. What remained was secondary infringement under sections 22 and 23 of the Copyright, Designs and Patents Act 1988, on the theory that the model is an "infringing copy" because making its weights would have infringed if done in the United Kingdom, trade mark claims under section 10 of the Trade Marks Act 1994 over Getty and iStock watermarks in outputs, and passing off. The judgment records that the model does not store training data. Mrs Justice Joanna Smith held that an "article" can be intangible, but that an infringing copy must at some time have contained or stored a copy of the work, and because Getty accepted, as both sides' experts agreed, that the weights contain no copies of its works, the secondary infringement claim failed. The trade mark claims succeeded only for a few watermarked outputs of the first two model versions, findings described as extremely limited; the claims for the SD XL and v1.6 models and under section 10(3) failed, and passing off added nothing. What it changes: The judgment accepts, on agreed expert evidence, that trained weights do not contain the works they were trained on, and it treats the model as the developer's own technical object rather than a copy of anyone's images. That supports protecting weights as a trade secret and claiming the training method and model structure as inventions in their own right. The case was decided on expert evidence of what the parameters represent, and a patent specification should carry the same kind of explanation, because what the weights encode is also what written description and sufficiency will be measured against. The territorial point cuts across patents too: Getty abandoned the training claim for want of evidence of training in the United Kingdom, and a training-method claim is likewise infringed only where training is done, so portfolios need claims to the trained model, the inference system and its uses in the markets where the model is deployed. This is a first-instance judgment on a narrowed case. Getty Images (US) Inc v Stability AI Ltd [2025] EWHC 2863 (Ch) (4 November 2025) (Joanna Smith J), Case No. IL-2023-000007 https://burbagelaw.com/entries/2025-10-17-uspto-director-takes-over-inter-partes-review-instit.html https://burbagelaw.com/entries/2025-10-17-uspto-director-takes-over-inter-partes-review-instit.html Fri, 17 Oct 2025 12:00:00 +0000 PTAB Director John Squires announced that from October 20, 2025 the Director will determine whether to institute inter partes review and post-grant review, in consultation with at least three judges of the Patent Trial and Appeal Board. Where institution is appropriate on at least one ground for one challenged claim, the Director issues a summary notice granting institution; otherwise a summary notice denying it. Routine institution decisions are limited to summary notices, the Director may write a decision on novel or important factual or legal issues, and may refer a petition to Board members for detailed treatment such as complex claim construction, priority or real party in interest questions. Instituted trials go to three-judge panels, which no longer decide institution on the merits; briefing under the March 26, 2025 interim procedures continues. The same day the United States Patent and Trademark Office published proposed rules that would bar institution unless each petitioner stipulates that it and its privies will not pursue invalidity under Section 102 or 103 of Title 35 of the United States Code in any other forum; bar it where a challenged claim was previously upheld by a district court, the International Trade Commission or a Board final written decision; and bar it where a district court trial or Commission determination on those grounds will more likely than not come before the final written decision. What it changes: For a company defending against software or artificial intelligence patents, the Board is now a forum the Director opens case by case, usually without explanation, and the invalidity case should be budgeted and built for district court from the start, with the petition as an option rather than the plan. If the proposed rules are adopted as written, an instituted petition would cost the petitioner its anticipation and obviousness defenses everywhere else, so the choice of forum for the best prior art would have to be made once, early, and with the trial team in the room. Defendants sued in fast districts would rarely qualify at all. Patent owners holding older software and machine learning patents gain bargaining power in licensing and should assert where trial dates come quickly. The memorandum is in effect; the rules are proposed and may change before adoption. Memorandum, "Director Institution of AIA Trial Proceedings" (Oct. 17, 2025), from John A. Squires, Under Secretary of Commerce for Intellectual Property and Director of the United States Patent and Trademark Office, to all Patent Trial and Appeal Board judges; Revision to Rules of Practice before the Patent Trial and Appeal Board, 90 Fed. Reg. 48335 (Oct. 17, 2025) (proposed rule) https://burbagelaw.com/entries/2025-09-26-appeals-review-panel-finds-machine-learning-claims-e.html https://burbagelaw.com/entries/2025-09-26-appeals-review-panel-finds-machine-learning-claims-e.html Fri, 26 Sep 2025 12:00:00 +0000 PTAB The USPTO Director convened an Appeals Review Panel to review a Board rejection of machine learning method claims under § 101. The panel disagreed with the Board's step-two analysis and held that technological advantages of the claimed method, including reduced storage requirements and efficiency gains, amount to a practical application of the abstract idea where those advantages are reflected in the claim language itself. What it changes: This is the single most useful citation available in prosecution right now for a machine learning claim carrying a resource-efficiency benefit. It works only if the efficiency is recited in the claim, not merely asserted in the specification or the response. Ex parte Desjardins, Appeals Review Panel decision (Sept. 26, 2025) (convened sua sponte by the Director) https://burbagelaw.com/entries/2025-09-25-federal-court-preliminarily-approves-anthropic-s-1-5.html https://burbagelaw.com/entries/2025-09-25-federal-court-preliminarily-approves-anthropic-s-1-5.html Thu, 25 Sep 2025 12:00:00 +0000 Copyright After Judge William Alsup's June 23, 2025 order held that training on books was fair use but that pirated library copies were not, the court on July 17, 2025 certified a "LibGen & PiLiMi Pirated Books Class," and trial was set for December 1, 2025. On September 5 the class moved for preliminary approval of a settlement. Anthropic will pay at least $1.5 billion plus interest into a non-reversionary fund; with around 500,000 works in the class, class counsel estimated a gross recovery of $3,000 per work, and Anthropic will pay a further $3,000 for each work added above 500,000. Anthropic will destroy the original files it obtained from Library Genesis and the Pirate Library Mirror and copies originating from them, subject to preservation obligations. The release covers only past conduct through August 25, 2025, and releases no claim, past or future, arising from allegedly infringing outputs of Anthropic's models. Class counsel called it the largest publicly reported copyright recovery and compared the per-work figure with the $750 statutory minimum and the $200 floor for innocent infringement. Funds are allocated equally per work. On September 25, after a September 8 hearing and written questions from the court, Judge Alsup granted preliminary approval of the class settlement, the plan of distribution and the plan for notice. What it changes: Pirated training data now has a price per work, and a model budget that does not carry acquisition or licensing cost for its corpus is missing a line. For the patent program the consequence is in diligence and continuity. A patent on a model, a training method or a product built on the model does nothing to offset liability for the corpus the model learned from, so any acquisition of a company, a portfolio or a model license should ask for the acquisition record of the training data alongside the patent schedule. The destruction term shows that a settlement can take datasets out of service, and a product line whose patented features depend on a model trained on those datasets needs a retraining plan. Output claims were expressly left open, so the filtering and attribution mechanisms that keep a model from reproducing its training text remain worth both engineering investment and patent filings. Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal. Sept. 25, 2025) (Alsup, J.) (minute entry granting preliminary approval of class settlement, Dkt. 427), on the Motion for Preliminary Approval of Class Settlement (Dkt. 363, filed Sept. 5, 2025) https://burbagelaw.com/entries/2025-09-24-xai-sues-openai-over-engineers-who-left-training-cod.html https://burbagelaw.com/entries/2025-09-24-xai-sues-openai-over-engineers-who-left-training-cod.html Wed, 24 Sep 2025 12:00:00 +0000 Trade Secrets xAI sued three OpenAI entities for misappropriation under the Defend Trade Secrets Act, intentional interference with prospective economic advantage, and unfair competition under California law. The complaint alleges that the same OpenAI recruiter pursued two early xAI engineers. Xuechen Li, who worked on training Grok, allegedly uploaded the entire xAI code base to a personal cloud account on July 25, 2025, reportedly admitted taking it with a slide deck on training and tuning methods, and in xAI's earlier suit against him was restrained on September 2 from any role at OpenAI until the material was deleted. Jimmy Fraiture, who led the production inference team, allegedly transferred source code by AirDrop on July 31, including inference cluster code, training and infrastructure code written by others, and co-founders' experimental folders. A senior finance executive allegedly took knowledge of xAI's data center deployment. The complaint pleads that "Trade secrets protect nearly all of xAI's developments—model weights, training data, tuning methods, know-how, and more," and lists its measures: security training, background checks, the NIST 800-171 Rev. 3 baseline, SOC 2 Type II compliance, full disk encryption, a written security policy, confidentiality and invention assignment agreements, and termination certifications. It seeks damages, injunctive relief, removal of xAI information from OpenAI's systems, and destruction of OpenAI technology or models developed with it. These are allegations, not findings. What it changes: The complaint reads as a checklist for any model developer that relies on trade secrets: name the secret, show who had access, show the measures, and prove the taking, here from forensic logs, cloud records and admissions. A company that cannot produce each of those for its training code, inference stack and infrastructure methods has a weaker claim than it thinks. The remedy requested, destruction of models built with the information, also shows what a recipient risks when it hires a competitor's engineers. For the patent decision, the techniques most likely to leave with an engineer, how inference clusters stay reliable, how training infrastructure is built, how data centers come up fast, are the ones where a patent pays, because it is enforced against a competitor without proving anyone took anything. Hiring from a competitor also calls for screening the new hire's invention disclosures for inventorship and derivation problems before they become applications. X.AI Corp. and X.AI LLC v. OpenAI, Inc., OpenAI Global, LLC and OpenAI OpCo, LLC, No. 3:25-cv-08133 (N.D. Cal. filed Sept. 24, 2025) (complaint); see X.AI Corp. v. Li, No. 3:25-cv-07292 (N.D. Cal. filed Aug. 28, 2025) https://burbagelaw.com/entries/2025-08-11-claim-reciting-motor-physically-moves-dumbbell-s-wei.html https://burbagelaw.com/entries/2025-08-11-claim-reciting-motor-physically-moves-dumbbell-s-wei.html Mon, 11 Aug 2025 12:00:00 +0000 Fed. Cir. United States Patent No. 7,578,771 claims a weight selection and adjustment system. Claim 1 recites a selectorized dumbbell with nested left and right stacks of weight plates, a handle, and a movable selector whose adjustment positions couple selected numbers of plates to each end, together with an electric motor "operatively connected to the selector" that "physically moves the selector into the adjustment position corresponding to the desired exercise weight." Claim 20 adds a data entry device through which the user inputs a weight selection that controls the motor. Judge Jill Parrish of the District of Utah dismissed claims 1 through 18 and 20 under Rule 12(b)(6) as directed to the abstract idea of automated weight stacking under Section 101 of Title 35 of the United States Code. The Federal Circuit reversed at step one. Claim 1 is "limited to a specific implementation of a technological improvement" to selectorized dumbbells and is "directed to an eligible mechanical invention," unlike the "do it on a computer" claims in University of Florida Research Foundation v. General Electric and the Chamberlain v. Techtronic claims, which did not recite the barrier they controlled. Citing Diamond v. Diehr, the panel declined to ignore limitations found in the prior art, and it cautioned against conflating novelty and obviousness with step one. Claims 2 through 18 and 20, not argued separately, stood with claim 1. What it changes: Machine learning in a physical product, in robotics, vehicles, medical devices or industrial control, is claimable at step one when the claim recites the machine the model acts on and the actuator that does the acting. A claim that says a controller issues a command based on a model's output, without the structure that receives it, reads like Chamberlain; a claim that names the mechanism, its coupling and the physical movement reads like this one. Draft the independent claim to include the plant, the actuator, the coupling and the physical result, and put the model's role in the step that selects the movement. In prosecution and litigation, Diehr is now applied in a precedential opinion against the argument that conventional components can be stripped out before asking what a claim is directed to, and that argument belongs under novelty and obviousness. The panel noted that this was a simple mechanical invention, decided on a motion to dismiss. PowerBlock Holdings, Inc. v. iFit, Inc., No. 2024-1177 (Fed. Cir. Aug. 11, 2025) (precedential) (Stoll, J., joined by Taranto, J., and Scarsi, District Judge, sitting by designation), reversing PowerBlock Holdings, Inc. v. iFit, Inc., No. 1:22-cv-00132-JNP-CMR (D. Utah Sept. 29, 2023) (Parrish, J.) https://burbagelaw.com/entries/2025-08-04-director-s-memorandum-narrows-mental-process-rejecti.html https://burbagelaw.com/entries/2025-08-04-director-s-memorandum-narrows-mental-process-rejecti.html Mon, 04 Aug 2025 12:00:00 +0000 USPTO The Director issued a memorandum to the software and business-method art units reminding examiners that a process falls in the mental-process grouping only if a human could practically perform it, not merely if a human could theoretically perform some version of it, and that an eligibility rejection must be supported under the same preponderance-of-the-evidence standard as any other rejection. What it changes: Cite the memorandum by name in the response and make the examiner carry the evidentiary burden. Where a claim recites a computation no person could practically perform — the data volume, the latency requirement, or the model size makes it impossible — say so on the record with the numbers from the specification. Memorandum, “Reminders on Evaluating Subject Matter Eligibility of Claims under 35 U.S.C. 101” (Aug. 4, 2025), directed to Technology Centers 2100, 2600, and 3600 https://burbagelaw.com/entries/2025-08-02-general-purpose-model-providers-european-union-must.html https://burbagelaw.com/entries/2025-08-02-general-purpose-model-providers-european-union-must.html Sat, 02 Aug 2025 12:00:00 +0000 EU Policy Chapter V of the Artificial Intelligence Act, on general-purpose models, applies from 2 August 2025 under Article 113(b), with the exception of Article 101 on fines. Article 53(1) requires every provider of a general-purpose model to keep technical documentation "including its training and testing process"; to give information and documentation to providers that integrate the model, "[w]ithout prejudice to the need to observe and protect intellectual property rights and confidential business information"; to "put in place a policy to comply with Union law on copyright and related rights," including reservations of rights under Article 4(3) of Directive (EU) 2019/790; and to "draw up and make publicly available a sufficiently detailed summary about the content used for training" on a template from the AI Office. Models released under a free and open-source licence are exempt from the first two duties only, and not if they carry systemic risk. Under Article 111(3), providers whose models were placed on the market before 2 August 2025 have until 2 August 2027. The Commission published the voluntary General-Purpose AI Code of Practice on 10 July 2025, with chapters on transparency, including a model documentation form, copyright, and safety and security, and on 24 July 2025 adopted the mandatory template, under which the summary lists the main data collections and explains other sources. What it changes: The training content summary is a mandatory publication for every general-purpose model placed on the European Union market, and anything it discloses about data selection, filtering, deduplication, synthetic data generation or handling of opt-outs is prior art from the day it is posted and no longer a trade secret. Patent counsel should see each summary before it is published. Any pipeline invention the summary would reveal should be filed first, and the summary should be written at the level of generality the template requires and no lower. The technical documentation and downstream information duties do not publish anything, but documentation handed to integrating providers travels under contract, so the confidentiality terms in those agreements need the same review as a license. For models already on the market the 2027 deadline leaves time to file on pipeline inventions that the first summary would expose. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 53, 111(3) and 113(b), obligations of providers of general-purpose AI models applicable from 2 August 2025; European Commission, General-Purpose AI Code of Practice (10 July 2025); European Commission, Explanatory Notice and Template for the Public Summary of Training Content for general-purpose AI models (24 July 2025)