Artificial intelligence patent developments

What it changesShowing users a model-written summary of a hidden chain of thought was public seven months before the provisional, in OpenAI's own words. Whether that post counts against OpenAI depends on the inventor's one-year grace period under Section 102(b)(1)(A) of Title 35 of the United States Code, but it counts against anyone who files on the idea now, so an application on reasoning summaries filed today has to claim how the summaries are produced, not that they are shown. That is what OpenAI did. Claim 1 recites the plumbing: a buffer, a determination of when to release it, a queue for tool results, and a summarization model that receives a subset of the reasoning tokens at each release. Upon review, I conclude that the release determination is likely where examination will turn, because claim 1 says that the system determines when to release the tokens without saying how, and I would expect the claim to end up closer to claim 6, which ties the release to a difference in token count or an elapsed time. The second lesson is in the continuation, filed twenty-four days after the parent. The server that generates a reasoning feed and the application that displays it are often built and sold by different companies, and a method claim whose steps are split between them risks having no single infringer; the parent claims the server side and the continuation the client side. When drafting for a system that splits work between a server and a client, consider an independent claim for each side.

What it changesUpon review, I conclude that the timing of the Greek application, a first filing in Greece by a Mountain View company with inventors in New Jersey, New York and California, one Monday before the preprint went up the next, points to a filing made to put a date in front of the paper. That date matters most outside the United States, where the European Patent Office and China's office have no grace period for a preprint an inventor posts; in the United States the paper was forgiven either way. The lesson for a research group is the one Google's scaling application taught on this page earlier this month: the first filing, normally a United States provisional, goes in before the paper is posted. The other lesson concerns clearance. This claim did not appear in a United States publication until October 22, 2026, twenty months after the paper, so a clearance search run in that interval found the paper and not the claim. Consider whether you should rely on a clearance search on a transformer architecture feature first described in a recent paper. When performing a clearance search now, search by class, not by keywords alone: the paper calls the architecture DeepCrossAttention, and that name appears nowhere in the application, which sits in G06N 3/08 and G06N 3/048. The antecedent problems in claims 2 and 7 are likely to draw a rejection under Section 112(b) of Title 35 of the United States Code, which Google can cure by amendment.

What it changesThe Court has declined, for now, to review the Director's use of a patent's age as a reason to refuse review. For the holder of an older artificial intelligence or software patent, the Office's practice, which the denial leaves in place, bears on the maintenance-fee decision: a claim that still reads on current systems is harder to challenge at the Board than it was two years ago, when the Office instituted 68 percent of petitions in all fields against 41 percent this year, which is a reason to pay the fee on it rather than let it lapse. For a company already accused of infringing such a patent, the Board is no longer the default venue, with just over a third of electrical and computer petitions instituted, and the realistic alternatives are an ex parte reexamination request, which turns on a substantial new question of patentability rather than on the patent's age, and the district court. Upon review, I conclude that the discretionary stage is where petitions now fail: across all fields, four of every five denials came at that stage, and 77 percent of the petitions decided at the merits stage were instituted. The Office's figures do not break that stage out by technology, so the same split cannot be confirmed for electrical and computer petitions alone.

What it changesThe 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.

What it changesAntedating 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.

What it changesFor a company whose model work is done partly in China, the patent file is now an export file. An assignment from a Chinese subsidiary to a United States parent of a patent or application on a listed technology is a transfer of ownership abroad that must be registered from October 16, and once the catalogue is final a restricted item will need a license, while a prohibited item cannot be exported at all. When an assignment or license from a Chinese entity is being prepared, consider whether it covers a technology on the draft list, because the registration requirement reaches transfers that have begun but are not finished. The draft is open for comment until November 15, 2026, and the final list may differ, but the registration requirement applies now.

What it changesThe 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.

What it changesFive 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.

What it changesThe 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.

What it changesClaim 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.

What it changesA 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.

What it changesFor 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.