Three new eligibility examples show the claims the Office will allow on network training, speech separation and treatment

2024-07-17USPTO

United States Patent and Trademark Office, "July 2024 Subject Matter Eligibility Examples," Examples 47 through 49 (July 17, 2024), issued with the 2024 Guidance Update on Patent Subject Matter Eligibility, Including on Artificial Intelligence, 89 Fed. Reg. 58128 (July 17, 2024)

The Office issued three hypothetical examples with its guidance update. Example 47, anomaly detection: claim 1, an application-specific integrated circuit with neuron arrays and synaptic circuits storing weights, recites no judicial exception and is eligible; claim 2, a method that discretizes continuous training data, trains an artificial neural network by backpropagation and gradient descent, detects anomalies and outputs anomaly data, recites mental processes and mathematical concepts and is ineligible because the computer and network are mere instructions to apply the idea; claim 3, which adds determining that an anomaly is linked to malicious packets and dropping them and blocking their source in real time, is eligible because those steps reflect the specification's network security improvement. Example 48, speech separation: claim 1, a short-time Fourier transform spectrogram fed to a deep neural network for embedding vectors, is ineligible as mathematical concepts; claim 2, which clusters the embeddings, applies binary masks, synthesizes separate waveforms and stitches a signal that excludes a target source, is eligible; claim 3, a medium that separates speakers with k-means clustering and masks and transcribes one source, is eligible as speech-to-text conversion. Example 49 holds a treatment claim that uses a model's polygenic risk score and recites "administering an appropriate treatment" ineligible as a generic instruction to apply the result, and claim 2, which specifies "Compound X eye drops," eligible as a particular treatment.

What it changesThe examples draw the line the Office wants claims drafted to. A training claim that recites the standard loop, backpropagation and gradient descent on discretized data, is treated as mathematics on a generic computer and fails; the same loop followed by a concrete technical action on the output, dropping packets and blocking a source, passes. Transform-and-embed is mathematics; transform, embed, cluster, mask and reconstruct a signal is an application. A model that scores risk for a physician fails; naming the drug passes. For every independent claim on a model, the drafting question is what the system does with the output that a person could not do by writing it down, and that step belongs in claim 1, not in a dependent claim. None of the examples finds eligibility in the training method alone, so a claim set built on a training technique needs a hardware claim like Example 47 claim 1 or an output-action claim to anchor it. The examples direct Office personnel and do not bind the courts.

More on Eligibility of software and AI claims

← All recent developments · Search all developments

Burbage Law is the practice of Jon Burbage, a registered patent attorney. 342 United States patents have issued from applications he wrote; the full list is on the Patents page. More about Jon Burbage.

This is general information about a published decision, agency guidance or piece of legislation. It is not legal advice and does not create an attorney-client relationship. Items concerning offices outside the United States are reported for information only and are not counsel on the law of those countries.