USPTO Pushes AI Patents Toward Measurable Technical Improvement
The USPTO’s recent memorandum on Rule 132 Subject Matter Eligibility Declarations puts a sharper lens on a familiar but often underdeveloped argument in AI, software and biotechnology patent prosecution: whether the claimed invention improves technology rather than merely using a computer to reach a desired result. The agency is not creating a separate patentability regime for AI. It is asking applicants to connect eligibility arguments to technical facts already disclosed in the application.
For applicants, the practical message is direct. Describing an AI model that classifies, predicts, generates or supports a decision will rarely be enough on its own. The specification and any supporting declaration should explain how the invention makes a computer system, data-processing workflow, training process or diagnostic procedure faster, cheaper, more stable or more efficient.
Technical improvement is becoming a fact record problem
The most useful part of the memorandum is not a new slogan about AI. It is the insistence that “technological improvement” should be supported by facts that a person of ordinary skill in the art could recognise from the specification. A Rule 132 declaration may help explain how the disclosure would be understood, but it cannot rescue a thin specification by adding the missing technical contribution after filing.
This changes the drafting sequence. In many AI and software cases, applicants first describe inputs, a model, outputs and a commercial use case, then try to add “improved efficiency” or “better accuracy” during prosecution. That approach is less comfortable when the eligibility argument turns on evidence. A stronger filing connects the technical problem, the system bottleneck, the claimed mechanism and the observable improvement before examination starts.
Using AI is not the same as improving technology
The weak version of an AI patent application says, in effect, that a model performs a task. It identifies images, predicts risk, ranks candidates or supports a diagnosis. If the claim and specification remain at the level of a business outcome, mathematical processing or generic computer implementation, the application may still face a subject-matter eligibility rejection.
The stronger version explains what changed at the technical level. Did the model architecture reduce inference latency? Did the training pipeline lower compute cost? Did data pre-processing improve stability under noisy inputs? Did deployment reduce storage, bandwidth or device requirements? In biotechnology and diagnostics, the same discipline matters. Saying that AI improves diagnostic accuracy is helpful but incomplete; explaining how feature selection, signal correction, sample handling or computation reduces false positives or testing time gives the argument more substance.
Declarations should stay focused on Section 101
The memorandum also points applicants toward cleaner evidentiary practice. A declaration addressing subject-matter eligibility should normally be kept separate from declarations on novelty, obviousness, written description or enablement. The reason is practical. Different rejections require different facts. When eligibility, obviousness and disclosure arguments are folded into one document, the record can become harder for the examiner to evaluate and harder for the applicant to rely on later.
A useful SMED does not simply state that the claim is patent eligible. It explains the relationship between the claim limitations and the technical disclosure, then shows why that relationship reflects a real-world technical application or improvement. Evidence of market demand, commercial success or user preference may have value elsewhere, but it is less persuasive under Section 101 unless it is tied back to the technical limitation that matters.
Drafting work needs to move earlier
For AI, software and biotech applicants planning US filings, the safest adjustment is to build a “technical improvement file” before the application is drafted. It does not need to look like an academic paper, but it should identify the prior technical constraint, the engineering step taken by the invention and the measurable or at least clearly explained effect. Quantified results are useful. A well-described causal chain is still valuable when hard numbers are not available.
Claims should carry some of that work. If the independent claim is drafted only as broad functional language, while the technical mechanism remains buried in the description, a later declaration has limited room to operate. Applicants should consider placing at least one or two system-level or process-level improvement features into the claims, with dependent claims covering alternative implementations, performance indicators and deployment settings.
Practical takeaways for applicants
Applicants should begin collecting evidence beyond accuracy. Latency, memory use, training time, labelling cost, hardware compatibility, error rates and test duration can all support a more concrete technical improvement story. In some cases, these metrics may matter more than a marginal increase in prediction accuracy.
They should also avoid treating a SMED as a legal brief. The more useful material is technical explanation from the perspective of skilled practitioners, supported by the original disclosure. Cross-border filers should take particular care here. The United States, Europe and China do not use identical eligibility tests, but all three systems tend to reward applications that describe technical effects with precision. AI patent drafting is moving away from “we used AI” and toward “we can show what the technology itself does better.”
