Software and AI claims must show the improvement
Describe how the technology itself gets better and claim the means, not just the result.
Lesson 5 of 9 · 4 min
An old problem in new clothes
Some subject matter cannot be patented at all, however clever it is. Laws of nature, natural phenomena and abstract ideas are excluded. Alice asks whether a claim is directed to one of those concepts and, if so, what else supplies an "inventive concept", and "wholly generic computer implementation" does not (Alice v. CLS Bank). This question is called patent eligibility.
The logic is older than computers. Morse and Alice share it: claiming a result or a generic implementation claims too much (O'Reilly v. Morse; Alice v. CLS Bank). The Supreme Court made the same point about natural laws: a patent "must do more than simply state the law of nature while adding the words 'apply it'" (Mayo v. Prometheus).
What the courts accept
Claims that improve computer functionality or technology are eligible under the line of cases running through Enfish and McRO (quoted in Ex parte Desjardins). So the question to ask of any software invention is plain: what does the machine do better, and how does it do it?
Two readers, two standards
An examiner and a judge may not reach the same answer, and USPTO guidance does not bind courts (USPTO 2024 AI update). What passes one may fail the other, so no drafting style can promise that a claim will survive.
How to write it
The following habits are our judgment, not a legal test:
- Name the technical problem with the earlier approach, not just the business goal.
- Explain the mechanism: the steps, data flow and structure that produce the gain.
- State the technical effect, such as less memory or fewer operations, if you can show it.
- Claim the means that produce that effect, and do not claim the result alone.
Compare two versions of one idea, in an example of our own. Predicting demand using a machine learning model states an outcome and a generic tool. A claim to a specific way of restructuring a model's training so that it uses less memory says what changed in the technology. The second gives a reader something to hold. If the honest answer is that nothing in the technology changed, then in our judgment you should expect a hard road, whatever the wording.
Timeless wisdom
Claiming a result, or a generic implementation of one, claims too much. That shared logic runs from Morse (O'Reilly v. Morse) to Alice, where "wholly generic computer implementation" did not supply an "inventive concept" (Alice v. CLS Bank). The other half of the principle is positive. Claims that improve computer functionality or technology are eligible under the Enfish and McRO line (quoted in Ex parte Desjardins).
What's current
As of November 6, 2026
As of November 6, 2026, the Office and the court converge on one point: the claim need not literally recite the improvement, but the disclosure must show it (MPEP notice, Dec. 5, 2025; '637 v. Google).
Desjardins credited an ML training method that learned new tasks "whilst protecting knowledge about previous tasks" (decision). Recentive v. Fox (Apr. 18, 2025) holds, narrowly, that generic ML on new data fails "without disclosing improvements to the machine learning models" (opinion). In our judgment, the two reconcile as improving a model versus applying a generic one, and drafters should write for the stricter court. Allowance does not predict survival in court.
Abroad, the EPC excludes mathematical methods only "as such" (EPC Art. 52). Commentary reports that China's 2026 Guidelines add AI ethics grounds (China IP Law Update).
Key takeaways
- Claiming a result or a generic implementation claims too much.
- Explain what the technology does better and how.
- USPTO guidance does not bind courts, so write for the stricter reader.
- No drafting style can promise that a claim survives.
Check your understanding
Describe your technical improvement
Sources
- Alice v. CLS Bank (U.S.)
- O'Reilly v. Morse (U.S.)
- Mayo v. Prometheus (U.S.)
- Ex parte Desjardins (PTAB Appeals Review Panel)
- USPTO 2024 AI subject matter eligibility update (Federal Register, July 17, 2024)
- USPTO MPEP notice on Desjardins (Dec. 5, 2025)
- '637 v. Google (Fed. Cir. Jan. 22, 2026)
- Recentive v. Fox (Fed. Cir. Apr. 18, 2025)
- European Patent Convention, Art. 52
- China IP Law Update on CNIPA's revised Guidelines (commentary)
The Academy is general education, not legal advice, and reading it doesn't create an attorney-client relationship. Patent law and fees change, so check the linked sources and have a registered patent attorney or agent review your situation before you file. Obsidyan sells patent software, including AI drafting tools, so weigh what these lessons say about AI tools with that in mind.