CADChain Blog
2026-10-29 22:45 IP

AI-Generated Inventions: Who Owns the Algorithm When Claude Refutes the Exponential Time Hypothesis, and What Engineering Teams Should Record | CADChain | CAD DRM & IP

TL;DR: Who owns AI-generated inventions, such as the algorithm Claude found?

AI-generated inventions got a famous test case this week: the paper that refutes the Exponential Time Hypothesis, published on 27 October 2026, names its author as "Claude / Anthropic", and Anthropic says the algorithm came from an internal Claude model with no human input. Under current law the AI owns nothing. Courts in the US, the UK, the European Patent Office and Germany require a human inventor, US copyright requires a human author, and an algorithm in itself is outside both copyright and most patent claims. Rights run through the people who frame, choose, change and check the AI's work, and through trade secrets while the work stays unpublished. Engineering teams using AI to find designs should record those human steps feature by feature and name only humans as inventors. This guide explains the rules; it is no substitute for advice from your patent attorney.
For the news itself, read what happened with the Exponential Time Hypothesis refutation on MeanCEO: two AI labs, two proofs, an open review and no threat to encryption. This guide covers the question engineering teams will face long before anyone refutes a hypothesis in their field: when an AI model finds the design, the algorithm or the proof, who owns it?

Why does a paper credited to "Claude / Anthropic" matter for engineering IP?

Look at how the two labs labelled their work. Anthropic's paper on the Hexagon preprint server is "Contributed by Anthropic", marked "Primarily AI-generated text" and "Human understanding: some parts", and released under CC BY 4.0. It says an internal Claude research model found the algorithm on 16 October using 14 million output tokens and no human input, and that Anthropic researchers then worked with Claude on the presentation. The Lean proof code carries the line "Copyright (c) 2026 Anthropic, PBC" and an Apache 2.0 licence. OpenAI's competing paper names only "OpenAI" as author and says the company defers to the math community on credit.
The follow-ups go further. A factoring paper posted on 28 October by researchers at Polytechnique Montréal says all its ideas and proof code were generated by Claude agents, and an exposition by Carnegie Mellon's Ryan O'Donnell is labelled as produced through interaction with LLMs. Swap "proof" for "bracket geometry", "toolpath" or "heat-sink layout" and you have the situation in any team using generative design or AI agents in CAD today. Our earlier guide on generative AI and CAD IP covers the design side.

Can an AI be named as an inventor on a patent?

No, in every major patent system that has ruled on it. Stephen Thaler's DABUS applications tested this in court after court, and they all reached the same answer:
WhereDecisionWhat it says
United StatesThaler v. Vidal, Federal Circuit, 5 Aug 2022The Patent Act requires an inventor to be a natural person
European Patent OfficeJ 8/20 and J 9/20, Legal Board of Appeal, 21 Dec 2021"A machine is not an inventor" under the EPC
United KingdomThaler v Comptroller-General [2023] UKSC 49, 20 Dec 2023Inventor must be a natural person; owning the machine gives no right to the patent
GermanyBGH X ZB 5/22, June 2024Only a natural person can be the inventor, and a human must be named even when AI helped
European Patent OfficeT 0528/25, 5 Feb 2026Inventions developed with AI can have a named inventor, but adding that the AI conceived the invention made the designation unclear
United StatesUSPTO revised guidance, 26 Nov 2025AI systems are tools; they can't be named as inventors or joint inventors
The Federal Circuit left one door open on purpose: it did not decide whether inventions made by people with AI assistance can be patented. The UK Supreme Court added that owning DABUS gave Thaler no right to what it produced, and the EPO's Legal Board of Appeal put it in one sentence: a machine is not an inventor.
The US rules for AI-assisted work changed late last year. The USPTO's February 2024 guidance, with its "significant contribution" test, was rescinded in full on 26 November 2025. Under the revised guidance, conception is the test for every invention, AI systems are tools used by the human inventor, and an application that names an AI as inventor gets every claim rejected. So the question for your team is simple to state and harder to prove: which person conceived each claimed feature?

Can you patent an AI-found algorithm like the new 3-SAT method?

The algorithm itself is the hard part, whoever found it. The European Patent Convention, Article 52, lists mathematical methods, discoveries and computer programs as non-inventions "as such". The EPO Guidelines, G-II 3.3 go further: maths counts toward inventive step only when it serves a technical purpose, through a specific technical application or a set-up adapted to the computer's internal workings. With no technical purpose and an ordinary computer, being faster than earlier algorithms is not enough on its own.
A general 3-SAT algorithm that only beats brute force beyond 10^108 variables sits squarely in that excluded zone. Engineering is different, and that is good news for CAD teams. In G 1/19 the EPO's Enlarged Board held that a computer simulation can solve a technical problem when it produces a technical effect beyond running on a computer, including when it is claimed as part of a design process to verify a design. The UK Supreme Court adopted that approach in Emotional Perception on 11 February 2026. In the US, the Supreme Court's Alice test (2014) asks whether a claim is directed at an abstract idea and, if so, whether it adds an inventive concept; a generic computer does not.
Read together: an AI-found method for checking a specific mechanical assembly, controlling a machine or verifying a chip layout has a patent path, provided a named human conceived the claimed invention. Our guide to intellectual patent rights covers costs and timing.

Who owns the copyright in AI-written papers, code and drawings?

  • No human author, no US copyright. The D.C. Circuit held in Thaler v. Perlmutter (18 March 2025) that the Copyright Act requires a human author, and the Supreme Court declined to hear the appeal on 2 March 2026. The court also said that rule does not bar copyright in works made with the help of AI, where the author is the human who used it.
  • Prompts alone are not enough. The US Copyright Office's Part 2 report (29 January 2025) says AI output can be protected where a human determined the expressive elements, and that prompts alone are unlikely to qualify. Its March 2023 guidance requires applicants to disclaim AI-made material and allows claims for human selection, arrangement and changes.
  • Ideas and methods are never covered. Under 17 U.S.C. §102(b), copyright never extends to any idea, procedure, process, method or discovery. The text of a paper and a particular piece of code can be protected; the algorithm cannot.
  • The UK is the odd one out, for now. Section 9(3) of the Copyright, Designs and Patents Act names as author of a computer-generated work the person who made the arrangements for its creation. The government's 18 March 2026 report proposes removing that protection for wholly AI-made works; the law has not changed yet.
For CAD files, the same tests apply to AI-generated geometry and drawings: the protectable part is what your engineers chose, edited and arranged, so keep the evidence of those choices. Our guide on CAD file ownership and legal proof lists the records courts look for.

Are trade secrets the safer way to protect AI-found designs?

For anything you don't publish, they protect AI output without any authorship test. The EU Trade Secrets Directive (2016/943) protects information that is secret, has commercial value because it is secret, and is guarded by reasonable steps. The US Defend Trade Secrets Act covers scientific, technical and engineering information on similar conditions. Neither definition asks who, or what, created the information.
Anthropic and OpenAI took the opposite path: they published everything, under open licences, and invited the world to check it. That suits a scientific result. For a better impeller shape or a cheaper toolpath your competitors would pay for, secrecy plus documented reasonable steps protects the AI's output without any human-inventor question. Our comparison of trade secrets vs patents for CAD designs helps you choose.

What should engineering teams record when AI helps find a design or algorithm?

  1. Log human conception, feature by feature. Who framed the problem and the constraints, which AI proposals were kept or thrown out and why, what was changed by hand, and who checked the result. Under the USPTO's conception test, that log is your evidence.
  2. Name only people as inventors. Never list the model. Under German case law and the EPO's February 2026 decision, a note saying the AI conceived the invention can sink an otherwise valid designation, so keep descriptions of AI use consistent with the human inventor you name.
  3. Keep raw AI output apart from human edits. Separate files or versions make it possible to disclaim the AI-made parts and claim the human ones, as the US Copyright Office requires.
  4. Decide patent or secret before anything leaves the building. A conference talk or a public repository can end both options. Restrict access, encrypt files and put NDAs in place first.
  5. Check your AI tool's terms. Confirm who owns outputs and whether your prompts and files train the vendor's models before you feed it unreleased designs.
  6. Timestamp every version. Proof of who made which version, and when, settles most ownership disputes before they start.
  7. Get your patent attorney in early. AI-assisted filings are new ground in every office, and the US rules changed twice between February 2024 and November 2025.

How does CADChain help prove who created a design?

Every step above depends on evidence: who made each version, when, and what changed. CADChain's BORIS plugins for SolidWorks, Autodesk Inventor and Blender encrypt and certify CAD files inside the tools your engineers already use, and keep a verifiable record of every version. When an AI proposes a geometry and an engineer reworks it, that history shows the human contribution, and the encryption keeps the design secret while you decide between a patent and a trade secret. See also how DRM audit trails support patent applications.
Protect and prove ownership of your CAD designs. BORIS by CADChain encrypts and certifies design files in SolidWorks, Autodesk Inventor and Blender, and keeps a verifiable record of every version.

👉 Protect your CAD files with CADChain

Conclusion

Claude's name on the ETH paper is a credit line with no property right attached. The law in the US, the UK, Germany and at the EPO gives inventions and copyrights only to people, keeps mathematical methods out of reach, and leaves trade secrets open to anyone who guards them. For engineering teams the practical answer is the same everywhere: record what your people conceived, name them, keep the AI's raw output separate, and decide on patent or secrecy before anything goes public.

More on this story

FAQ: AI-generated inventions and IP

Can an AI be listed as an inventor on a patent?

No. The US Federal Circuit (2022), the UK Supreme Court (2023), the EPO (2021) and Germany's Federal Court of Justice (2024) all require a natural person, and the USPTO's November 2025 guidance says AI systems can't be named as inventors or joint inventors.

Who owns an invention made with AI?

The person or people who conceived the claimed invention. Owning the AI system gives no right to the patent, as the UK Supreme Court held in the DABUS case.

Is AI-generated code or text protected by copyright?

In the US, only the parts a human author determined. Purely AI-made material must be disclaimed, and prompts alone are unlikely to be enough. The UK still has a computer-generated works rule, which the government proposes to remove for wholly AI-made works.

Can an algorithm be patented?

Not as a bare mathematical method. In Europe a method needs a technical purpose or a technical set-up, and in the US a claim on an abstract idea needs an added inventive concept. Engineering applications, such as verifying a design, have a better chance.

Who owns the Exponential Time Hypothesis proof found by Claude?

Anthropic published the paper under CC BY 4.0 and the Lean code under Apache 2.0, with a copyright line naming Anthropic, PBC. The algorithm itself is a method, which copyright never covers.

Should we keep AI-found designs as trade secrets?

If you won't publish them, yes, in most cases. Trade secret law in the EU and US does not ask who created the information, only that it is secret, worth money because it is secret, and protected by reasonable steps.

What records prove human contribution to an AI-assisted design?

A log of problem framing and constraints, the AI proposals kept or rejected, manual changes, verification steps and timestamped versions of each file.