Your Engineers Invent With AI Now. Your Inventorship Records Need to Catch Up.
By Brandon Furdock, Registered Patent Attorney, Chicago Patent Law LLC
Walk any AI company's engineering floor and you'll find the same workflow: models proposing architectures, coding assistants generating implementations, and agents running experiment sweeps overnight. The technology being invented is AI, and increasingly, so is the process of inventing it. Patent law has a settled headline rule for this world, which is that only natural persons can be inventors, and an unsettled operational question underneath it: when a human and an AI system produce an invention together, what must the human have contributed, and how do you prove it?
The legal frame, in brief
U.S. law is clear that an AI system cannot be a named inventor. That question was litigated and closed. The harder question is what a human must contribute when the invention emerges from a human-plus-AI workflow, and here the ground moved recently. In February 2024, the Patent Office answered with a test borrowed from joint-inventorship law: for each claim, some natural person had to have made a significant contribution to conception. In November 2025, the Office rescinded that guidance and replaced it with something simpler (Revised Inventorship Guidance for AI-Assisted Inventions, Nov. 28, 2025): the same inventorship standard applies to every invention, AI-assisted or not. A named inventor must have conceived the invention, a definite and permanent idea of the complete and operative invention, clearly enough in mind to describe it. Using an AI tool disqualifies no one, the tool is treated like any other tool. And the Office continues to presume that the inventors named on an application are the right ones.
Read one way, that is a loosening, and it is. Your examiner is unlikely to ask about AI use, and there is no AI-specific paperwork duty. But notice what did not change. Conception is still a factual question about what was in a human mind, and when it matters, it is still decided on evidence. What changed is where it will matter: less at the Patent Office, more in diligence and litigation.
Why this is a business problem, not an academic one
Inventorship defects are not paperwork errors. Incorrect inventorship can render a patent unenforceable, and deliberate misstatements flirt with inequitable conduct. Now consider who will be probing: an acquirer's diligence team asking pointed questions about AI tool usage (they've started); a litigation opponent deposing your engineers about how much of claim 1 came from a model (they will); a foreign patent office applying its own AI-inventorship rule, which may not match ours, to an application whose facts were established, undocumented, in 2026.
Every one of those encounters goes fine for the company with records and badly for the company with recollections.
The documentation program (it's smaller than you fear)
This does not require a bureaucracy. It requires four habits:
1. An AI-use disclosure line in your invention disclosure form. Which tools were used, for what, at which stage. Inventors fill out disclosure forms anyway; this is two additional fields. Its mere existence proves the company took the question seriously, which is an asset in itself.
2. Conception notes that name the human contribution. When the disclosure describes the invention, it should also describe, in a sentence or two per inventor, what each person conceived, decided, selected, or redesigned, specifically including how they directed or built upon AI-generated material. "The model proposed six candidate architectures; Dr. John Doe recognized that the third could be modified with X to solve Y" is exactly the kind of sentence that decides these questions.
3. Preserved artifacts for crown-jewel inventions. For the filings that matter most, consider keeping the working record: significant prompts and outputs, experiment logs, design documents, the trail from AI suggestion to human-directed final form. Not everything, just the important things.
4. Inventorship analysis by counsel at filing, documented. For AI-assisted inventions, the who-is-an-inventor call should be an explicit, recorded legal judgment, not a default listing of whoever was in the room. Ten minutes per application, and it converts a future deposition topic into a settled record.
The honest case against writing everything down
Documentation advice sounds costless. It isn't, and you should hear the other side before adopting the program.
First, records cut both ways. Everything you keep is discoverable. If the honest record of a crown-jewel invention shows that the model did nearly everything and the humans mostly watched, you have not protected the patent; you have assembled the challenger's exhibit list. Silence, by contrast, leaves an opponent fighting your engineers' testimony and the legal presumption that named inventors are correct. A record only helps if the facts in it help.
Second, the Patent Office no longer asks for any of this. After November 2025 there is no AI-specific documentation duty, and your inventorship is presumed proper. The return on these habits comes almost entirely from two audiences you may never face: an acquirer's diligence team and a litigation opponent. A company that never sells and never asserts is buying insurance that may never pay out.
Third, bad records are worse than none. A disclosure form filled out carelessly by an engineer who wanted to get back to work becomes impeachment material: wrong tool names, overstated AI roles, understated human ones. Adopting the habit means adopting the review that keeps it accurate.
So why recommend the program anyway? Because the record is also a routing device. If contemporaneous notes show real human conception, they are the whole rebuttal to a your-inventors-just-pressed-a-button challenge. If they show the human contribution was thin, far better to learn that at the disclosure meeting, when you can still choose trade secret protection or a defensive publication, than in a deposition after spending six figures asserting the patent. Either way the record earns its keep, it just earns it differently. The rule underneath: keep records where the asset matters and the human contribution is real, keep them accurate, and never keep them for show.
The fine print
Watch the interaction with your patent/trade-secret split. Some of the most sensitive material in these records, meaning prompts, training methodology, and experiment infrastructure, may itself be trade secret. The documentation program may need the same access controls as the secrets it describes.
AI-assisted invention isn't a corner case anymore. The law spent 2024 and 2025 revising what it means and will keep refining it. Your job is simpler: make sure the beneficial facts are written down.
Brandon Furdock is a registered patent attorney and computer engineer (M.S., Northwestern University). Chicago Patent Law helps AI companies build invention-disclosure and inventorship-documentation practices as part of its fractional IP counsel retainers. General information, not legal advice. Attorney Advertising.
Building something worth protecting? A 15-minute intro call costs nothing: furdock@chicagopatent.law.