A patent agent drafting a set of claims pastes the inventor's disclosure into ChatGPT to tighten the language before filing. It's a Tuesday-afternoon task, done to save an hour. But novelty is a fragile asset: an invention is patentable only while it hasn't been disclosed, and the moment that disclosure leaves the firm's control, the question of whether secrecy still holds is no longer entirely the firm's to answer.
This is the specific exposure that patent prosecution work carries and that general "AI safety" advice misses. The risk isn't a fabricated citation or a wrong answer. It's that the raw material of a patent application — a description of something not yet public, not yet filed — becomes subject to a third party's terms of use the instant it's submitted. And unlike a filed application, a pre-filing invention has no protection to fall back on once it's out.
Why pre-filing is the one moment that can't be undone
Patentability turns on novelty. Most systems require that an invention not have been disclosed to the public before the priority date, and even jurisdictions with a grace period treat inventor-controlled disclosure very differently from disclosure into a system whose retention and reuse the firm doesn't govern. Once a description of an unfiled invention is pasted into a public LLM, it becomes subject to the provider's terms — which can grant broad rights to retain the content, process it through sub-processors, and use it to train the provider's models.
The hard part is irreversibility. After submission, there is nothing left to control. You cannot recall the text, audit where copies went, or prove it was never surfaced elsewhere. For a filed application that's an inconvenience; for a pre-filing disclosure it can be the difference between a patentable invention and one whose novelty is now in question.
The Samsung episode from 2023 is the plainest version of the pattern: within roughly twenty days of allowing ChatGPT internally, engineers had pasted source code and a defect-detection algorithm into it, and the data couldn't be recalled. The relevant detail for a patent practice isn't the scale — it's the speed. The gap between "we're just experimenting" and "we have an irreversible disclosure problem" was measured in weeks.
The courts are starting to treat AI submission as voluntary disclosure
Early decisions are directional, not settled law, but they point in a consistent and uncomfortable direction. In Trinidad v. OpenAI (N.D. Cal., Jan 2026), a trade-secret claim was dismissed because developing the alleged secrets through ChatGPT counted as voluntary disclosure — the secrecy that the claim depended on had already been lost. In U.S. v. Heppner (S.D.N.Y., Feb 2026), documents created with a public GenAI tool were held not protected by attorney-client privilege.
Read those together from a prosecution desk. A trade secret and a pre-filing invention are both protected by the same thing: they haven't been disclosed. If a court is prepared to treat submission into a public tool as the act that ends secrecy, the same logic threatens the novelty an application is built on. No firm should want to litigate whether pasting an invention into a chatbot started the clock — the safer position is to make sure the invention never reaches the tool in the first place.
This is why the framing that resonates with practitioners is due diligence, not cybersecurity. The duties already exist: confidentiality to the client, and the newer duty of technological competence that expects a lawyer or agent to understand the tools they use. Keeping an unfiled invention out of a public LLM is simply what doing your due diligence looks like when the tool in question retains what you feed it.
Bans don't survive contact with a busy prosecution practice
The reflex — prohibit AI tools — fails for the same reason it fails everywhere. Gartner's 2026 survey found that 88% of employees with enterprise AI access also use personal AI tools for work, and Cyberhaven reported that 82.8% of legal documents entered into AI tools go to non-corporate accounts. A policy that says "don't use ChatGPT for client work" doesn't stop the associate drafting claims at 9 p.m.; it just moves the disclosure onto a personal account where the firm has no visibility at all.
Consider how the exposure actually happens. A junior agent uses AI to rephrase a clunky background section — pasting in the full inventor disclosure for context. A paralegal summarizes a batch of invention-disclosure forms ahead of a filing deadline. A partner uploads a draft specification to check claim consistency. None of these people are being reckless; they're being efficient under deadline pressure, which is exactly the condition under which pre-filing material leaks.
What works is governance at the level of the data, not the tool. That means visibility into where risky AI use concentrates across the practice, sensitive content caught and redacted before a prompt reaches the tool, and in-the-moment explanation so the person learns why a particular paste was stopped. A partner who can see policy events across their staff — the fact that a risky prompt was caught, never the contents of what someone typed — has both a supervisory control and, should it ever matter, a record that the firm took reasonable steps to prevent disclosure.
This is the principle Sanitized AI is built on: act before submission, because after submission there is nothing left to protect. For pre-filing inventions, that ordering isn't a nicety — it's the whole game.
What to check before your next filing cycle
One concrete step this quarter: ask how invention disclosures actually move through your firm before a priority date is secured. Not the policy — the practice. Where does the raw description of an unfiled invention travel, who touches it, and which of those steps could involve a public AI tool on a personal account you can't see? If you can't answer that with confidence, the novelty you're trying to protect is running on trust rather than control.
If you'd like to see how catching sensitive prompt data before it leaves the firm works in a prosecution workflow, request a demo.