Key Takeaway: As AI-drafted invention disclosures become more common, counsel are reading past the polished prose, and thinking harder about privilege, and confidentiality.
If invention disclosures sound different lately, look to the writer, not the research. Disclosures now reach the attorney polished, dense, and confident. They are often structurally complete, but materially harder for counsel to triage before drafting.
An AI tool drafting from an inventor’s notes tends to generalize away the reasoning that may have carried the most weight: why a structure works, which substitutions failed, what the inventor tried first. Alternative embodiments and working examples may be missed. Materials and citations are sometimes created outright. Often important details have been stripped out from the original disclosure; those important details that counsel may have otherwise relied on to locate the point of novelty. This may leave counsel reverse-engineering the document to recover an inventive concept and identify risks from the AI-drafted disclosure.
Beyond the drafting challenges, AI generated invention disclosures may also carry risks of losing attorney-client privilege or confidentiality.
Privilege. In United States v. Heppner, Judge Rakoff held that a defendant’s exchanges with a public AI platform were protected by neither attorney-client privilege nor the work product doctrine.[1] Privilege failed in this case because the platform was not an attorney, thus disclosure to the platform waived confidentiality in the inputs, and later routing the output through counsel could not confer privilege retroactively. Work product failed on a separate ground: the defendant acted on his own initiative, not at counsel’s direction. One district court’s ruling, however, is not a per se bar. Whether attorney-directed use of publicly available AI tools is protected remains an open question, and therefore both prompts and outputs may be discoverable as the courts continue to wrestle with these ideas.
Confidentiality. Pasting a full disclosure, structures, or assay data into a publicly available AI tool may undercut the reasonable-measures element of trade secret status and, depending on the platform’s retention and training terms, may qualify as a disclosure that counts against novelty. Both the United States and Canada offer a general twelve-month grace period; Europe and China do not have a grace period, and their narrow exceptions (evident abuse and recognized exhibitions) may not rescue such a prompt.[2]
To avoid some of the risks identified above[3], biopharma companies should build a few additional questions into intake of invention disclosures:
- Ask for the raw version. Request unedited notes or lab data alongside any polished draft. Important inventive concepts may live in the unfiltered version.
- Treat fluency as a flag. These tools produce authoritative, citation-rich conclusions regardless of accuracy.
- Separate AI validation from novelty. An assertion of novelty is a hypothesis, not a substitute for patentability and freedom-to-operate searching.
- Watch for over-abstraction. AI can strip the concrete inputs and specific steps that distinguish eligible subject matter from a natural law or abstract idea.
- Run the question in reverse. If a disclosure sounds patentable but the inventor cannot articulate their own contribution to conceiving it, pause. Find out where the gap is coming from. AI is a tool; only a natural person can be an inventor[4].
Editor: Brenden S. Gingrich, Ph.D.
[1] United States v. Heppner, No. 25-cr-00503-JSR (S.D.N.Y.) (bench ruling Feb. 10, 2026; written mem. Feb. 17, 2026)
[2] 35 U.S.C. § 102(b)(1); Patent Act, R.S.C. 1985, c. P-4, s. 28.2(1)(a) (Canada); EPC arts. 54–55; Patent Law of the P.R.C. art. 24
[3] Once the record is straight can the protection question be answered and it is a genuine fork. A patent is a bargain: enabling disclosure in exchange for a time-limited right to exclude. A trade secret demands the reverse: confidentiality with no fixed term, but protection that ends the moment the information escapes. Which route fits turns on the technical detail that reaches counsel, the very detail an AI draft may have flattened, or already disclosed.
[4] The best control is upstream: build these questions into the intake form, treat confidential data entered into a public tool as potentially compromised, and enforce internal AI policies accordingly. The fastest way to separate real innovation from confident-sounding output remains the oldest one — talk to the scientist who did the work.
Subscribe for News and Updates
