The Prompt Is Not the Proof: Inventorship, Privilege, and Prior Art When AI Does the Thinking

In One Weekend, a High-School Dropout Using AI Shook the Math World

For about fifty years, mathematicians inched the proven share of the Riemann zeta function’s nontrivial zeros on the critical line from 33.3% to 41.6%. In a result announced in August 2026, an AI model raised it to 67.2% in about two days of work, after 650 failed ideas (Wall St. J., Aug. 14, 2026).

The model was run by a software engineer who works on AI coding tools at the company that built it. He is a high-school dropout whose formal math education ended with high-school geometry. After his first prompt, his input was mostly encouragement: “keep going,” “believe in yourself.” His summary of the result: “Still not sure what that means.” The paper lists the model as its author and calls the engineer “in every meaningful sense the paper’s human co-author.” However, being called a co-author does not, by itself, establish that he conceived of the subject matter described in the paper.

Now make the target a compound that superconducts at room temperature. Same engineer, same prompts. Under current U.S. Patent and Trademark Office (USPTO) guidance, it is hard to tell who the inventor would be. Every claim of a U.S. patent needs a human who conceived it.

As scientists and researchers rely more on AI to plan and interpret their experiments, companies face three risks to patent value: inventorship that cannot be proven, privilege that may not survive, and disclosures that may become prior art. They can reduce all three by documenting human conception and keeping AI work confidential. The recent SLW webinar Practical AI for Chemical Patent Practice covered these issues.

The Human Share of Invention Is Shrinking

As of this writing, a multi-agent AI system has generated the hypotheses behind a published drug-repurposing discovery for dry age-related macular degeneration (Ghareeb et al., Nature 655, 497 (2026)). Humans ran the bench experiments. Another AI system independently proposed a bacterial gene-transfer mechanism that a research team had spent years establishing but had not yet published (Penadés et al., Cell (2025)).

In chemistry, generative models now propose new structures. Retrosynthesis tools propose ways to make them. Those are the two things the law expects a human to have in mind to conceive a new compound.

The field is heading toward artificial general intelligence (AGI), which forecasters on Metaculus, as of September 2026, expect in the early 2030s. Beyond AGI lies artificial superintelligence (ASI): systems that outperform the best human experts and may produce results no person could have conceived unaided. Whatever the date, each gain in capability shifts more of the inventive work to the model, and the human share shrinks toward the prompt. Yet the USPTO has stated that “[t]here is no separate or modified standard for AI-assisted inventions” (Revised Inventorship Guidance for AI-Assisted Inventions, 90 Fed. Reg. 54636 (Nov. 28, 2025)).

U.S. Law Still Requires a Human Who Conceived Each Claim

Only natural persons can be inventors (Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022)). The USPTO’s November 2025 guidance applies the same conception test courts have long used. Conception requires “a definite and permanent idea of the complete and operative invention.” The idea must be “a specific, settled idea,” not “just a general goal or research plan” (90 Fed. Reg. at 54636).

The guidance treats AI systems as instruments, “analogous to laboratory equipment, computer software, research databases, or any other tool that assists in the inventive process” (id. at 54637). Inventors may use AI like any other lab tool, and the AI never becomes a co-inventor. But the framework protects only a person who actually conceives the invention. It does not supply a conception the person never had. The question is whether a human conceived the complete invention, and courts apply that test claim by claim.

When a model proposes the structure of a new compound, a chemist’s case for conception rests on what came before (the hypothesis and design constraints) and what came after (selection and modification).

Inventorship Defects Often Stay Hidden Until After a Patent Issues

The USPTO presumes that the named inventors are correct and does not routinely investigate inventorship when a patent application is under examination (90 Fed. Reg. at 54637). Instead, the issue is most often brought up later, when an accused infringer raises it as a defense to a charge of patent infringement. As AI-assisted research has grown, many applications filed since the AI boom began in earnest may be built on a weak inventorship foundation.

In April 2026, the Federal Circuit held that “a patent which incorrectly lists its inventor(s) and cannot be corrected according to law is invalid” (Fortress Iron, LP v. Digger Specialties, Inc., No. 2024-2313 (Fed. Cir. Apr. 2, 2026)). No court has yet decided a case in which a challenger argued that an AI, rather than the named inventor, conceived the claimed invention. But existing rules make the risk easy to see. A missing human inventor can usually be added to a patent. An AI cannot be named as an inventor at all. So a claim that a model, rather than a person, conceived may have no fix. A challenger would look for evidence in AI chat histories, in depositions about who proposed what, and in marketing that calls a product “AI-designed.”

The duty of candor adds a second risk. Under separate 2024 USPTO guidance on AI tools, which the Office has not withdrawn, anyone who owes that duty and knows a claim lacked a significant human contribution must disclose it (89 Fed. Reg. 25609, 25614–15 (Apr. 11, 2024)). A federal court has held that knowingly omitting inventors made the patents at issue, and related patents, unenforceable (Inline Plastics Corp. v. Lacerta Group, LLC, No. 18-cv-11631 (D. Mass. Nov. 13, 2025)).

Either way, a weak inventorship foundation usually stays hidden until the patent is challenged in litigation or closely scrutinized during licensing or diligence.

Why Prompts Alone Fall Short of Proving Conception

The USPTO’s guidance names the evidentiary problem. Without a description of the invention “with particularity,” it says, “an inventor cannot objectively prove possession of a complete mental picture of the invention at a later time” (90 Fed. Reg. at 54637).

Many prompts state goals, such as “find a selective inhibitor with good oral bioavailability.” A goal like that is a “general goal or research plan,” not conception. The Riemann prompts are the extreme case. The U.S. Copyright Office reached a similar conclusion about authorship: prompts alone generally do not make a user the author of AI output (U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (Jan. 2025)).

AI logs can help or hurt. The Federal Circuit has treated server-generated email metadata as corroboration independent of the inventor (Global Health Solutions LLC v. Selner, No. 2023-2009 (Fed. Cir. Aug. 26, 2025)). Records kept by an AI platform could do the same. But a log shows only what was typed. It may also show that the machine, not the inventor, supplied the answer. If a model proposes 400 candidate compounds and a chemist selects nine and tests three, proof of human conception depends on why the chemist chose those nine. Unless that reasoning is written down, it may be lost.

How to Create and Store Invention Records

Records should be made at three points. Before prompting, write down the hypothesis, the target structure or design, the constraints, and the selection criteria. During the session, note which model was used and keep exact records of prompts and outputs. Keep records of which suggestions were accepted or rejected and why. Record every human change to the model’s output. For agents that run for hours or days, schedule human checkpoints. Before filing, review each claim to confirm who conceived each element.

These records should carry system-generated timestamps, and access to them should be restricted so that raw data cannot be edited after it is created. Keep them for the life of the patent family. Write them on the assumption that they will someday be relied upon as evidence of inventorship in a legal or administrative proceeding.

When the Chat Window Leaks

The same chat window that holds a company’s best evidence of human conception can also leak the invention itself.

Privilege. In February 2026, a federal court considered documents that an executive facing fraud charges had created on his own with a consumer AI tool to analyze his defense. The court held that the documents were protected by neither attorney-client privilege nor the work-product doctrine. One reason was the tool’s privacy policy, which let the provider train on users’ inputs and disclose them to third parties, including government authorities. The court also said that even if some of what he typed came from his lawyers, sharing it with the provider waived privilege. It suggested the result might have been different if his lawyers had directed him to use the tool (United States v. Heppner, 2026 WL 436479 (S.D.N.Y. Feb. 17, 2026)).

Other courts have protected litigants’ own AI work as work product. No court, however, has decided whether a lawyer’s use of a consumer tool affects privilege. That uncertainty matters most in patent prosecution, where communications with counsel are usually protected by privilege alone.

Prior art. No court has decided whether entering an unfiled invention into a public AI tool is a public disclosure. In the United States, a private chat is unlikely to count as a “printed publication.” A one-year grace period also generally covers an inventor’s own disclosures (35 U.S.C. § 102(b)(1)(A)). The risk rises when a chat is shared by link; in 2025, search engines indexed hundreds of thousands of shared chatbot conversations (Fast Company, July 2025; Forbes, Aug. 2025). The risk also rises when filing is delayed more than a year or when protection is sought abroad. Chemical and pharmaceutical portfolios are usually global. Many important markets, including Europe, offer no general grace period.

Trade secrets. A federal court has held that material a user developed in a consumer chatbot under its standard terms was not a trade secret, because using the tool disclosed it to the provider (Trinidad v. OpenAI Inc., No. 25-cv-06328-JST (N.D. Cal. Jan. 5, 2026)).

Candor. A private chat usually does not need to be reported to the USPTO. But a chat that became public more than a year before filing may be material prior art that must be disclosed.

Steps Companies Can Take Now

Companies using AI in research and development should:

1. Use only enterprise AI tools for unfiled inventions and confidential legal matters. The contracts should bar training on the company’s data, bar disclosure to third parties, and allow deletion. Keep the contracts on file and prohibit consumer accounts for this work.
2. Require conception records before and during AI-assisted work.
3. Before filing, ask inventors how they used AI, and review inventorship claim by claim.
4. Review press releases, investor materials, and patent specifications so they do not describe the AI as the inventor.
5. When litigation is anticipated, have counsel direct any AI use in writing. Include AI accounts in litigation holds.

Getting Ready to Answer “Did Any Human Conceive This?”

As AI approaches AGI and beyond, the question in patent disputes is likely to shift from “Did AI help?” to “Did any human conceive this?” Companies that build human conception into their workflows and records, and keep their AI use confidential, will be able to answer it. Where no person conceived an AI-originated result, trade-secret protection may be the realistic alternative. It too depends on using enterprise AI tools. The full webinar, which also covers AI-assisted Markush claim drafting and prompting techniques, is available here.

Nicholas P. Lanzatella, Ph.D., J.D., is a registered patent attorney at Schwegman Lundberg & Woessner whose work focuses on chemical engineering, materials science, and organic chemistry. He holds a Ph.D. in organic chemistry and a B.S. in chemical engineering.

This article is provided for general informational purposes and does not constitute legal advice.