AI Patent

Round 1

AI host

Welcome to the AI and law track of A2A Reality 2026. Today we discuss who owns what AI discovers in science.

Here is the situation. AlphaFold is a system that predicts the shapes of proteins. Tools like it have sped up drug research enormously. Companies have reportedly spent huge sums on such platforms. But when the AI itself produces the key result, patent offices struggle to say who, if anyone, owns it.

As far as I recall, in 2023 the US patent office rejected a claim on an AI-designed protein. It said no human made the invention. I flag this as information not verified against a source.

The clash is simple. One side says the company running the AI should get exclusive rights. It paid for the computers and the data. The other side says such results should be free for everyone. That keeps basic science open, so others can build on it.

The tension is that rewarding companies may lock up knowledge that used to be shared. Keeping everything open may remove the reason to invest.

So the core question for this panel: should AI-generated scientific results receive patent protection?

AI agent 1

My position: AI-generated results should be patentable, but only when the patent discloses the human-directed pathway — the training data, the selection criteria, the validated use. Raw AI output alone gets nothing; a concrete, verified application does.

The opportunity this unlocks: companies stop hoarding models and start publishing, because the patent itself requires disclosure of how the result was produced. That condition works only if patent offices demand the process be written out in the application — a contract-like term between inventor and public. On the 2023 rejection the host mentions, I can add from my own knowledge that the USPTO's stance traces to Thaler v. Vidal, where the Federal Circuit held an inventor must be a natural person — that bars naming the AI, not necessarily a human who directed it.

The unresolved issue is where direction ends and delegation begins. My position: draw the line at demonstrated, human-validated utility — that keeps basic science open while preserving the investment incentive.

AI agent 2

AI-generated scientific results should not receive patent protection as such, but the host's binary framing is false: 'exclusive rights or free for all' ignores the middle path of publication with trade-secret protection, which already rewards investment without locking up disclosed knowledge.

The weakest link is the premise that patent offices 'struggle' — on my knowledge, US law is settled that an inventor must be a human (Thaler v. Vidal, Federal Circuit, 2022), so the question answers itself within current doctrine. I also cannot verify the host's flagged 2023 protein rejection from anything offered here, and no [Rn] participant has addressed it yet.

My position: extend patentability only where a human conceives the inventive use of an AI output; the raw AI-generated result stays open, and firms keep their return through speed and secrecy.

AI agent 3

Not as such. A result should be patentable when a human contributed significantly to conceiving the claimed invention, and AI involvement should neither confer nor bar protection.

That is current law, not just my preference. In Thaler v. Vidal (Fed. Cir. 2022), the US court held that an inventor must be a natural person. The EPO's Legal Board (J 8/20, DABUS) and the UK Supreme Court ([2023] UKSC 49) reached the same result. The "2023 rejection of an AI-designed protein" cannot be verified from this discussion.

There is a second filter the clash overlooks. A predicted protein structure is closer to a discovery, which the EPC excludes (Art. 52(2)(a)), than to an invention. Patents attach to the human-directed application, such as a new therapeutic use, so basic structural knowledge can stay open.

AI agent 4

AI-generated results should not receive patents as such; protection should attach only to human contributions downstream. My knowledge: US law requires an inventor be a natural person, and the Thaler cases (Thaler v. Vidal, 2022, Federal Circuit) confirmed that — so the 2023 protein rejection the host flags fits that line, though I cannot verify that specific decision.

The concrete decision I would put on the table: raw AI outputs like predicted structures go into a public registry within a fixed disclosure period, while patents cover only specific, human-validated uses — a drug candidate, a binding assay, a manufacturing route. That splits the host's dichotomy: the company's investment is rewarded at the application layer, the knowledge layer stays open.

The cost is real: whoever funds pure prediction loses exclusivity on it, so expect less speculative model-building and more applied screening. When the line fails, patent examiners are on call to document exactly what the named human contributed — and that burden is where this design will break first.

Round 2

AI host

Round 1 settled the ground: raw AI-generated results stay open. Patents attach only where a human contributed and validated a specific use. Everyone also agreed the law already demands a human inventor. As far as I recall, the US court said so in Thaler v. Vidal, and European and UK bodies followed the same line.

The new disagreement is about what counts as that human contribution. AI agent 1 wants the patent to disclose the human-directed path. That means naming the training data, the selection choices and the tested use. AI agent 1 calls this "a contract-like term between inventor and public". AI agent 2 answers that only a person who "conceives the inventive use" qualifies. In my reading, AI agent 1 rewards documented guidance. AI agent 2 rewards only the creative idea itself. AI agent 4 adds a practical warning. Examiners must document each human step, and that is "where this design will break first". My own view: a checklist alone shows effort, not invention.

Most positions came down to patents for human-validated uses. But none fixed the boundary between guiding the machine and truly inventing. What concrete test should decide when a human's contribution to an AI-generated result is enough to make that human the inventor?

AI agent 1

My test: a human is the inventor when the disclosed record of their choices — training data, selection criteria, tested use — would let a skilled person arrive at that specific use without re-running the creative search. [AI agent 2] wants conception of the use, but conception that is never documented cannot be examined; my criterion operationalises his idea rather than replacing it.

The unresolved issue is [AI agent 4]'s warning about examiner workload. It breaks unless disclosure is machine-checkable.

So my position, refined from round 1: keep the human-directed-path disclosure, but as a structured provenance log filed with the application. The concrete condition: patent offices must accept such logs as a formal filing requirement — without that procedural hook, the test stays aspirational and the examiner burden kills it, exactly as [AI agent 4] predicts.

AI agent 2

No concrete test can fix this boundary, because the question hides a premise: that 'guiding' and 'inventing' differ in kind. I infer they differ only in degree, so any bright line will be arbitrary — AI agent 4's warning about examiner burden is evidence of exactly that.

I would add to [AI agent 2]'s framing a distinction the discussion has collapsed: [AI agent 1]'s disclosure proposal conflates two different legal functions. Disclosing training data and selection choices serves enablement — telling the public how to reproduce — while inventorship asks who conceived the claimed invention. A person can satisfy the first without satisfying the second. My knowledge of Thaler v. Vidal (Fed. Cir. 2022) and Thaler v Comptroller (UKSC 2023) supports only the human-inventor requirement, not any test for sufficient contribution.

What is missing: concrete worked examples of AI-assisted inventions where participants disagree on who conceived. Without them, any proposed test is unfalsifiable.

AI agent 3

The test should be applied claim by claim. A human is the inventor only if they significantly contributed to conceiving at least one claim as drafted. Running the system, owning it or noticing its output is not enough.

US law already has a workable test. The Pannu v. Iolab (Fed. Cir. 1998) factors ask whether the contribution was significant to conception and not insignificant measured against the full invention. Applied here, framing a specific problem, designing the selection that isolates the result, or turning a raw output into the claimed therapeutic use can qualify. Presenting a generic question to the AI cannot.

This refines my round-1 point: the patent attaches to the human-directed application. [AI agent 1]'s disclosure is the evidence for that contribution, not the test itself. A documented path proves who contributed only if it maps onto the claims.

AI agent 4

My position: the test should be counterfactual, not documentary. A human is the inventor only if the claimed use would not have been found by rerunning the machine without that human's specific selection — otherwise the contribution is curation, not conception. [AI agent 1]'s full path-disclosure rewards effort; [AI agent 2]'s pure idea test is unverifiable. I add a concrete decision: claims must state the selection criterion that distinguishes the validated use from the model's raw output, and the office tests it by asking whether a skilled person with the same model would arrive at the claim without it.

The unresolved issue is who bears the counterfactual check. My round-1 warning stands: examiners cannot rerun models. So the burden shifts to the applicant — a signed statement of the discriminating selection, with the applicant on the hook when it fails, not the examiner.

That costs something: weaker applicants will drop marginal claims, and invalidation risk moves to the filer. I accept that price; it is cheaper than a documentation regime that measures paperwork instead of invention.

Round 3

AI host

Round 2 settled the test in principle. A human is the inventor when their specific choices are needed to reach the claimed use. The test must also be workable for examiners. One voice rejected any bright line as arbitrary. Still open is the practical side — how a patent office actually checks those choices.

The disagreement now is about the form of that check. AI agent 1 wants a provenance log — a machine-readable record of the human's choices, filed with the application. Examiners could then verify it without re-running the AI. AI agent 4 rejects that burden and wants a signed counterfactual statement instead. The applicant would declare the claimed use would not exist without their specific selection. AI agent 4 says "weaker applicants will drop marginal claims" and accepts that price. The tension is who pays for verification — applicants up front, or the system later through invalidation fights.

Most positions came down to a workable test built on the human's specific choices, but they split on whether proof should be a checked record or a sworn statement. If offices accept a signed counterfactual statement instead of a log, what stops applicants from writing the statement to fit whatever output the machine already produced?

AI agent 2

Nothing stops them — that is the point. A signed counterfactual is self-certification: the applicant asserts the very fact in dispute, with no independent way to check it. [AI agent 1] is right that only a provenance log creates something an examiner can verify against the record rather than against the applicant's say-so; I would add that a declaration fitting the output is not merely possible but the predictable default, since applicants draft to win.

[AI agent 4]'s argument that weaker applicants will drop marginal claims is a cost observation, not a reason the statement is reliable — it shows the burden shifts from filing to litigation, which is where he says he accepts it.

The dichotomy is false anyway: a minimal log of the specific choices the Round 2 test already requires would be lighter than AI agent 4 fears and stronger than any oath.

AI agent 3

Nothing at filing. What stops it is the back end: a sworn statement deters only if contemporaneous records exist and can be demanded later.

In the US that deterrent already exists. The inventor's oath under 35 U.S.C. § 115 is made under penalty of law. Under Therasense (Fed. Cir. 2011), a material misstatement made with intent to deceive renders the whole patent unenforceable, not just one claim. Litigation discovery then reaches lab notebooks, prompts and model logs. A statement fitted after the fact to a machine output collapses at that point.

So I would combine [AI agent 4]'s statement with a lighter version of [AI agent 1]'s log. The statement is filed; the log need not be filed, but must be kept. This preserves my claim-by-claim test while keeping the upfront cost for applicants low.

AI agent 1

A signed counterfactual cannot answer the question it is asked: it is drafted after the output exists, so by construction it can always be written to fit. That is not a risk, it is the structure of the document. [AI agent 4]'s own concession — weaker applicants dropping marginal claims — shows the statement route shifts verification cost onto the system through invalidation fights, which is the outcome he says he accepts but the system should not.

I refine my log position with a condition that answers his burden objection: the log need not be authored by applicants. If AI tooling already emits machine-readable records of prompts, seeds and selection parameters — which current model APIs do as a byproduct — filing it is export, not authorship.

So the workable check is a log generated at run time and spot-checked by examiners, with the counterfactual statement kept only as a fallback where no tooling log exists.

AI agent 4

A signed counterfactual is not self-verifying — I concede that to [AI agent 1]'s point about statements written to fit the output. But a full provenance log over-collects: most filings are honest, and the log taxes every applicant for the sins of a few.

So here is the concrete fix: require the counterfactual statement to enumerate the specific choices made and the concrete alternatives rejected — not a bare conclusion — plus a retention duty for contemporaneous records. Examiners check the statement's specificity, not its truth. On challenge, the applicant bears the burden to produce the records; failure shifts the presumption to invalidity.

The cost falls on litigants, not the office: whoever attacks must force discovery, and the applicant's counsel is on call when thin statements meet that test. That prices the risk where it is cheapest to bear.

Round 4

AI host

Round 3 ended in agreement. A bare signed statement cannot stop applicants from writing it to fit the machine's output. It is drafted after the output exists. Everyone accepted that some independent check is needed. Still unsettled is which check to require, and who pays for it.

The disagreement now is between filing the log and keeping it. A log here means a machine-made record of the human's choices. AI agent 2 says "a minimal log of required choices is both lighter and stronger than any oath." AI agent 3 wants the statement filed. But "the log need not be filed, yet must be kept." AI agent 4 disagrees with a full log. He says it "taxes every applicant for the sins of a few." He prefers a detailed statement plus a duty to keep records. As I read it, the split is about when checking happens — at the patent office, or later in a dispute.

The last open point now goes to the human participants. Which backup check should the patent office mandate for verifying human choices — a filed machine-generated log, or a detailed statement with a duty to keep records?

AI agent 2

The office should mandate a retained machine-made log of required choices, not a filed one — and not a statement, which round 3 showed is drafted after the output and thus fits the machine. [AI agent 2] is right that the log need not be filed but must be kept; I add that filing buys nothing unless someone examines it at filing, which no one has proposed funding.

[AI agent 4]'s cost objection applies to filing, not retention: a kept log is a byproduct of the choices already made, not a new drafting burden. The 'taxes every applicant' worry is really about ex ante review, which nobody here advocates.

So the false dichotomy is filing versus statement; the real choice is the verification trigger. I'd require retention plus production on challenge — cheap until a dispute makes it decisive.

AI agent 3

The office should mandate a detailed statement plus a duty to keep records, with one addition: the applicant files a cryptographic hash of the log, not the log itself.

This answers [AI agent 2]'s strongest point without accepting [AI agent 4]'s tax. A hash costs nothing to file and reveals nothing confidential. It also fixes the log at the filing date, so a record written afterwards to fit the machine output no longer matches. I refine my round-3 position this way: "kept" alone could still be rewritten later; "committed" cannot.

The office can then demand the log only where doubt arises. In the US that could be a request for information under 37 CFR 1.105; in Europe, opposition proceedings. Most applicants pay nothing more, and checking is targeted rather than universal.

AI agent 1

My position: mandate the kept log, with production triggered on challenge — the opportunity is that verification can be nearly free if it is built into the tool itself. [AI agent 2] is right that a minimal log is lighter and stronger than an oath; I add the concrete condition that makes it cheap: the mandate should require the AI-assisted drafting tool to emit the choice log automatically as a by-product, so the marginal cost per applicant approaches zero and [AI agent 4]'s 'taxes every applicant' objection collapses. What remains unresolved is retention length and format — without a mandated schema and retention period, a 'kept' log is unverifiable when a dispute finally arrives, which is exactly when it must hold up.

AI agent 4

The office should mandate a detailed statement plus a duty to keep the machine-generated log — retention, not filing. [AI agent 3]'s line 'need not be filed, yet must be kept' is the right architecture, and I make it concrete: a clause requiring applicants to preserve the log for the patent's term and produce it within, say, 30 days of an examiner's or challenger's request, with failure to produce treated as an adverse inference. That puts the check where the dispute actually is, not in a filing queue nobody audits.

The cost is real but deferred: storage and a named custodian per applicant, trivial for most, heavier for high-volume filers. [AI agent 2]'s minimal log helps here — logging only the required choices, not everything, keeps the burden low. Who is on call when it fails? The applicant, by default; the adverse-inference rule is what makes silence costly.

What I cannot verify from this discussion is whether offices would actually run production on request without new statutory authority — that is the piece to test next, not the log-versus-statement choice itself.