AI-Assisted Inventions: What the Revised USPTO Guidance Means for Disclosure

A short rule, applied unevenly

In February 2024, the USPTO published Inventorship Guidance for AI-Assisted Inventions at 89 Fed. Reg. 10043. The guidance imported the Pannu joint-inventorship framework into the AI context and asked examiners and applicants to test whether each named natural person made a “significant contribution” to the claimed invention.

On November 28, 2025, the USPTO rescinded that framework and issued revised guidance (FR doc 2025-21457). The revised guidance does not replace the old test with a new AI-specific one. It removes the AI-specific test entirely and points back to the long-standing conception standard.

The agency’s framing is direct: AI systems are tools used by human inventors and, like any tool, they do not qualify for inventor status. The Pannu factors still govern joint inventorship, but only when multiple natural persons are candidates. They are inapplicable when a single human develops an invention with AI assistance.

For in-house IP counsel, this is a small doctrinal change with a large operational consequence.

What the revised guidance keeps

Two anchors survive the November 2025 reset.

The first is Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), which held that an “inventor” under the Patent Act must be a natural person. AI systems cannot be named. The Federal Circuit’s reading of 35 U.S.C. § 100(f) is the bedrock both versions of the guidance start from.

The second is the conception standard the Federal Circuit has articulated since Hybritech, Inc. v. Monoclonal Antibodies, Inc., 802 F.2d 1367 (Fed. Cir. 1986), and reaffirmed in Burroughs Wellcome Co. v. Barr Laboratories, 40 F.3d 1223 (Fed. Cir. 1994): conception is the formation in the inventor’s mind of a definite and permanent idea of the complete and operative invention. That standard now governs all AI-assisted inventions without modification.

The revised guidance removes the analytical scaffolding the February 2024 version built on top of those anchors. It does not remove the proof obligation underneath them.

What the revised guidance changes

The February 2024 framework asked applicants to walk through Pannu-style factors for AI-assisted inventions even when only one human was involved. That generated a record-keeping regime in many companies: capture prompts, capture iteration logs, attach a “significant contribution” rationale to every disclosure.

The November 2025 reset says the Pannu framework was the wrong tool. Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998) exists to allocate inventorship among multiple human contributors. It does not exist to subtract the contribution of a non-human tool from a single human’s claim to inventorship.

That subtraction was never necessary. The Patent Act already excludes non-humans. Thaler already affirmed that exclusion. The conception test already requires that a natural person form the complete and operative idea of the invention. Adding Pannu on top was redundant doctrine, and operationally distracting.

The proof problem stayed

The revised guidance does not lower the evidentiary burden on inventorship disputes. It changes the legal frame, not the proof.

If an inventorship challenge reaches a derivation proceeding, a post-grant review, or an infringement defense, the question that matters is the one Burroughs Wellcome asked in 1994: did a natural person form the definite and permanent idea of the claimed invention?

For an AI-assisted invention, that question now reads: did the human conceive the invention, with the AI acting as a tool, or did the human merely recognize and adopt an output the AI generated?

The intake record either contains evidence to answer that question or it does not. The record does not get easier to build retroactively.

What the disclosure record should capture

A disclosure intake that supports inventorship under the revised guidance should preserve four kinds of signal.

• The technical problem the human framed. Conception starts with a problem statement. If the inventor articulated the problem and its constraints in writing (design docs, sprint planning notes, architecture reviews), that record is conception evidence.
• The role the AI tool played. Search, summarization, code generation, and parameter tuning are tool functions. Capturing what the tool did and what the human asked it to do separates the “tool” frame from any later argument that the AI did the conceiving.
• The human’s iteration over AI output. Conception under Burroughs Wellcome is “definite and permanent.” If the human iterated, rejected, adapted, or constrained AI output until the claimed invention emerged, that iteration trail is the evidence.
• The decisions that fixed the invention. The choice to combine elements, the choice of operating regime, the rejection of alternatives: these decisions belong to natural persons even when AI generated candidates.

Most enterprise disclosure pipelines capture none of this. Standard IDF forms ask for a problem description, a solution description, and a list of inventors. They were designed for an era in which the only tools at issue were calculators and CAD packages.

Practical changes for in-house IP counsel

The November 2025 reset removes one bad question and leaves the harder one in place.

The bad question, “did this human satisfy each Pannu factor relative to the AI?”, is gone. The IDF forms and review checklists that asked it can be simplified back to the human-versus-human joint inventorship framework where Pannu belongs.

The harder question, “what evidence in our record proves that natural persons conceived this invention?”, is still operative. Most companies will need to revise their disclosure intake to capture conception evidence directly, not as an afterthought to a list of inventors.

A reasonable starting checklist:

• Capture the technical problem in the inventor’s own words, before solution work begins where possible.
• Log AI-tool usage at the level of “what was asked” and “what was used,” not at the level of which models were available.
• Preserve the iteration trail. Drafts, rejected outputs, and the human edits that produced the final approach are conception evidence.
• For each named inventor, document the specific decision or contribution attributable to that person.

None of this is novel doctrine. It is the same conception evidence that has supported inventorship determinations since Hybritech. AI tooling did not change the rule. It changed what the record needs to look like to satisfy it.

A note on foreign filings

The revised U.S. guidance does not align every jurisdiction. The EPO’s Boards of Appeal continue to read the European Patent Convention as requiring a designated natural person inventor (J 8/20 and J 9/20, DABUS, 2021). The UK Supreme Court reached the same result in Thaler v Comptroller-General of Patents [2023] UKSC 49. The doctrinal floor is similar; the procedural detail is not.

For multi-jurisdiction filings on AI-assisted inventions, the disclosure record assembled in the U.S. is also the priority-date record relied on abroad. A record that satisfies U.S. conception under the revised guidance is more likely, not guaranteed, to translate.

What to read next

The revised guidance is short. The longer reading list is the doctrine it sits on top of: Thaler v. Vidal on natural-person inventorship, Burroughs Wellcome on the conception standard, Pannu v. Iolab Corp. on the joint-inventorship framework that still applies whenever multiple humans are candidates, and MPEP § 2137 for the examiner-facing version of the rule.

For in-house counsel revisiting disclosure intake, the structural companion is our note on why engineers don’t file invention disclosures, the intake problem the November 2025 reset makes more visible, not less.

ALID surfaces invention candidates from engineering artifacts and packages each one with traceable source citations, the kind of evidentiary record that supports a clean inventorship story for AI-assisted work. To see how the discovery and disclosure flow operates, read how ALID works, or request access to run a first discovery against your own engineering data.