AI for IP Lawyers: The USPTO Rules, the Export-Control Trap and Trademark Clearance Limits

Suppose a patent attorney in Austin pastes an unfiled invention disclosure into a chatbot at 11pm to tidy the background section. The tool is on a paid tier, the vendor promises not to train on inputs, the draft comes back cleaner. If the vendor’s servers sit outside the United States, the attorney has also just exported technical data about an unfiled US invention. The scenario is hypothetical; the risk is not. The USPTO wrote it into its April 2024 guidance.

No other practice area has that trap, which is why AI for IP lawyers needs its own guide: the most specific regulator text on AI in the profession, an examiner side that adopted AI faster than the bar, and a naming problem where the tools pull in exactly the wrong direction.

The USPTO’s April 2024 guidance: six rules for practitioners

The guidance at 89 Fed. Reg. 25609 (11 April 2024) creates no new rules; it explains how the existing ones apply with AI in the room. Its most-quoted sentence is the whole philosophy: “Simply relying on the accuracy of an AI tool is not a reasonable inquiry.”

The IDS row is easy to miss: a model that finds forty plausible references is not helping if you cannot say what each is for. The foreign filing licence row gets its own section.

The export-control trap: AI tools on non-US servers

Three consequences. First, “no training” is not the question; region is. A vendor can honour a no-training clause perfectly while inferring on infrastructure in Frankfurt. Second, get the processing region in writing: Anthropic’s API, for instance, lets a customer pin inference to the United States at a price uplift, and the data residency guide has the vendor-by-vendor position. Third, the same disclosure bites twice in absolute-novelty jurisdictions: practitioners warn that sharing invention details with public AI tools risks third-party disclosure that matters in Europe.

The safe default for unfiled matter is the one the confidentiality guide recommends everywhere, plus one line: US-region processing, confirmed in writing, kept with the file.

Inventorship after the November 2025 revised guidance

The USPTO’s February 2024 inventorship guidance is gone. The Revised Inventorship Guidance for AI-Assisted Inventions, 90 Fed. Reg. 54636 (28 November 2025) rescinds it “in its entirety” and replaces it with one principle: “The same legal standard for determining inventorship applies to all inventions, regardless of whether AI systems were used in the inventive process.” AI systems “are analogous to laboratory equipment, computer software, research databases, or any other tool that assists in the inventive process.”

If the tool is lab equipment, the inventor is whoever conceived the claimed invention using it, and the file should show who that was, claim by claim, at the time. A model will not flag an inventorship error, and correction under 35 U.S.C. § 256 is not where you want to discover one.

What AI drafts badly: sweeping claims and unsupported specification detail

Drafting is where general models look most impressive and are most dangerous: the output is fluent and the defects surface in litigation. Thompson Patent Law’s analysis of AI-drafted applications names two failure modes. Models cannot balance claim breadth against prior art, producing “sweeping claims that read straight onto known systems and die under §§102 and 103”. And they fill specification gaps with “plausible-sounding but unsupported detail”, creating § 112 defects that “survive prosecution unnoticed, then surface during licensing, validity challenges, or enforcement”.

The workflows that survive are those where the model checks a draft against a document rather than generating from nothing.

For each independent and dependent claim in <claims>[paste]</claims>, quote the passage(s) in <specification>[paste]</specification> that provide written-description and enablement support, with paragraph numbers. Where support is partial, say what is missing. Where you find none, write NO SUPPORT FOUND rather than proposing language. Then list every technical statement in the specification not tied to a figure, example or data, so I can decide whether each is inventor knowledge or filler. Do not add, rewrite or broaden anything.

Using the Office Action, the claims as filed and the full text of each cited reference in <refs>[paste]</refs>: for each rejected claim, state the rejection type and quote the examiner's reasoning; chart the examiner's mapping against the passages actually cited; identify elements you cannot find at those passages, quoting what is there instead; list candidate arguments (teaching away, missing element, no motivation to combine, unexpected results) each tied to a specific passage; and list candidate amendments with paragraph-numbered support. Do not draft the response. Flag any argument that would need a declaration.

Both are closed-universe tasks: the model reads what it was given and reports; the practitioner argues and signs. The contract drafting guide shows the same pattern in transactional work.

Trademark examination has changed: Class ACT and Scout LLM

The examiner side moved faster than the bar. On 19 March 2026 the USPTO launched Class ACT, the Trademark Classification Agentic Codification Tool, which assigns international classes, design search codes and pseudo-marks in minutes rather than up to five months. By July it had analysed 250,000 applications, and on 1 July 2026 the Office’s Scout LLM reached full adoption across Trademarks under a carefully chosen slogan: “Examining attorneys lead, tools support.”

For applicants: classification, design coding and pseudo-mark assignment are now proposed by a model and confirmed by an examining attorney, so identifications of goods and services should be machine-legible as well as legally precise. The Office’s own rule, tool proposes and attorney decides, is the standard it will expect of you too.

AI clearance searches: recall, common-law gaps and the 8-12 week rule

AI clearance tools are good at one thing: semantic recall across large registers. The 2026 comparison at AI Trademark Review reports word-mark recall above 90% and design-mark recall of around 70 to 80%, with a semantic catch a character search misses, “KwikKopy” against “Quick Copy”. It also gives the failure: a coffee-shop logo flagged against a car-repair logo, since “the AI may not understand the commercial context”. Its verdict: “AI can identify potential conflicts, but it cannot assess the commercial context, the strength of a mark, or the likelihood of confusion in the marketplace.”

The timing advice is the part clients ignore: run AI searches “8-12 weeks before you plan to file” and treat every result as preliminary. Two gaps do not close with better models: the USPTO’s own AI image search is US-only with no common-law coverage, and aural and conceptual similarity need a human ear.

Here are the results of our clearance search for [MARK] in classes [X] in [jurisdiction]: <results>[paste]</results>. Group each hit as High, Medium or Low conflict risk using: mark similarity (visual, aural, conceptual, one line each), goods and services proximity, and status (registered, pending, abandoned). For each High hit, give the arguments for and against likelihood of confusion and the further information we need (use, dates, coexistence). Add no hit that is not in the results. End with the questions to put to the client before we opine.

AI-generated names and logos: ownership and descriptiveness

Ownership is the easy question. As a Chicago practitioner put it, the owner is “the person or entity that controls the nature and quality of the goods or services sold under the mark, not the software that drew the logo”. Two checks follow: read the platform’s terms for any claim over outputs, and document the human decision to adopt the mark.

Here are [12] candidate names a client generated for [goods/services]: <names>[paste]</names>. For each, classify along the spectrum generic / descriptive / suggestive / arbitrary or fanciful, with one sentence of reasoning and the specific word that drives the classification. Flag any candidate that describes an ingredient, quality, function or geographic origin of the goods. Rank the candidates by likely registrability on that ground alone. Cite no cases or TTAB decisions; I will add authority.

IP litigation: the Concord v. Anthropic declaration

A formatting request is still a generation request. The fix is mechanical: paste the real citation, ask only to reformat the exact text, then check the output against the source. The citation verification guide sets out the six-layer check; if a fake has already been filed, the what-to-do guide covers the candour courts reward.

Expert declarations deserve their own clause. In Kohls v. Ellison (D. Minn., January 2025) a Stanford expert typed “[cite]” as a placeholder, GPT-4o filled it with fabricated studies, and the declaration was excluded. Proposed Federal Rule of Evidence 707, which would subject machine-generated evidence to Rule 702-style reliability requirements, was deferred by the Advisory Committee to Fall 2026. Until it lands, the expert’s engagement letter is the rule: say what AI use is permitted and require disclosure.

A compliance checklist for patent prosecution

Short, because the duties are old; only the tool is new.

1. Verify everything the signature covers, including every IDS reference, by opening the source, not by asking the model.
2. Read the vendor’s retention and training terms before any client matter goes in (the vendor due-diligence checklist has the questions).
3. Confirm the processing region in writing for any tool that touches an unfiled invention; a non-US region is a § 5.11© issue until a licence or filing exists.
4. Document inventor contribution per claim at drafting time; the tool will not flag errors.
5. Consult the client on AI use as part of the means of representation (§ 11.104) and reflect the answer in the engagement letter.
6. Never let the model fill a placeholder. Search the final document for “[” before filing.

One table makes this concrete: every AI tool the practice uses, its tier, its training terms, its processing region and the USPTO duty it touches. IP practitioners in AI Lab for Lawyers build it in one session with the vendor pages open, and most find a tool in the wrong column.

Where to go next: the practice-area hub has the sibling guides, including employment lawyers for the trade-secret and non-compete side of IP work, and the prompt library has the prompts above in reusable form.