What the USPTO's AI does in trademark examination, and what it doesn't

Two claims about the USPTO and artificial intelligence are going around. One says the agency now uses AI to examine trademarks. The other says AI will finally make examination consistent. The USPTO's own announcements support neither claim as stated. They describe tools that prepare and organize applications and help staff find things. None of them describes a tool that decides whether a mark is registrable. On consistency, the best evidence we have predates these tools. Nobody has yet published data that would show whether they change anything.

What the tools do

Class ACT. On March 19, 2026, the USPTO announced the Trademark Classification Agentic Codification Tool. It does three pre-examination jobs. It assigns international classes to applications that arrive unclassified. It assigns design search codes, the numeric labels that make logos findable. It generates pseudo marks, the alternative spellings that make wording like "4EVER" turn up in a search for "forever." The Office says this work used to take about five months and now takes minutes. It says the output is "still reviewed by humans." By July 13, 2026, the agency reported that Class ACT had analyzed 250,000 applications.

Search and filing aids. On April 2, 2026, the USPTO added an image search, labeled beta, to its Trademark Search system. It requires a USPTO.gov login and returns marks with similar design elements, "the same as searching by an image on an internet search engine." The same notice announced a mark-description and color-claim generator for Trademark Center, scheduled for April 23.

Scout. On July 1, Trademarks opened Scout, the agency's general large-language-model tool, to all its work units, examining attorneys included. The USPTO says Scout "summarizes information, analyzes data, researches topics, and drafts selectable content." The announcement puts the division of labor in four words: "Examining attorneys lead, tools support."

The Docket · WeeklyAnalysis like this, plus every TTAB decision summarized. Fridays.

Email addressSubscribe

What they do not do

None of these announcements says that a tool decides a refusal, and the statute assigns that decision to a person. The Lanham Act sends each application to an examiner, and registration follows only if "it shall appear" to the examiner that the applicant is entitled to it (15 U.S.C. §1062(a)). A final refusal can be appealed to the Trademark Trial and Appeal Board (§1070), and the Board's decision can be appealed to the Federal Circuit. The Class ACT announcement presents the tool as a way to free examining attorneys to apply "experienced judgment and reason to the substantive issues in examination."

There is also something the announcements leave out. The USPTO says Class ACT works "at a high level of accuracy," but it has published no accuracy rate, error rate or rate of human correction for Class ACT, the image search or the description generator. We found no reported bias, error pattern or litigation involving the Office's trademark AI. That is not evidence that there are none. With no published measurements, nobody outside the agency can check.

The consistency problem predates the tools

The case for consistency tools rests on a real finding. Michael Frakes and Melissa Wasserman studied more than 7.8 million applications from 1982 onward. They used the USPTO Office of the Chief Economist's Trademark Case Files Dataset and covered 1,308 examining attorneys, who are assigned applications at random. The study is Are There as Many Trademark Offices as Trademark Examiners?, 69 Duke L.J. 1807 (2020). Its findings:

• Publication. The average examining attorney approved 78 percent of applications for publication, with a standard deviation of 11.5 points. A typical examining attorney therefore falls somewhere between about 67 and 89 percent (at 1831). At the tails, the 2nd percentile published about 56 percent and the 98th percentile published every application it reviewed (at 1833).
• First actions. The spread was wider for the first substantive decision. Rates ran from 11 percent at the 2nd percentile to nearly 84 percent at the 98th (at 1840).
• Registration. Registration rates averaged about 58 percent, with a standard deviation of 13.6 points (at 1842).
• Oppositions. The average examining attorney's approved applications drew a sustained opposition about 7 times in 1,000. At the 98th percentile the figure was about 22 in 1,000, roughly three times the average (at 1845).
• Convergence. The spread did not narrow over time. Variation in publication rates was "just as high in the 2010s as it was in the 1980s" (at 1852).

The authors adjusted for application characteristics, and the spread remained. The adjustment cannot capture everything that differs between applications. Still, because applications are assigned at random, the variation is hard to explain away.

The courts have now made the same point in doctrinal terms. In In re Brunetti, No. 2023-1539 (Fed. Cir. Aug. 26, 2025), the court noted that the Office had more than 700 examining attorneys in 2022. It accepted that the agency cannot guarantee that every examiner decides alike. But it treated "the apparent inconsistency of past examining attorneys' decisions" on all-purpose words as "strong evidence of the failure to articulate coherent rules." It vacated the Board's failure-to-function decision as an "I know it when I see it" approach. The court's remedy was a clearer stated standard, not a new tool.

What the numbers show so far

The figures the USPTO has published are about speed. In a May 5, 2026 notice, the Office reported that first-action pendency fell from 5.6 to 4.45 months in the first half of fiscal 2026, and total pendency from 11.7 to 10.03 months, while filings rose 7 percent. The notice does not attribute the gains to AI. It mentions hiring and an "automated processing tool" that routes statements of use. Pendency, then, is falling at the same time AI is arriving. The published data do not show that one causes the other.

What would answer the question

This section is the editors' opinion.

Whether the tools make examination better or more uniform is an empirical question, and the data to answer it mostly exist. Three measures would go a long way:

1. Accuracy of pre-examination coding. How often staff change Class ACT's classes, design codes and pseudo marks, broken down by class and by mark type.
2. Search recall. For the image search, how often marks later cited against an application under §2(d) appear in its results for the applicant's own logo.
3. Examiner spread, before and after. The Frakes and Wasserman measures, repeated for applications examined before and after Scout became available. The Case Files Dataset now covers 12.7 million applications through March 2024. It can supply the baseline. The later years will have to come from the Office's future releases.

None of these measures requires disclosing anything confidential about an application. Publishing them would let practitioners judge the tools by their results.

AI on the filer's side

The documented trademark harms from AI so far come from filers, not from the Office. The USPTO's April 2024 guidance (89 Fed. Reg. 25609) warns against submitting "AI-generated specimens" that do not show actual use of the mark in commerce (at 25616). It also says that "simply relying on the accuracy of an AI tool is not a reasonable inquiry" under the signature certification in 37 C.F.R. §11.18 (at 25614). Press reports describe a 2025 public reprimand of a practitioner whose TTAB brief cited decisions that do not exist. The same rules apply to everyone who files, whatever tools they use.

What to do differently

• Search closer to filing. Class ACT codes new applications almost immediately, so a clearance search goes stale faster than it did. Run it, or update it, shortly before you file.
• Check the codes on your own application. Classification and design codes are now assigned by machine and then reviewed. If the design codes on your record do not describe your logo, people searching for marks like yours may not find it. The USPTO's contact page gives a mailbox for design code corrections.
• Treat image search as a lead, not a clearance. The beta tool has no published recall figures. Use it alongside a coded design search and a word search.
• Argue the record, not the tool. Refusals are still written and signed by examining attorneys and reviewed by the Board. Build responses from the statute, the TMEP and precedent, as before.

This is general information, not legal advice. The Before You File guide on searching before you file explains the searches in plain language.