AI patent search is now a daily task for most IP teams, yet a study dated 13 October 2026 found that the usual AI class code, G06N, missed 44% of AI-related US applications from seven large filers. The data covers mid-July to early October 2026. It suggests that a prior art search built around one class code leaves a large part of the field unread.
What the new data shows
Patent attorney Jon Burbage compared two publication windows a year apart, using USPTO Open Data Portal records. Publications in G06N, the main AI class, grew from 3,868 to 6,017, a rise of 56%, while all published utility applications grew by 28%.
Across seven companies with at least 20 AI-related publications each, G06N missed 419 of 949 applications. The miss rate was 77% for Apple, 71% for Meta and 37% for Alphabet. Many of these filings sit in natural language processing (G06F 40), computer vision (G06V) or speech (G10L), so a search limited to G06N never reaches them.
Titles with agent or agentic terms at eight companies rose from 14 to 47, which suggests where new filings are heading. The author advises adding the other three classes and searching titles and claims for model and agent terms.
Offices are adding AI to search, and the query still matters
WIPO announced on 2 July 2026 an assistant in PATENTSCOPE that turns plain-language instructions into structured queries. Version 1.0 handles keywords, date ranges and country codes. WIPO says the model runs on its own premises and keeps queries confidential.
The USPTO has taken a different route for applicants. Its ASAP! pilot gives an AI-generated prior art search before examination, and the office extended it until 1 June 2026. The tool ranks prior art from US and foreign databases, and replying to its notice is optional, according to Finnegan.
Two practitioner views
Michael Anderson of Marshall, Gerstein and Borun wrote in Bloomberg Law on 13 July 2026 that current AI search tools have not yet changed examination at a basic level. He expects better tools to bring more relevant prior art, extra rounds of prosecution and closer scrutiny, and he advises detailed specifications that describe particular technical improvements.
An MLex commentary by Freny Patel on 21 August 2026 reports that India's Patent Office has told examiners they may use AI but must not let it replace their judgement. The piece asks whether human oversight is enough, because the office's own experiments reportedly produced opposite novelty conclusions and inconsistent classifications.
How AI patent search should change in practice
We think the lesson is about method more than tools. A natural-language assistant such as the one in PATENTSCOPE can draft a first query in minutes. The person running the search still decides which classes and terms define the field. Class codes work best as a starting filter, and the edge of a topic has to come from reading the claims.
Consulting and market intelligence teams face the same gap when they count competitors' AI filings. A count built on G06N alone would make Apple and Meta look far smaller than the data suggests. That error can flow straight into a market sizing or a competitor ranking.
One practical step for this week is to run a live search twice. Use G06N only the first time, then G06N plus G06F 40, G06V, G10L and title terms such as "language model" and "agent", and read what the second run adds. Keep a human reviewer on the final list, because AI-drafted queries and AI-ranked results still need a check from someone who knows the technology.
The takeaway is to use AI to build and widen searches faster while keeping the class strategy in human hands. Watch for more data on agent-related filings, since that is where the title counts moved fastest.

0 Comments