Connect with us

AI

The Patent Office Made AI Search Mandatory After AI Flooded It

Patent due diligence now happens before a design freeze, after 209,518 AI filings and a USPTO order that examiners search with AI.

Published

on

Rob Maier, chair of Baker Botts’ New York intellectual property group, wrote on September 8 that artificial intelligence is reshaping patent practice from two directions at once. IFI CLAIMS counted 209,518 patent applications in artificial intelligence filed around the globe in 2025. The U.S. Patent and Trademark Office has required examiners to run AI search tools on every application since July 2025.

Patent due diligence used to wait for a deal or a lawsuit. It now has to happen while a product can still change, because the file is too large to read by hand and too noisy to trust a chatbot with unsupervised.

Global AI Filings Hit 209,518 in a Single Year

IFI CLAIMS, a patent-data firm, published those 2025 counts on July 28. Twenty-three percent of the global pile was generative AI, the large-model systems that write text, images, and video from a prompt. Agentic AI, systems that plan and run multi-step tasks, took 9% of global applications, up from 5% in IFI’s prior study.

Inside the United States the mix is sharper. Generative AI was 16% of the AI patent domain. Agentic AI was 15%, up from 7%. Worldwide, agentic applications rose 59% over two years. In the U.S. they rose 40%.

AI PATENT FILINGS IN 2025

Slice Share of AI filings
Global generative AI 23%
Global agentic AI 9%
U.S. generative AI 16%
U.S. agentic AI 15%

Overall U.S. AI filings, IFI found, climbed hard from 2018 to 2023 and then flattened. Generative and agentic claims did not flatten with them. Global AI grants have risen at a 35% compound annual rate, and applications at 29%. Generative grants worldwide expanded 54% a year, with applications up 44%. U.S. generative grants rose 31%, and applications 29%.

The World Intellectual Property Organization put a finer grain on the generative slice. Published GenAI patent families rose from about 14,000 in 2023 to over 37,000 in 2025. In 2024 and 2025 alone WIPO counted more than 56,000 new GenAI patent families, more than the entire 2014 to 2023 total. GenAI was 8.7% of all published AI patent families, up from 6.1% in 2023. In 2025, large-language-model families outnumbered generative-adversarial-network families by nearly three to one, 14,100 against 5,200.

U.S. GenAI families rose from about 1,185 in 2023 to over 4,380 in 2025, a 92% compound annual growth rate in WIPO’s tally. SoftBank became the largest single GenAI holder by volume, with nearly 3,000 families, almost all published in 2025. Alphabet was the largest U.S.-based GenAI owner, with 1,083 families and 640 published in 2024 or 2025. China-based inventors published over 43,000 GenAI families in 2024 and 2025, more than that country’s output from 2014 to 2023.

That is the first half of Maier’s two directions: more claims, in more subfields, landing in the same offices that already had a backlog.

The Patent Office Now Forces Examiners to Search With AI

John A. Squires, Under Secretary of Commerce for Intellectual Property and Director of the USPTO, told a House judiciary subcommittee on March 25 that the office received about 475,233 serialized patent applications in fiscal 2025, up 2% from fiscal 2024. The unexamined backlog fell from an all-time high of almost 838,000 applications in January 2025 to 788,229 at the end of fiscal 2025. The fiscal 2026 target is about 688,000.

First-action pendency still grew, from 20.5 months in January 2025 to 22.2 months in February 2026, because the office pushed resources at its oldest cases. Total pendency moved from 26.2 months to 27.9 months. Squires said the office is using AI so patents are “born strong,” with a human still in the loop.

THE USPTO’S AI SEARCH ROLL-OUT

  1. 2021: An autoclassification system and an AI “More Like This Document” search tool go live.
  2. 2022: AI-powered Similarity Search is added for examiners.
  3. July 2025: Use of the AI patent search tools is made mandatory for every patent application.
  4. August 2025: Specialized AI image search is deployed for design patents.
  5. October 8, 2025: The Artificial Intelligence Search Automated Pilot, ASAP!, launches for applicants.
  6. February 2026: Three enhanced Similarity Search tools are deployed.
  7. June 1, 2026: The extended ASAP! petition window closes.

The mandate is the tell. Examiners no longer get to skip the machine when the pile is this size.

ASAP! Gave Applicants a Top-Ten List

On October 8, 2025, the USPTO said ASAP! would test its internal AI on pre-examination prior art search. The tool reads the Cooperative Patent Classification, the specification, the claims, and the abstract, then sends the applicant an AI-Assisted Search Results Notice with a top ten list of potential prior art, ranked by relevance. No reply was required. The notice was not itself an office action.

To the Senate Judiciary Committee and inventors of all stripes, I promised to ‘lean-in’ to AI. Our ASAP! announcement today is the first of many planned AI pilots designed to help examiners and applicants alike ensure their patents are born strong.

John A. Squires, USPTO Director, October 8, 2025

Petitions were accepted for original, noncontinuing, nonprovisional utility applications filed on or after October 20, 2025. The first window ran through April 20, 2026, with a plan to take at least 1,600 applications. On April 16, 2026, the office extended petitions through June 1, 2026, raised the target to at least 3,200, and waived the petition fee for filings on or after March 23, 2026. The program is now closed.

WHAT THE TOP-TEN NOTICE LET AN APPLICANT DO

  • Amend early: File a preliminary amendment to put the claims in better shape before an examiner is assigned.
  • Build a record: Marshal evidence for affidavit practice if the AI hits look close.
  • Wait: Request deferral of examination and rethink the filing.
  • Walk away: Expressly abandon and seek a refund of certain fees if the hits kill the case.

Squires said quality starts at filing. The pilot was the office handing applicants a machine-ranked short list and asking them to react before the slow part of prosecution began.

Examiners Lost the Option to Skip the Tools

The July 2025 order is broader than the pilot. Every application, not just the ones that paid a petition fee, now goes through the AI search stack. The office also said AI has increased how often foreign prior art is put on the record. That helps validity. It also means a product team clearing a feature in 2026 is searching a file the office itself no longer pretends a person can finish unaided.

Why Early Clearance Moved Into Product Design

Maier’s second direction is the practical one. AI search is fast enough to run while a design is still plastic. A late opinion, after tooling and marketing are locked, is an invitation to strip a feature or pay a verdict.

Freedom-to-operate work used to mean a lawyer, a keyword search, and weeks of claim charts. The new sequence is a machine pass over millions of documents, a short list of claims that look close, and a human who still has to read those claims against the product. General-purpose chatbots are a poor substitute for that second step. They do not hold the structured patent record, and they will invent a citation when the prompt is hungry.

A first-pass scan is a screen. It is not a clearance opinion. Legal AI products now say as much in their own labels, calling patent checks “triage.” The expensive part has not moved. Judgment still sits with counsel who will sign the letter. What moved is the calendar. The scan is cheap enough to run at schematic review, so engineering, not litigation, becomes the place where a blocking claim is designed around.

Companies that wait until a launch freeze are buying a different product: a redesign under deadline, or a fight over a feature that already shipped.

Generic Machine Learning Claims Keep Getting Thrown Out

Speed does not tell you which of those 209,518 filings will survive. On April 18, 2025, the U.S. Court of Appeals for the Federal Circuit decided Recentive Analytics, Inc. v. Fox Corp., a first-impression case on machine-learning patents. Recentive owned four patents on using machine learning to schedule live events and to build television network maps. The district court in Delaware held the claims ineligible under Section 101. The Federal Circuit affirmed.

Today, we hold only that patents that do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models to be applied, are patent ineligible under § 101.

U.S. Court of Appeals for the Federal Circuit, Recentive Analytics, Inc. v. Fox Corp., April 18, 2025

The panel said iterative training on new data is “incident to the very nature of machine learning” and does not make machine learning better. Faster-than-human scheduling was not enough. Field-of-use limits, using the same model on a new kind of data, were not enough. The court left room for patents that improve the models themselves.

That holding cuts both ways in diligence. A portfolio stuffed with “apply AI to X” claims is weaker than a count of issued patents suggests. A product team can also over-read a hit. A claim that looks broad on a dashboard may be the exact class of claim Recentive just lost. The work is not finding the document. The work is reading whether the claim is a real fence or an abstract idea with a neural-net sticker on it.

A Watch Feature That Cost $634 Million After It Shipped

Late clearance has a published price in hardware. In November 2025 a federal jury in California awarded Masimo $634 million after finding that Apple Watch workout mode and heart-rate notification features infringed U.S. Patent No. 10,433,776, a pulse-oximetry patent that expired in 2022. Masimo called the verdict a win for its ability to keep building patient technology. Apple said the verdict was “contrary to the facts,” that Masimo “does not sell any products to consumers,” and that it would appeal. An Apple spokesperson said Masimo had asserted over 25 patents over six years and that most had been found invalid.

The International Trade Commission had already sided with Masimo in 2023 and restricted imports of Apple Watches with blood-oxygen sensing. Apple later moved the measurement and calculation onto a paired iPhone. In July 2026 U.S. District Judge James V. Selna refused Apple’s bid for judgment as a matter of law or a new trial.

WHAT LATE CLEARANCE COST IN THE WATCH FIGHT

  • The award: $634 million on a patent that had already expired.
  • The import hit: A 2023 ITC restriction that pulled blood-oxygen sensing out of U.S. watches.
  • The redesign: Sensor math moved off the watch and onto the phone so the feature could return.
  • The appeal path: Apple still contests the verdict after the trial judge left it in place in July 2026.

That fight is medical hardware, not a large language model. The diligence lesson is the same. A feature that ships without a hard read of the claims can be disabled, redesigned, or paid for at jury scale. AI search can surface the Masimo-class patent while the schematic is still on a whiteboard. It cannot decide whether the product is a “patient monitor.” That call is still a person with the file.

Samsung’s 682 U.S. Filings Dwarf OpenAI’s 27

IFI’s 2025 applicant table is the other half of the irony. The companies building the models are not the ones blanketing the office.

U.S. AI PATENT APPLICATIONS IN 2025

Applicant U.S. AI applications
Samsung 682
Google 671
Microsoft 585
IBM 460
Nvidia 401
Qualcomm 300
OpenAI 27
Anthropic 7

Worldwide, Samsung led with 2,162 AI applications, then Huawei with 1,822 and Google with 1,672. Nvidia led agentic filings, 225 worldwide and 128 in the U.S. Google took the top generative slot. Samsung’s 682 U.S. AI applications were more than 25 times OpenAI’s 27. Anthropic filed 8 patents worldwide and 7 at the USPTO. IFI found no AI patents for DeepSeek or for xAI.

On February 4, 2025, OpenAI said it would use its patents only defensively unless a party threatens or sues, helps someone else do so, or harms OpenAI or its users. That pledge does not thin the file around the models. Device makers, chip vendors, and platform companies are still filing into the same art units. A product that wraps a frontier model still has to clear Samsung, Google, Microsoft, Nvidia, and a long tail of application patents that Recentive says may not even be eligible.

Anthropic’s seven U.S. filings also sit beside live fighting over AI patents. The University of Tennessee Foundation’s suit against Anthropic is one reminder that a light portfolio does not mean a quiet docket. Diligence for a company shipping on top of these models is less about what the lab patented and more about what everyone else patented around it.

Two Models Checking Each Other Is Not Review

The USPTO’s Office of Enrollment and Discipline made the human-read rule public on July 27, 2026. In proceeding D2026-16, it reprimanded California patent attorney Brian E. Mitchell, the first discipline order predicated on reckless AI use. Preparing a joint claim-construction chart in Magpul Industries Corp. v. Mission First Tactical Group, Inc., in the Eastern District of Pennsylvania, Mitchell used one generative tool to draft constructions and a second tool to review them. The draft cited the patent’s specification, figures, and prosecution history. Some of those cites did not exist.

He corrected the chart the next day. The district court did not sanction him. The client was not harmed. OED still found violations of competence, diligence, and misrepresentation rules, and treated the second model as no review at all. “I had another model check it” does not discharge the duty to verify against the file.

That order is the diligence standard in one page. The office now requires AI search on every application. It will also discipline a lawyer who lets AI invent the intrinsic record. Product teams using the same class of tools to clear a feature are in the same bind. The scan is mandatory in spirit even when it is not mandatory in the CFR. The read is still the job.

Maier’s September 8 point lands there. AI stuffed the file, then became the only way to move through it. Early patent due diligence is now a design-freeze gate because the alternative is a Watch-style retrofit or a Mitchell-style citation that no one caught until after it was filed. He fixed the chart the next day. The USPTO still put the reprimand on the public list.

Disclaimer: This article is news reporting and analysis of patent-office practice, court holdings, and published filing data. It is for information only and is not a freedom-to-operate opinion, a patentability opinion, or legal advice for any product, filing, or dispute. Readers who need to clear a feature, value a portfolio, or respond to a claim should consult a registered patent attorney about their specific facts before they act. Figures and program statuses reflect the USPTO, WIPO, IFI CLAIMS, and court materials cited here and can change as new office actions, verdicts, and filing tallies issue.

Logan Pierce is a writer and web publisher with over seven years of experience covering consumer technology. He has published work on independent tech blogs and freelance bylines covering Android devices, privacy focused software, and budget gadgets. Logan founded Oton Technology to publish clear, no nonsense tech news and reviews based on real hands on testing. He has personally tested and reviewed dozens of mid range and budget Android phones, written extensively about app privacy, and built and managed multiple WordPress publications over the past decade. Logan holds a bachelor's degree in English and studied digital marketing at a certificate level.

Continue Reading
Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending