Published GenAI patent families doubled in a year. The prior art that will be used against your next filing is already dated, already indexed, and still invisible.
The generative AI race is no longer defined solely by who builds the best models. Increasingly, it is also about who secures the intellectual property surrounding them.
WIPO's July 2026 update to its generative AI patent landscape puts a number on the shift. More than 56,000 GenAI patent families were published in 2024 and 2025 alone, exceeding the entire cumulative output of the preceding decade from 2014 through 2023. Publications roughly doubled year over year, from 18,862 in 2024 to 37,808 in 2025, growth that in WIPO's words "far surpasses the long-term average of 43% annually since 2014."
For companies integrating AI into their products, those numbers are not trivial. Every newly published application expands the body of prior art against which future applications will be examined. As that body grows, meaningful patent protection depends increasingly on identifying and claiming an actual technological advance rather than an application of someone else's.
The window has not closed. It has, however, become considerably more competitive.
The Patent Landscape Visible Today Is Already Out of Date
Under 35 U.S.C. § 122(b)(1)(A), an application publishes 18 months from the earliest filing date for which a benefit is sought. That clock runs from the priority date, not the filing date of the application actually pending. A nonprovisional filed on the one-year anniversary of its provisional therefore surfaces publicly about six months later, and everything filed in the preceding 18 months remains invisible.
WIPO makes the same point about its own data, attributing the 2024–25 surge to "the typical 18-month lag between patent filing and publication." The clearest illustration in the report is SoftBank, now the single largest GenAI patent holder at 2,985 families, "virtually all of which were filed in 2023 and published in 2025."
The doctrinal consequence is the part that companies and their IP counsel should not miss. Prior-art status does not attach when a document becomes public. Under 35 U.S.C. § 102(a)(2), a published application or issued patent is prior art if it "was effectively filed before the effective filing date of the claimed invention," and § 102(d)(2) fixes that effective date as "the filing date of the earliest such application that describes the subject matter." The 37,808 families published in 2025 are a lagging readout of a prior-art pool that already existed, already carried its dates, and was already operating against applications filed in 2024.
This pool of so-called "secret prior art" in AI is measurable and growing. An April 2026 empirical study of roughly 10,000 randomly sampled office actions per year found that nearly 30% of rejections now cite at least one legally secret reference, up from about 20% a decade ago, and that about 25% of rejections rest on prior art that was not available at the time of filing. Seventy-nine percent of those secret references are pre-grant publications, with an average secrecy gap of 371 days. Electrical engineering and telecommunications show the highest rates; mechanical engineering the lowest.
Nor is the problem confined to examination. In Lynk Labs, Inc. v. Samsung Electronics Co., 125 F.4th 1120 (Fed. Cir. 2025), the Federal Circuit held that a published application can qualify as a "printed publication" in an inter partes review even though it was not publicly accessible until after the challenged patent was filed. The statutory term "printed publication" is "otherwise temporally agnostic," the court explained; the timing comes from elsewhere in § 102, which in that case was pre-AIA § 102(e)(1). The post-AIA analogue is § 102(a)(2)'s effective-filing-date rule, which reaches the same place by a different route. The Supreme Court denied certiorari on March 9, 2026, leaving the Federal Circuit's holding the current law. Invisible prior art is an invalidity risk, not merely an examination inconvenience.
First-Inventor-to-File Is a Tiebreaker, But Not Dispositive
The United States has operated on a first-inventor-to-file basis since March 16, 2013, under 35 U.S.C. § 102 as amended by § 3 of the America Invents Act. Filing timing is a genuine strategic variable. It is not, however, a naked race, and the qualifications matter because one of them answers the fear the data above creates.
Section 102(b)(1) preserves a one-year grace period for the inventor's own disclosures. Section 102(b)(2)(B) goes further: a patent applicant's own prior public disclosure removes an earlier-filed, later-published third-party application as prior art altogether. Sections 102(b)(2)(C) and 102(c) supply common-ownership and joint-research-agreement shields, which matter to startups working under sponsored research or university collaborations. Section 135 provides a derivation remedy where an earlier applicant took the invention from someone else, available within one year of that application's publication or grant.
The urgency to file thus comes from § 102(a)(2)'s effective-filing-date rule and the invisible pipeline behind it — not merely from a footrace in which whoever reaches the filing window first prevails regardless of provenance.
What the Federal Circuit Held in Recentive
In Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. Apr. 18, 2025), the Federal Circuit addressed what it called a question of first impression: whether claims that do no more than apply established machine learning methods to a new data environment are patent eligible. The court found they are not, and the Supreme Court denied certiorari on December 8, 2025.
The holding is narrower than has been reported, however, and the court said so:
"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." 134 F.4th 1216.
The sentence immediately before it is the counterweight that most summaries drop: "Machine learning is a burgeoning and increasingly important field and may lead to patent-eligible improvements in technology."
The defect was in the claiming and the disclosure, not in whether the patentee had built something. Recentive conceded that it was "not claiming machine learning itself," and the court found that "neither the claims nor the specifications describe how such an improvement was accomplished." Id. at 1212–13. Iterative training and dynamic adjustment did not help, because those features are "incident to the very nature of machine learning." Id. at 1212. Nor did the new field of use: "[a]n abstract idea does not become nonabstract by limiting the invention to a particular field of use or technological environment." Id. at 1213.
Cases applying Recentive sharpen the practical rule. In Dental Monitoring SAS v. Align Technology, Inc., No. 2024-2270 (Fed. Cir. July 7, 2026), the court extended Recentive from a Rule 12 dismissal to summary judgment and rejected the argument that specialized training data cures the problem: "[t]hat the 'deep learning device' here is trained on specific images does not make the 'deep learning device' itself non-generic." In Rensselaer Polytechnic Institute v. Amazon.com, Inc., Nos. 2024-1725, -1739 (Fed. Cir. Feb. 24, 2026), the court held that "[a] conventional application of case-based reasoning, even to a novel environment, is abstract." Both dispositions are nonprecedential: under Federal Circuit Rule 32.1, they are freely citable and the court may look to them "for guidance or persuasive reasoning," but they do not bind a future panel. No precedential Federal Circuit decision has yet applied Recentive to machine learning claims, which is itself worth noting before treating the case as a settled rule.
Machine learning claims have nonetheless survived, and how they survived is the useful part. In Aon Re, Inc. v. Zesty.ai, Inc., No. 1:25-cv-00201 (D. Del. July 15, 2025), a district court denied a motion to dismiss: "Aon's patents do not offer a new twist on machine learning itself. But that is not fatal, because we agree with Aon that the patents recite the patent-eligible arrangement of two independently trained classifiers." The court identified the second route expressly: "claiming a specifically arranged implementation of machine learning, analogous to the patent-eligible arrangements of McRO and Koninklijke."
The District of Delaware likewise denied summary judgment in Nielsen Co. (US), LLC v. Hyphametrics, Inc. (D. Del. July 22, 2025), where neural-network claims recited "specific steps for improving computer functionality." That decision carries particular weight on where the Recentive line stops, because it was written by Judge Gregory B. Williams — the same district judge who granted the Rule 12(b)(6) dismissal in Recentive that the Federal Circuit affirmed. Distinguishing his own decision, he wrote that "this case is not one where the patents merely 'apply machine learning to this new field of use.'"
The operative instruction for anyone drafting patent applications on AI-enabled technology: claim the arrangement and the mechanics, not just the application. "Apply a language model to [vertical]" is likely to be rejected. Technical implementations, like a specific orchestration architecture, a described routing mechanism, or a verification pipeline whose operation is disclosed, tend to survive patent-eligibility challenges.
The USPTO Has Moved in the Opposite Direction
While the Federal Circuit tightened, the Patent Office loosened.
An August 4, 2025 memorandum to the examining corps instructs that a § 101 rejection is appropriate only where it is "more likely than not (i.e., more than 50%)" that a claim is ineligible, that "[u]ncertainty alone doesn't justify rejection," and that "[c]laim limitations that encompass AI in a way that cannot be practically performed in the human mind do not fall within" the mental-process grouping.
Ex parte Desjardins, Appeal No. 2024-000567, went further. An Appeals Review Panel convened by Director Squires vacated the Board's § 101 rejection on September 26, 2025, and the decision was designated precedential on November 4, 2025. Although the claim recited a mathematical calculation, the panel held the claim as a whole was not directed to an abstract idea: adjusting model parameters "to optimize performance on the second task while protecting performance on the first," which overcomes catastrophic forgetting, is an improvement in how the model itself operates. A December 5, 2025 advance notice revises MPEP §§ 2106.04(d), 2106.05(a), and 2106.05(f) and adds two examples recognizing improved machine learning training methods and parameter adjustment as eligible improvements.
The divergence is the fact that should govern portfolio decisions. An AI patent is easier to obtain in 2026 than it was in 2024, and no easier to enforce. Examination guidance binds examiners; it does not bind a district judge, who must apply Recentive. A portfolio built to the agency's standard rather than to the court's risks creating a portfolio of assets that issue and then fail when tested.
Building on Foundation Models Does Not Eliminate Patent Opportunities
Many companies assume that reliance on foundation models from companies like OpenAI, Anthropic, Google, or Meta leaves little to patent. In practice, the commercially valuable engineering sits well above the model layer: agent orchestration, retrieval-augmented generation, model routing and selection, hallucination detection and verification, security and privacy controls, workflow automation, and domain-specific deployment architecture.
Such aspects can be patentable subjects when claimed as technical arrangements and disclosed at the level of how they work. The same paragraph of engineering can yield an eligible claim or an ineligible one depending on whether the specification adequately discloses the mechanism or merely describes the desired outcome. (See Recentive).
Some recent cases are illustrative. FriendliAI Inc. v. Hugging Face, Inc., No. 1:23-cv-00816 (D. Del.), asserted U.S. Patent No. 11,442,775, directed to inference scheduling and continuous batching for large language model (LLM) serving, against a range of Hugging Face products before settling in January 2025. Inference scheduling is an implementation-layer invention. While not fully adjudicated, the asserted patent was enforceable enough to produce a settlement in this case.
As a counterpoint, WIPO reports that OpenAI held 35 patent filings globally as of late 2025, concentrated on "product-level innovations" rather than foundational architectures, which the report reads as "a continued reliance on trade secrets and speed of execution as its primary competitive advantages, with patents serving a complementary role." For a company whose moat is release velocity, that is a defensible allocation. For a company whose enterprise value will be assessed in a diligence room, investors may desire more tangible protection like patents.
Where the Filing Activity Actually Lies
Large language models have overtaken generative adversarial networks (GANs) as the largest GenAI model category, at roughly 20,900 published families cumulatively against about 18,800 for GANs. The 2025 gap is wider still: 14,100 LLM families against 5,245 for GANs. Diffusion models grew from 441 families in 2023 to nearly 4,000 in 2025 and now rank third at roughly 6,500 cumulative. WIPO cautions that not every GenAI family maps to a model type, since some claim a use case rather than an architecture.
Reasoning models. WIPO identifies inference-time reasoning as "a new frontier for patenting activity around reasoning systems, inference-time scaling and model efficiency." This is the newest category in the data and the least crowded.
Agentic systems. WIPO describes agentic AI as "still a nascent field" with early activity "from companies such as Alphabet (Google) and Nvidia among the first movers." Commercial data puts numbers on it: agentic AI accounted for roughly 9% of the 209,518 AI patent applications published worldwide in 2025, up 59% year over year and 137% over two years, with U.S. agentic applications rising about 40%. Nvidia led with 128 U.S. applications and 225 worldwide, followed by Microsoft at 112 and Google at 91. Those counts rest on a vendor's own classification of "agentic AI," which is not published. The combination is what matters strategically: small in absolute terms, the fastest-growing segment in the field, and already incumbent-led. The implementation layer is where the patent land grab is happening, and it is where the 18-month blind spot is densest.
Efficiency in training and inference. Techniques that reduce inference cost, compress models, improve latency, or allocate compute are, in WIPO's framing, "likely to become increasingly important directions for both research and patenting" as enterprises pursue scalable deployment.
Multimodal systems. Systems combining text, images, video, audio, code, and sensor data continue to expand the innovative scope of GenAI. Image and video remains the largest modality at roughly 40,000 cumulative families, with over 13,800 published in 2025 alone; text more than tripled in two years, from 3,400 families in 2023 to nearly 11,800 in 2025.
The applicant base is broadening as well. WIPO reports that "telecommunications firms, infrastructure providers, industrial companies and utilities" are now joining traditional software and internet companies as leading applicants: State Grid Corporation of China at 1,144 families, China Mobile at 573, Bosch at 368, Ant Group at 353. WIPO expects integration into healthcare, manufacturing, finance, and energy to expand that base further.
Filing Strategy: Early, and Deep
Waiting until a product reaches maturity allows competitors to establish earlier effective filing dates on the same technical concepts. That much is straightforward. The less obvious point is that filing early is not always sufficient, as the strength of a filing date turns on what the earliest document actually describes. Two lines of authority now converge on that single control point, and one of them moved just yesterday.
An applicant's own provisional supports later claims only to the extent it describes them. Under 35 U.S.C. § 119(e), a provisional patent application must adequately describe the claims for which the priority date is sought. New Railhead Mfg., L.L.C. v. Vermeer Mfg. Co., 298 F.3d 1290 (Fed. Cir. 2002), is a cautionary case: a provisional that did not describe the claimed feature cost the patentee the priority date and, with it, the patent. In re Riggs, 131 F.4th 1377 (Fed. Cir. 2025), went further on the prior-art side, holding that support for a single claim is not enough because "the portion of the application relied on by the examiner as prior art must be supported by the provisional application." Riggs is a pre-AIA § 102(e) decision and does not address the post-AIA statute in effect since 2013.
In a newly decided case, a competitor's provisional now faces the same test. The Board had held in Penumbra, Inc. v. RapidPulse, Inc., IPR2021-01466 (PTAB Mar. 10, 2023) (precedential) that under post-AIA § 102(d)(2) a reference need only satisfy "the ministerial requirements of 35 U.S.C. §§ 119 and 120," with no inquiry into whether its provisional supported any claim. On August 10, 2026, in Dental Monitoring SAS v. Align Technology, Inc., No. 2025-1752 (precedential), the Federal Circuit rejected that interpretation and vacated the Board's decision. Judge Lourie, writing for a panel that included Judges Schall and Taranto, held that "the statutory text requires § 112(a) support for at least one of the prior art patent's published claims before that reference may obtain an earlier filing date for prior art purposes." A challenger "must show that [the reference's] provisional application provides written description support for at least one claim" of the reference, and "[t]he requirement to do so is not merely 'ministerial.'"
The court disposed of the argument that Dynamic Drinkware, LLC v. National Graphics, Inc., 800 F.3d 1375 (Fed. Cir. 2015), died with the AIA: its footnote declining to reach § 102(d) was "a reservation of decision," not "a holding on the merits," and "even assuming Dynamic Drinkware did not control because it addressed pre-AIA § 102(e), that conclusion does not suggest that the AIA altered the underlying principle." The court also removed the defense that its own earlier affirmance had blessed Penumbra: a Rule 36 judgment "does not endorse or reject any specific part of the [Board's] reasoning" and "has no precedential value."
The practical consequence is a single rule running in both directions: a provisional is worth exactly what it describes. A thin provisional will not support the filer's later claims, and it will no longer hand a competitor an early prior-art date either. What remains open is whether Riggs's second layer — support for the specific passages relied upon, not merely one claim — migrates into the AIA regime as well. Dental Monitoring did not reach that question and did not cite Riggs.
Two consequences follow immediately. Written description in the earliest filing is now the highest-leverage decision in an AI patent program, as it controls both what the filer can claim and what a competitor can assert. And any invalidity position built on the Board's Penumbra framework over the last three years is worth re-examining, as the premise it rested on has now been rejected by the Federal Circuit.
Section 112 is also the next constraint to arrive on the eligibility side. Amgen Inc. v. Sanofi, 598 U.S. 594 (2023), requires enablement across the full scope of a claimed class, and PTAB practice already draws the line for machine learning: written description was upheld where inputs, outputs, and methodology were described, and rejected where a specification said only that weights "may be specified by a subject matter expert." As patent eligibility under § 101 loosens at the agency, written description and enablement under § 112 increasingly becomes a binding limit, and claims that survive one may not necessarily survive the other. For applicants filing in Europe, EPO Guidelines G-II 3.3.1 makes the requirement explicit: where a technical effect depends on characteristics of the training dataset, those characteristics must be disclosed, though the dataset itself need not be.
Cost and timing cut the same way. The USPTO reported on April 10, 2026 that the unexamined application inventory had fallen to 776,995 as of April 6, down from a January 2025 peak of 837,928. Pendency, however, moved the other way: Director Squires' March 25, 2026 testimony to the House Judiciary IP Subcommittee reports first-action pendency rising from 20.5 months in January 2025 to 22.2 months in February 2026, and total pendency from 26.2 to 27.9 months. The fee schedule effective January 19, 2025 added a surcharge under 37 C.F.R. § 1.17(w) of $2,700 for presenting a benefit claim more than six and no more than nine years after the earliest benefit filing date, and $4,000 beyond nine years, net of any amount already paid. That clock runs from the earliest benefit date, which is another reason to place the anchoring application early rather than to build a long provisional-to-continuation chain. Track One prioritized examination, available to applicants with additional fees, targets final disposition in about twelve months, and the annual cap on such applicants rose from 15,000 to 20,000 requests on July 8, 2025.
What a Patent Will Not Do
The same invisible pipeline that argues for filing early means freedom to operate cannot be cleared either. A 371-day average secrecy gap makes a clearance search in generative AI incomplete by construction, and a patent confers a right to exclude, not a right to practice. 35 U.S.C. § 271. The prevalence of unpublished patent applications in AI is a further argument for a defensive patent portfolio.
No startup can file on every implementation detail, and two complements close the gap. Publication under § 122(b) destroys trade secrecy as a matter of law whether or not a patent ever issues, and the nonpublication route under § 122(b)(2)(B) is unavailable to any applicant who will file abroad, which describes most AI companies with enterprise ambitions. For disclosures that will not be claimed, defensive publication costs almost nothing, becomes § 102(a)(1) prior art against everyone else, and under § 102(b)(1)(B) neutralizes a competitor's earlier-filed but later-published application on the same subject matter.
Practical Steps
- Identify the technical innovation, and identify it as a mechanism. Companies often describe products in terms of customer outcomes. Patent protection depends on how the technology achieves those outcomes: for example, the architectural arrangement, the routing or scheduling logic, the verification pipeline, the training methodology, the security mechanism. Recentive has made this distinction between the outcome and the mechanism dispositive rather than stylistic.
- Put the implementation details in the first filing. After Dental Monitoring, written description in the earliest document controls both the filer's own priority and a competitor's ability to use its filings against others. A provisional drafted as a placeholder now underperforms in both directions.
- Run landscape reviews, and treat patentability and freedom to operate as separate questions. Landscape work reveals both crowded areas and underdeveloped ones. It also separates two inquiries that companies routinely merge, and both grow more consequential as AI portfolios expand.
- Understand what the current inventorship rule actually requires. The USPTO's Revised Inventorship Guidance for AI-Assisted Inventions, 90 Fed. Reg. 54636 (Nov. 28, 2025), rescinded the February 2024 guidance in its entirety and withdrew the practice of applying the Pannu factors to AI-assisted inventions. There is "no separate or modified standard": AI is an instrument, "analogous to laboratory equipment, computer software, research databases," and the question is whether a natural person conceived the claimed invention under the traditional standard, meaning a "definite and permanent idea of the complete and operative invention." Pannu now governs only the allocation of inventorship among multiple natural persons. AI still cannot be named as an inventor, because inventorship is limited to natural persons. Thaler v. Vidal, 43 F.4th 1207, 1212 (Fed. Cir. 2022), cert. denied (2023). Neither USPTO guidance nor any rule requires an applicant to document AI use or to memorialize human conception. The reason to keep records is evidentiary, and the guidance supplies it: conception "turns on the ability of an inventor to describe an invention with particularity," and "[a]bsent such a description, an inventor cannot objectively prove possession of a complete mental picture of the invention at a later time." Contemporaneous records of prompts, data selection, and refinement of model output are how that burden gets carried if inventorship is later contested. As of August 2026, no court or PTAB decision has invalidated a patent on an inventorship theory premised on AI use. The exposure is structurally available and untested.
- Revisit the portfolio on a schedule. Features like orchestration, security, optimization, and deployment frequently mature after an initial filing. Periodic invention review identifies continuations, new filings, defensive publications, and subject matter better left as a trade secret. The benefit-claim fee tiers make the timing of that review a budget question as well as a legal one.
Looking Ahead
The AI patent surge WIPO documented is unlikely to slow any time soon. As foundation models become commodity infrastructure, differentiation shifts to the technologies that make AI efficient, reliable, secure, explainable, and deployable. Those are implementation-layer technologies, and they sit where most enterprise value in this cycle will be created and where the prior-art pool is thickening fastest.
Notwithstanding the volume of filings, the harder problem for technology companies and the counsel who advise them is that the institutions governing AI patents, at least in the U.S., are diverging. The Patent Office has spent the last year lowering the eligibility bar for machine learning claims, through the August 2025 examiner guidance, the precedential designation of Ex parte Desjardins, and MPEP revisions that recognize training-method and parameter-adjustment improvements as eligible. The Federal Circuit has spent the same period holding that generic machine learning applied to a new data environment is abstract, and the Supreme Court declined to revisit it. Congress is unlikely to close this gap any time soon: the Patent Eligibility Restoration Act received its first full Senate Judiciary Committee hearing on July 14, 2026 without a vote, the Senate is in recess until September 11, and the bill's lead sponsor across four Congresses leaves the Senate at the end of this term.
An AI patent is therefore easier to get and no easier to enforce. That gap is where portfolios built on volume are fraught, and portfolios built on concrete implementations are more likely to persist. And recent Federal Circuit decisions on a single patent make the point better than any general observation: claims of U.S. 10,755,409 to Dental Monitoring were held ineligible on July 7, while others were rescued from a prior-art challenge on August 10, because the challenger's prior art reference did not fully support its own provisional filing date. Eligibility and prior art are separate fights, and robust disclosure is more likely to overcome both.
For IP counsel, the near-term work is unglamorous and specific: audit pending applications for claims that recite an outcome without a technical arrangement; include sufficient implementation details (including contemplated alternatives) in the anchoring application, rather than waiting until design freeze, and supplement these details with later filings as needed; re-examine any invalidity position that relied on the Board's Penumbra framework; and decide deliberately which advances are worth publishing, which are worth claiming, and which are worth keeping secret. The applications being drafted this quarter will become the prior art of 2028, and they are being examined under one standard and litigated under another.