Generative AI Patent Trends 2026: What the Filing Data Shows, and What It Doesn't

Published generative AI patent families roughly doubled in a year, from 18,862 in 2024 to 37,808 in 2025, according to WIPO’s 2026 update. That follows growth from fewer than 800 in 2014 to about 14,000 in 2023. The surge is real, but the counts trail real filing activity by about 18 months, count families rather than individual applications, and say nothing on their own about which inventions matter commercially.

What the 2026 data shows

The volume step-change

WIPO’s 2024 landscape report counted 54,358 GenAI patent families from 2014 to 2023, with growth averaging about 45% a year since 2017. The 2026 update found 18,862 families published in 2024 and 37,808 in 2025. WIPO’s July 2026 announcement notes that the two years together exceed the entire previous decade. GenAI’s share of all AI patenting rose from 6.1% in 2023 to 8.7% in 2025.

The model mix flipped

Generative adversarial networks (GANs) used to be the largest GenAI model category in patents. By 2025 large language models outnumbered GANs nearly three to one, with diffusion models third. LLM-related families went from 881 to more than 14,100 in two years. Text-related filings more than tripled. Image and video remain the largest output mode, with about 40,000 families from 2014 to 2025, over 13,800 of them published in 2025.

New kinds of applicant

SoftBank tops WIPO’s 2025 owner ranking with 2,985 families, ahead of Tencent, Ping An, Baidu and the Chinese Academy of Sciences. State Grid Corporation of China, Zhejiang University, Alphabet/Google, Microsoft and IBM complete the top 10. WIPO’s September 2026 analysis reads the arrival of utilities, Nvidia, Inspur and Bosch as GenAI spreading into established industries. China remains the dominant origin. The United States and Japan are among the fastest growing.

The gap in absolute terms is wide. For 2014–2023, WIPO’s 2024 report counted more than 38,000 GenAI families with inventors in China and around 6,300 with inventors in the United States. The United Kingdom had 714 and Germany 708. Inventor location is not the same as owner location, though, and a family counts once however many countries it reaches.

Science runs ahead of patents

The same WIPO report found GenAI scientific publications rising from around 100 in 2014 to more than 34,000 in 2023. That is more than double the patent family count for the same year. Papers usually become public sooner, while patent applications generally stay unpublished for about 18 months. For an early read on where filings will go next, the research literature is often the better leading indicator.

The view from IFI Claims

IFI Claims measures AI differently, by documents rather than families. Its May 2025 analysis found nearly 190,000 AI applications worldwide over a 14-month window. GenAI made up 28% of them, and 24% of AI applications at the USPTO. Over 10 years, IFI measured compound annual growth of 38% for AI grants and 58% for GenAI grants. Agentic AI was already 7% of US AI applications.

Meanwhile, the overall US system was not growing. IFI’s 2025 report counted 323,272 US grants, down 0.24%, and 393,344 applications, down 9%.

What the data doesn’t show

WIPO’s own commentary lists the limits clearly.

• Quality and value. Patent counts don’t establish quality, commercial success or market share.
• Purpose. Filing data can’t tell you why an applicant filed: to protect a product, to build a defensive position, or to signal to investors.
• Timing of the work. SoftBank’s portfolio was largely filed in 2023 and published in 2025. The 2025 count reflects a concentrated filing wave, not a single year of research.
• Geographic reach. Only about 9% of GenAI families published in 2025 appeared in two or more jurisdictions, against about 15% for all patent families. Many are single-office filings. Some may add countries later.
• Model coverage. WIPO assigns patents to model types from titles, abstracts and claims, and not every GenAI patent can be assigned to one of the five core models.

Patent counts are not market share

One contrast in WIPO’s September analysis makes the point. SoftBank holds 2,985 GenAI families. OpenAI had 35 patent filings worldwide as of late 2025. WIPO reads SoftBank’s portfolio as consistent with a broad platform and infrastructure strategy, and OpenAI’s as more selective and product-focused. Neither count tells you who leads the market. Filing behavior varies by company, by country and by business model.

How to read AI patent counts critically

1. Check the date basis

Applications are generally published about 18 months after the earliest filing date. Count by priority or filing year and the most recent 18 months look artificially low: a cliff that is really a lag. Count by publication year, as WIPO’s figures for families published each year do, and you get no cliff. Instead the chart shows filings made 1–2 years earlier.

2. Check the counting unit

A patent family groups the filings for one invention across offices. WIPO counts families. IFI counts application documents and grants. One family can produce a US application, a Chinese application, a European application and several grants. Never put a family count and a document count on the same axis.

3. Check the definition

“AI” and “GenAI” are search queries, not official categories. IFI’s AI set covers machine learning, neural networks, natural language processing and biological-model computing. WIPO’s GenAI set is built for its landscape. Two credible sources can differ widely and both be right about their own definitions.

4. Watch the base effect

IFI’s fastest-growing technologies list for 2025, ranked by five-year compound growth in US applications, is led by electrolysis and metal-recovery codes, not AI. Small, emerging fields post high growth rates more easily than large ones. A field that is already large can add thousands of filings a year and still miss a growth ranking.

5. Look past the totals

The decision-relevant signal sits below the headline number: who is filing, in which subgroups, with what claim scope, and in how many countries.

Using AI patent trends in decisions

Different readers need different cuts of the same data:

• R&D leaders: which assignees are filing in the CPC subgroups closest to your roadmap, and whether their pace is rising.
• Investors: whether a target’s GenAI families are single-office filings or extend to the markets its business plan depends on.
• Licensing and business development teams: which incumbents outside software (utilities, industrial firms, device makers) are now filing, and may prefer to license technology rather than build it.

For an R&D team, the useful question isn’t “how many GenAI patents exist” but “who is filing near our roadmap, and how fast?” That calls for a landscape scoped to your technical problem, with assignee normalization and claim-level reading of the families that matter. After that comes a watch that flags new publications without burying your team. Our guide to patent monitoring without alert fatigue explains how to set one up. If you’re writing a grant proposal, the same data strengthens a commercialization section; see patent data for SBIR and STTR proposals.

How Meta Dragons can help

Filings grow like coral: slowly, then everywhere. Our Patent Landscape & Whitespace service maps a GenAI subfield by assignee, CPC subgroup and filing trend in 3–6 weeks. It states every query, date basis and counting unit in the audit trail. Competitive Patent Intelligence then profiles the filers that matter to you and keeps watch with analyst-written digests.