The typical 18-month publication delay for filed patent applications means that the increased activity in the space beginning in late 2022/early 2023 is now becoming visible. As a result, there has been a significant and sustained increase in published patent applications within the GenAI space in 2024 and 2025, as noted in WIPO’s report on patenting trends in GenAI over 2024/25.

In particular, during 2024 and 2025, over 56,000 new GenAI patent families were published, exceeding the total number of patent families published over the preceding ten years.

In addition to the substantial increase in published patent applications in 2024 and 2025 compared with the previous decade, data from 2025 indicates continued growth from 2024.

In 2025, the number of published patent families in this space increased by 100%, from 18,862 in 2024 to 37,808, far exceeding the long-term average annual growth rate of 43% since 2014.

What is being protected?

GenAI often makes use of large language models (LLMs), which are trained on billions of input sources (books, text and websites) to learn patterns in human language. The number of published patent families for large language models (LLMs) in 2024/25 has substantially exceeded that of any other GenAI model, indicating a strong preference for this model type that aligns with commercial growth in the sector. LLMs have seen a substantially greater increase in published patent families than other GenAI models. Notably, since 2023 LLMs have overtaken generative adversarial networks (GANs) as the model category with the largest volume of published patent families. From 2023 to 2025, the number of LLM patent families as a percentage of GAN patent families increased from 37% to 270%.

The broad applicability of an underlying architecture to complete a large variety of tasks (compared to GANs, which usually have specific use cases) is a reason why LLMs are so useful but also presents issues with patentability, with patent offices often requiring a technical application to be clearly defined in patent applications. Meanwhile, the output of GenAI may typically be considered a presentation of information and the underlying process may be considered a mathematical method. These issues can create significant difficulties in patenting innovations in LLMs.

In addition, many patent offices require details of LLM training to be provided to meet sufficiency requirements. While patent offices do not require applicants to provide access to the proprietary dataset (which may be one of the most valuable aspects of the model), applicants must characterise the dataset sufficiently for a skilled person to curate an equivalent dataset, with details of any selection rules used in training and parameters used. The EPO is particularly strict on this sufficiency requirement, requiring high levels of detail to meet the bar, as illustrated in T 0161/18 and T 1669/21.

Within this still-developing field, case law is continuing to evolve, and the large influx of LLM-related published patents may lead to further significant decisions providing clarity on the rules around patenting such inventions.

Applicants must determine whether the risk of disclosing details of their training dataset is a worthwhile trade-off for gaining patent protection. While LLMs present patentability challenges, with appropriate claim construction, technical features of LLMs may obtain patent protection.

The substantial increase in patent filings from early 2023 indicates that many proprietors consider the risk of disclosure worthwhile.  

Who is filing, and where?

Of the top 25 patent owners by number of published patent families between 2014 and 2025:

• three owners are Japanese (including the largest owner SoftBank); 
• 15 are from China (including those ranked 2-6);
• five are from the USA;
• one is from the Republic of Korea; and
• one is from Germany (the only European entity in the top 25). 

The top 20 owners are responsible for 20% of the patent families published in the 2024/25 period.

Japanese company SoftBank, a significant investor in OpenAI, has substantially increased its published patent families in GenAI from only seven at the end of 2023 to 2,978 at the end of 2025, pushing it into first place as the largest individual owner of published patents in the space.

However, while shareholders of OpenAI such as SoftBank may be focusing heavily on the GenAI patent landscape, OpenAI has yet to publish a meaningful number of patents, with only 35 patent filings globally as of late 2025. This limited patenting activity may be linked to a desire to protect its intellectual property via trade secrets.

The bigger picture

The number of international patent families remains a small proportion of GenAI filings at the end of 2025, with the majority of filings remaining domestic at this stage.

As patents confer negative rights, allowing their owners to prevent competitors from acting, the key jurisdictions to file in are those where competitors are located. With most innovation in GenAI being limited to a small number of countries (namely China, USA, Japan and the Republic of Korea), it is not surprising that filings have mostly been directed at these locations. Filing domestically can be a key part of defensive strategy, ensuring that competitors cannot crowd an existing entity out of the space.

We also note that another motivation for filing domestically can be to take advantage of Patent Box type systems, which can offer tax incentives for entities based in that jurisdiction with granted patents. Since GenAI models can be costly to develop, it may be advantageous to file domestically and benefit from such systems.

Related reading: Patent Box FAQ Guide

From understanding what the Patent Box is to calculating potential tax savings and navigating the application process, our introduction to the UK Patent Box addresses the most common questions businesses ask about the scheme. Learn how patents can support both innovation and tax efficiency, explore the qualifying criteria, understand the typical patenting timeline, and discover how patent-protected inventions can help reduce corporation tax. The guide also explains how our specialists can work alongside accountancy teams to help maximise available benefits. Read the full Patent Box FAQ guide online:
https://dycip.com/patent-box 

WIPO projects that this filing pattern will shift over the next few years as the field develops and more patentees turn their attention towards international filing strategies and expansion. This suggests that commercialisation in the GenAI space remains at an early stage, and that further growth and expansion are still to come and may occur rapidly.

The rapid rise of SoftBank to the top of the list of published patent owners indicates that in this young and rapidly evolving field, it is still possible for new competitors to innovate within the market.

For businesses operating in the GenAI field, the rapid growth in patent publications is a clear signal that the competitive landscape is becoming more crowded.

Companies should review whether their GenAI-related implementations might contain differences that they may wish to patent, particularly where those features improve performance, training efficiency, data handling, user interaction, system architecture, security or deployment in a specific technical environment.

Drafting patent applications for such innovations can, however, be challenging, particularly in Europe and the UK.

In practical terms, companies should capture inventions early, before public disclosure; coordinate patent strategy with product roadmaps and target markets; and avoid assuming that GenAI inventions are unpatentable purely because they involve software or AI. The strongest cases are likely to be those framed around a specific technical contribution rather than the mere use of a generative model.

Useful links

• WIPO, Patent Trends Update in Gen AI, 2026: https://dycip.com/wipo-patent-trends-ai-2026 
• D Young & Co guide to The Patent Box: https://dycip.com/patent-box  
• Issues concerning the grant of machine learning patents at the EPO: 
  https://dycip.com/epo-ml-t0161-18-t1191-19 
• EPO legal texts, case law of the Boards of Appeal, 6.1.1 artificial neuronal network: https://dycip.com/epo-case-law-ann  
• EPO decisions of the Boards of Appeal, T 0161/18 (Equivalent aortic pressure/ARC SEIBERSDORF): https://dycip.com/epo-t0161-18  
• EPO decisions of the Boards of Appeal, T 1669/22 of 05.12.2024: https://dycip.com/epo-t1669-22