What are the major IP issues driven by AI?
The rapid advancements in AI systems and widespread availability of AI tools have raised many challenges to some of the most fundamental tenets of IP. These are summarized below. Later, this guide addresses them in greater detail, with a specific lens on business concerns.
Copyright owners vs. GenAI developers
GenAI models require vast amounts of training data. For example, OpenAI’s original GPT-3 model was trained on 570 gigabytes of text. In some cases, large portions of these training data sets were scraped from the internet and contain copyright-protected text and images. This has raised issues of potential copyright infringement. As such, there are two main conflicting viewpoints regarding the balance between the interests of copyright owners and the need for GenAI developers to use copyrighted works to train their models.
The developers of GenAI models sometimes argue that the models do not keep a permanent copy of any copyright-protected work and only access the information contained in the materials temporarily: in this way, they argue that the use of copyrighted training data sets is like a human learning by reading books or viewing artworks. When a copy is made, either temporarily or permanently, developers argue that those copies fall under existing copyright exceptions and limitations, and therefore no infringement occurs.
Many copyright owners, on the other hand, assert that unauthorized use of their works as part of these training data sets is copyright infringement. Several lawsuits have been initiated by copyright owners against AI developers, but clear answers to these complex legal questions are unlikely to emerge in the immediate future. A particular point of contention is that there is no easy way to assess how much a single work contributed to training an AI model and how to potentially compensate the copyright owner. Also, copyright registration is not compulsory in many jurisdictions, making the identification of copyright works difficult at times.
Copyrighting GenAI outputs
GenAI can produce an extensive array of creative outputs, including text, audio, video, and images. There is an ongoing international discussion on whether these AI-generated outputs should benefit from copyright protection.
For example, a high-profile case in the United States of America involved an application for copyright protection for a work entitled "Théâtre D'opéra Spatial", which was created almost entirely using the AI image generator Midjourney. This application was rejected on the basis that the human creative input was de minimis. However, in a more recent case centered around a work called "A Single Piece of American Cheese, the author was able to demonstrate sufficient human intervention combined with AI tools. The US Copyright Office was satisfied that the work met the requirements for registration.
The "Théâtre D'opéra Spatial" was generated in 2022 by Jason M. Allen using the AI image generator Midjourney.A landmark case that came before the Beijing Internet Court in 2023 concluded that an image generated by Stable Diffusion (another AI image generator) counted as an original work because the human prompts required to create it amounted to sufficient human creativity. This decision contends that GenAI can be used as a tool for humans to fuel creativity, and that if a human creator is using GenAI as a tool while also making a significant creative contribution, the resulting work could potentially be awarded copyright protection. How to measure such contributions and establish a threshold remains an open question.
AI and inventorship
Can an AI system hold a patent? Can AI be credited as the inventor of a new technology or tool?
DABUS (“Device for the Autonomous Bootstrapping of Unified Sentience”) is an AI system created by Dr. Stephen Thaler that reportedly conceived of two inventions without human input. Thaler filed patent applications for these inventions, listing DABUS as the inventor, around the world. Most patent offices have rejected the DABUS applications on the grounds that patent laws require the naming of a human inventor. South Africa’s patent office became a notable exception by granting the patent, albeit without examination. Beyond these legal rejections, many computer scientists remain skeptical that current AI systems are truly capable of inventing autonomously.
What counts as a human contribution?
While the concept of a fully autonomous AI invention remains contentious, policymakers still must engage with the reality that AI is playing an increasingly large role in the inventive process. This raises critical questions, such as:
When AI is used in the inventive process, what should be the threshold of human contribution for a patent to be granted?
What kind of human contributions count as part of the inventive process, as opposed to the mere use of a tool?
Should the AI model’s developers and data providers have entitlement to patents arising from inventions made using their systems?
Each question carries widespread implications for innovation incentives, patent validity and the fundamental purpose of the patent system.
Disclosure and transparency concerns
Patent law requires that applications name an inventor and disclose the invention in enough detail for others to reproduce it. However, they do not require disclosure of how the invention was made. This means that the role of an AI system in the inventive process can potentially remain hidden both from patent examiners and the public.
If jurisdictions decide that the extent of AI involvement is material to patentability, and if human contribution thresholds matter for granting patents, then the current disclosure framework is insufficient. Patent examiners would need mechanisms to ascertain the respective contributions of humans and AI systems when evaluating applications. Without such transparency, whatever standards that emerge for AI-assisted invention cannot be applied consistently, creating uncertainty for both patent applicants and examiners.
If there are no changes to the current IP laws, and the patent law thresholds for human contribution are high, will the lack of patent protection for AI-generated inventions lead to an increased reliance on trade secrets protection? This, in turn, could drive major transparency concerns and compound the “black box” problem of AI. Would investment in AI be disincentivized? Would there be grounds for patent revocation if a human were wrongly named as an inventor of an AI-generated invention?
Potential policy solutions
Potential policy solutions include:
Exploring joint inventorship between AIs and humans
Removing the requirement to name an inventor at all
Naming the person with the closest connection to the AI as the inventor
Naming a human inventor but requiring disclosure of the involvement of AI
A brand new IP framework for AI-generated inventions
Whatever the outcome, policymakers will need to go back to basics, with analyses of the economic and social purposes of the patent system, i.e., to incentivize human creativity and the disclosure of new inventions by granting inventors the exclusive right to exploit their invention for a set period.
What are current major legislative and other developments around IP and AI?
China
China has taken a pragmatic and increasingly permissive approach to AI and IP, and has been implementing comprehensive AI governance through sector-specific regulations.
In November 2023, in a landmark case, the Beijing Internet Court ruled in an infringement lawsuit that an AI-generated image is copyrightable and that a person who prompted the AI-generated image is entitled to the right of authorship under Chinese copyright law. A similar decision was made by the Changshu People’s Court in March 2025. The court considered that an image created through AI can be an original intellectual achievement as it involves making design choices, writing prompts, setting parameters and selecting the final output.
"Spring Breeze Brings Tenderness" was created in 2023. The disputed image in Li v. Liu was generated by Li Yunkai using the AI image generator Stable Diffusion.On December 31, 2024, China's National Intellectual Property Administration issued “Guidelines for Patent Applications for AI-Related Inventions”. These guidelines emphasize that the inventor(s) listed in a patent document must be a natural person(s). In September 2025, “Measures for the Labeling of Artificial Intelligence-Generated and Synthetic Content” came into effect, mandating explicit and embedded labeling of all AI-generated text, images, audio, video and immersive virtual scenes distributed online within China.
European Union
The European Union's AI Act is a comprehensive regulatory framework that became law on August 1, 2024, with a phased implementation approach. It placed a ban on AI systems posing unacceptable risks, starting in February 2025, with codes of practice applying nine months after entry into force and rules for general-purpose AI models applying 12 months later.
The AI Act has two provisions that address copyright: Article 53(1)(c) and (d). The first requires “general purpose AI” developers to comply with copyright law and the conditions set out in the Directive on Copyright and Related Rights in the Digital Single Market to allow text and data mining. In essence, this means that copyrighted materials can be used to train AI models so long as right holders do not express their refusal. The second provision requires that AI developers report on the data sets used to train general-purpose AI models. This creates transparency obligations that could expose potential copyright infringement in training data sets.
In February 2026, the European Parliament’s Committee on Legal Affairs adopted a report addressing the intersection of generative AI and copyright law, signaling its intent to take a stronger stand in favour of copyright holders.
Japan
Japan was the first country in the world to explicitly exempt text and data mining from copyright liability. Notably, unlike similar exemptions in other countries, this applies to both commercial and non-commercial uses. This was done to encourage the growth of Japan’s digital economy. Since then, Japan has expanded exemptions for AI-based activities, with new laws exempting the following activities from copyright infringement:
certain uses of copyrighted works for machine learning;
the making of incidental electronic copies of works; and
the use of copyrighted works for data verification.
Japan has maintained this relatively permissive approach to AI training data, though it may be moving toward a more nuanced stance as litigation increases globally.
United Kingdom
The United Kingdom has protections for "computer-generated works," a somewhat unique form of copyright available in the United Kingdom, Ireland and a few other jurisdictions. Simply put, these rules give copyright protection to works created entirely by computers with no human involvement. The rule solves the "who owns it?" problem by naming the human who created the computer system that generated the work as the copyright beneficiary.
On the surface, these laws may seem like they solve the conundrum of copyright protection for AI-generated works. However, this approach remains fundamentally at odds with copyright laws worldwide, notably the Berne Convention. These are designed around the core assumption that creators are humans who benefit from protection. The United Kingdom’s approach goes against the basic "author-centric" nature of copyright law, and so functions more as a problematic workaround rather than a proper solution to the AI authorship challenge
In December 2024, the UK Government published a consultation on “Copyright and Artificial Intelligence” that proposes, among other things, to reform protections for computer-generated works, which could have significant impacts for AI. Following these consultations, in March 2026, the Government published its Report on Copyright and Artificial Intelligence. This report made no concrete proposals, abandoned some its previous preferences (such as text and data mining exceptions and right holders opt-out), and overall adopted more of a wait-and-see approach to policymaking in this area.
The UK Data (Use and Access) Act 2025 does not impose immediate legal obligations on AI developers to secure consent from right holders, but strongly signals the potential emergence of a formal framework for AI and IP soon.
United States
In February 2025, in Thomson Reuters v.. ROSS, a US federal court issued the first major AI copyright decision of 2025, upholding the plaintiff’s copyright infringement claim and rejecting the defendant’s fair use defense. It is worth noting that the AI tool in this case was not a GenAI model writing new content, but rather a search engine style model that functioned as a directly competing product to the copyrighted material it was trained on. Since this case, other decisions in the US (notably Kadrey v. Meta and Bartz v. Anthropic) have upheld the fair use defense.
Executive Order 14179, issued in January 2025, reoriented US AI policy by revoking previous directives emphasizing data protection and transparency and aiming to eliminate federal policies perceived as impediments to innovation. A subsequent order in December 2025 was aimed at blocking states from enforcing their own AI regulations.
The US Copyright Office published a three-part report on Copyright and Artificial Intelligence. The third part, on GenAI training, was released as a pre-publication version in May 2025, and essentially concluded that legislation was premature at this time. Instead, it recommended that fair use doctrine continue to be applied on a case-by-case basis, with licensing as a fallback where fair use does not apply.
Cross-jurisdictional concerns
When an AI system potentially infringes on copyright and the case reaches a court, it raises some challenging practical issues rooted in private international law. The core question here is: does the court have the competence to hear the case?
This is a complex issue that involves looking into not only the intricate corporate structures of the companies creating AI products (often spanning several jurisdictions) but also where complex GenAI models are trained, refined and deployed (again, often in several jurisdictions).
The EU AI Act has tried to solve this in Recital 106. It states that:
“providers of general-purpose AI models should put in place a policy to comply with Union law on copyright and related rights.”
It goes on to say that:
“any provider placing a general-purpose AI model on the Union market should comply with this obligation, regardless of the jurisdiction in which the copyright-relevant acts underpinning the training of those general-purpose AI models take place.”
In plain language: “If you want to sell your AI model in the European Union, you must follow EU copyright law, regardless of where the model was trained.” The underlying logic here is that no company should gain an unfair advantage in the European market by training their AI in countries with weaker copyright protections.
However, Recital 106, while sweeping, is not legally binding and ignores the fact that copyright law is governed by the principle of territoriality. In practice, if the text and data mining activities take place outside of the European Union and the model is placed in the European Union market only afterwards, then the provider will be compliant with the recital. On the contrary, should any of the training stages happen within European Union territory, such as web-scraping, then the recital kicks in and the provider will have to respect European Union law.
Despite these provisions being narrowly focused on copyright law and the European Union, it would not be surprising to see this approach exported in the future in other areas of IP law and beyond the territory of the European Union. The goal is to avoid a "race to the bottom" mentality, where AI companies shop for the most permissive jurisdictions to develop their models while still accessing stricter markets for sales.