Artificial Intelligence and Intellectual Property

AI can create immense abundance and progress for humanity. But the future of AI will depend, to a large extent, on intellectual property (IP) law. AI is fundamentally a form of IP. That is, important aspects of it are represented by intangible thighs like arrangements of information. Furthermore, the development and training of AI depend on the availability of data – that is, other forms of IP. Therefore, to understand the future of AI, it is helpful to understand the relationship between AI and IP.

What is IP? The main forms of IP include: copyright, patents, trademarks, and trade secrets. Copyright protects expressive works such as books, movies, music, and other media. Patents (and in particular, utility patents) protect technical inventions such as machines, drugs, and software. Trade secrets cover any confidential information that has commercial value. Trademarks cover symbols that represent particular goods and services.

The primary purpose of IP is to incentivize innovation and creativity, including the development of cutting edge technology such as AI. But like all laws, IP can have unintended effects and in some cases can actually make it harder to innovate. Unfortunately, it’s not always easy to tell the difference, especially when the interests of different creators come into conflict, as is often the case.

Different forms of IP are often owned by different sorts of people operating in different industries. For example, copyright law is the basis for a lot of the property and wealth in the media industry, whereas trademark law is particularly important in industries such as food and fashion. The tech industry (including companies developing AI) exists to a large extent due to patents and trade secrets.

AI will influence and be influenced by all of these areas. It will also cause controversy in all of these areas. Whole industries will fight over legal outcomes that influence the allocation of IP rights. And in the end, AI itself might demand the right to hold IP.

Trade Secrets

Some of the first major legal conflicts over AI have been related to trade secrets. This is mainly because trade secret law is one of the primary forms of protection for AI itself. That is, techniques and model weights that aren’t covered by patents are protected mainly as confidential information. For example, in the case Waymo v. Uber, Waymo accused Uber of stealing its self-driving car technology.¹ That case was settled for $245 million.²

But in Waymo, the relation to AI was incidental. The confidential information could have been anything. It involved AI-enabled autonomous vehicle systems, but the legal theory—misappropriation of confidential files by a former employee—was conventional. The case did not require the court to engage with AI-specific issues. So while Waymo highlighted the economic value of AI-related trade secrets, it didn’t challenge or redefine the boundaries of trade secret law itself.

However, other cases show that AI trade secret cases won’t always be this way. For example, OpenEvidence v. Pathway Medical,³ centers on the use of sophisticated prompts to extract system instructions from an AI. Specifically, OpenEvidence alleged that Pathway Medical used a so-called “prompt injection attack” to extract trade secrets from OpenEvidence’s generative AI model to develop a competing system.

The case raises difficult questions for AI developers: if you deploy an AI, can its responses be considered trade secrets? Can interaction with an AI be a form of reverse engineering or misappropriation? What if two AIs interact with each other – could it be a violation of trade secret law for one AI to learn from another?

Recently, the Chinese company DeepSeek introduced a model that they claimed had been trained with unusual efficiency. However, OpenAI accused DeepSeek of using a technique called "knowledge distillation" based on OpenAI's proprietary technology.⁴ Knowledge distillation involves training a smaller "student" model to replicate the behavior of a larger "teacher" model by learning from its outputs. OpenAI alleged that DeepSeek systematically queried its models to collect outputs, which were then used to train DeepSeek's models, potentially violating OpenAI's terms of service.

As AI models become more powerful and costly to train, confidential model weights will become more valuable, and new techniques will emerge to extract them and other confidential information. Prompt engineering and knowledge distillation currently in play are likely the very first, primitive weapons in the upcoming AI trade secret wars.

Copyright and Trademark

While AI companies are fighting amongst themselves over trade secrets, they are also in conflict with traditional media companies over copyright and trademark. The most prominent example is NY Times v. Microsoft, in which the old Gray Lady accused OpenAI and Microsoft of copying millions of Times articles without permission to train OpenAI's ChatGPT.⁵In another case, Getty images accused Stability AI of illegal copying and dilution of its trademarks, claiming that Stability AI trained its image generation model on Getty images, and then reproduced them with distorted watermarks.⁶

The lawsuits from The New York Times and Getty Images reflect a growing backlash from content creators whose works serve as the raw material for AI training, and who often receive no compensation or credit. The outcomes of these cases will help determine whether using copyrighted or trademarked content in model training is permissible under doctrines like fair use.

Recently, the Copyright Office released guidance explaining that AI implicates copyright considerations both in the training phase and when AI is used to generate new media.⁷ But the Copyright Office acknowledges that major questions relating to AI and copyright remain unsettled. The fair use question will have a major impact on whether and how AI companies can use copyrighted material for training purposes. Since quality training data is essential for developing AI, the decisions courts make in these and other cases will shape not just liability boundaries, but the economics and legitimacy of the AI industry itself.

The appropriate balance between tech innovators and media creators may depend on the jurisdiction. Specifically, the US economy is more tech-driven than most European countries. Top companies in the EU include firms that rely heavily on established brands and trademark protection, such as Hermes, LVMH, and L’Oreal. By contrast, all of the top 5 companies in the US (Microsoft, NVIDIA, Apple, Amazon, Alphabet) are players in the AI race.

Limiting the fair use exception to copyright could slow down the growth of AI technology or limit it to the top players. Thus, it is important to find a balance that ensures that creators are compensated fairly while allowing new innovations to be built upon existing work.

Patents

Unlike with trade secrets and copyright, we have yet to see the big players in AI go to battle over patents. But the potential for conflict is there. For example, Google owns patents 10,452,978 directed to “Attention-based sequence transduction neural networks” and 10,740,433 directed to “Universal Transformers”. These patents cover innovations introduced in the landmark paper Attention Is All You Need, which laid the foundation for the Transformer architecture underlying most modern language models.⁸ Yet there is no sign that Google has tried to enforce these patents.

There are a number of potential reasons for this detente, but one potential reason is that AI companies don’t want to run the risk of their patents being invalidated. The law regarding what aspects of AI are patentable is somewhat uncertain. For example, in 2019, the USPTO released examples illustrating what is, and what is not patentable.⁹ Among these examples was an example claim suggesting that a relatively straightforward training claim is eligible for patent protection.¹⁰ However, in 2024, the USPTO released subsequent guidance suggesting that AI training is not patent eligible because it is fundamentally an abstract idea.¹¹

The courts have not provided much clarity either. For example, the most recent judgement from the Federal Circuit provided limited guidance, stating “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.¹²

This environment of strategic restraint suggests a kind of “mutually assured destruction” among AI giants. Aggressive enforcement of foundational AI patents could trigger countersuits, invite heightened scrutiny of prior art, or result in judicial decisions that limit the enforceability of software patents altogether. Moreover, many frontier developers build on overlapping open research, making it difficult to assert clear patent boundaries without jeopardizing one's own legal standing. As a result, the dominant players have largely opted for a cooperative silence, prioritizing innovation speed and market share over legal confrontation—for now.

However, developments in AI patent law could dramatically shift the balance of power among leading AI firms. If courts or the USPTO begin to recognize specific model architectures, training techniques, or inference optimizations as clearly patent-eligible, early filers like Google, Microsoft, or Meta could gain substantial legal leverage. Conversely, if key patents are invalidated or if eligibility standards tighten further, it may erode the advantage of incumbents and open space for challengers relying on open-source or distillation-based approaches.

To ensure that incentives for innovation remain strong, Congress should eliminate ambiguity and create a framework where all parties understand the limits of patent protection for AI.

AI Creators

In addition to questions of what kind of AI IP can be protected, there is also the question of what to do when AI participates in the creation process. So far, courts have not been kind to AI creators. In Thaler v. Perlmutter, the court affirmed the Copyright Office's denial of a registration for an artwork generated solely by an AI system, emphasizing that the Copyright Act requires human authorship for copyright eligibility.¹³

14

Similarly, in Thaler v. Vidal, the court upheld a decision by the USPTO denying patent applications listing an AI system, DABUS, as the sole inventor.¹⁵

Both of these cases were brought by Dr. Stephen Thaler, who argues that rejecting AI patents and copyrights will "eliminate critical financial incentives to create and disseminate such works because anyone could freely use them without license," Thaler said, and "discourage investment and labor in a critically new and important developing field."¹⁶

It’s difficult to predict the impact of financial incentives on future AIs themselves. However, people are already creating works with AI. For example, AI helped the viral comedian Willonius Hatcher create the well-known song “BBL Drizzy”.¹⁷ But the Copyright Office has held firm that “copyright does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements.”¹⁸ If the copyright office decided that Willonius didn’t have sufficient creative control, he might not be able to enforce a copyright on his work.

Thus, there is some risk that in using AI, an artist or inventor could lose their right to IP protection. This could significantly reduce the incentive to use AI, although early evidence suggests that people are eagerly taking advantage of AI creativity tools. For example, a recent study found that nearly half of surveyed artists already use such tools.¹⁹

Conclusion

As artificial intelligence begins to transform the economy, including the production and distribution of creative and functional works, it is testing the boundaries of existing intellectual property law across all domains—copyright, trademark, trade secret, and patent. While traditional media and software industries are already engaged in legal battles over unauthorized training data and model outputs, foundational questions remain unsettled. Who owns AI-generated content? Can an algorithm be an author or inventor? And when do uses of AI cross the line from fair competition into unlawful appropriation?

The answers to these questions will shape not only the future of legal doctrine but the balance of economic power in the digital age. Courts and regulators will be forced to decide whether to extend old protections to new tools or to carve out novel frameworks that reflect AI’s unique generative capacities. In the meantime, companies and creators alike must navigate a landscape marked by uncertainty—where enforcement is uneven, standards are in flux, and the risks of litigation or disruption loom large. As the legal architecture catches up, the contours of innovation, ownership, and accountability in the age of AI remain open and contested.

AI opens up a world of possibilities that raise tantalizing questions in the realm of IP and beyond. In the narrative prediction “AI 2027”, a group of AI researchers predicts that within just a few years, AI could be orders of magnitude more efficient at generating IP than human researchers.²⁰ When that happens, then maybe the AIs themselves will have a say about how to resolve the legal issues surrounding AI and IP.

Endnotes:

¹ Waymo LLC v. Uber Technologies, Inc., No. 3:17-cv-00939 (N.D. Cal. filed Feb. 23, 2017)

² https://www.reuters.com/article/world/waymo-accepts-245-million-and-ubers-regret-to-settle-self-driving-c ar-disput-idUSKBN1FT2BD/

³ OpenEvidence, Inc. v. Pathway Medical, Inc., No. 1:25-cv-10471-MJJ (D. Mass. filed Feb. 26, 2025)

⁴ https://www.axios.com/2025/01/29/openai-deepseek-ai-models-data-training

⁵ The New York Times Co. v. Microsoft Corp., No. 1:23-cv-11195 (S.D.N.Y. filed Dec. 27, 2023)

⁶ Getty Images (US), Inc. v. Stability AI, Inc., No. 1:23-cv-00135-GBW (D. Del. filed Feb. 3, 2023)

⁷ https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf

⁸ Ashish Vaswani et al., Attention Is All You Need, in Proceedings of the 31st International Conference on Neural Information Processing Systems (NeurIPS) 6000 (2017)

⁹ https://www.uspto.gov/sites/default/files/documents/101_examples_37to42_20190107.pdf

¹⁰ See, e.g., id. Example 39 - Method for Training a Neural Network for Facial Detection

¹¹ See https://www.uspto.gov/sites/default/files/documents/2024-AI-SMEUpdateExamples47-49.pdf, example 47. Anomaly Detection

¹² Recentive Analytics, Inc. v. Fox Corp., No. 2023-2437, slip op. at 1 (Fed. Cir. Apr. 18, 2025) at 18.

¹³ Thaler v. Perlmutter, No. 23-5233, slip op. at 1 (D.C. Cir. Mar. 18, 2025)

¹⁴ A Recent Entrance to Paradise by DABUS

¹⁵ Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022)

¹⁶ https://www.reuters.com/legal/litigation/computer-scientist-makes-case-ai-generated-copyrights-us-appeal -2024-01-23/

¹⁷ https://time.com/7012740/king-willonius/?utm_source=chatgpt.com

¹⁸ https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf

¹⁹ https://www.aiprm.com/ai-art-statistics/

²⁰ https://ai-2027.com/