AI has become a controversial issue in recent policy discussions. While some may point to its negative impact on entry-level jobs, others point to how AI is a tool that undermines creativity in art and writing. LLMs (Large Language Models) like ChatGPT and Claude rely heavily on a corpus of information, much of which exists in the public domain, to generate information.

Take the issue of visual art. If you are an artist with your own style of creating an image. Imagine if now, AI can take your art, analyze the patterns, and generate a variation of it for the user? Does the original artist not deserve a certain payment or commission for being the inspiration of that output? Are AI systems effectively stealing from existing artistic work, whether written or visual, to generate output?

Fair Use – The Four Factors That Decide Infringement 

In order to decide whether AI infringes on Copyright law, we must first understand what “fair use” constitutes under the Copyright Act. There are four “non-exclusive” factors that courts weigh to determine whether AI systems break Copyright laws:

“1. the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;

2. the nature of the copyrighted work;

3. the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and

4. the effect of the use upon the potential market for or value of the copyrighted work” (Congressional Research Service, 2025)

Out of this, factor 4 is often most important to the courts, which looks at whether the use of the copyrighted work affects the market for this work and how it does so. The danger is, if copyrighted work is used to dilute the market, then the original work’s value is displaced. And so, the broader danger with AI, is that outputs can dilute the markets for the works they learned from. Market substitution or displacement, which is what factor 4 is about, is where AI could pose the highest risk of violating copyright law.

Concretely, if there is a specific style or niche in artistic expression, which an artist owns, and AI replicates variations of them that are not “substantially” different, the market for that niche gets displaced and the value of that work weakens. 

A Divided Bench – The Limits Of Precedent 

Unfortunately, courts are not united on how to treat AI. On the one hand, in Bartz v. Anthropic, the court ruled in 2025 that training a model on lawfully acquired books was fair use, yet downloading pirated copies to do so was not (Congressional Research Service, 2025). On the other hand, in Kadrey v. Meta Platforms, Inc., another judge ruled that Meta did not violate fair use when its generative AI system used pirated books for its training data. However – unlike in Bartz – in Kadrey’s case, the court defined market dilution more broadly under factor 4, where AI would create market harms by generating items within the same topic or issue area.

In Bartz, Anthropic had to pay a hefty sum for training on pirated data. But, on the issue of market dilution, the court ruled that Claude AI outputs did not displace or dilute demand for books. So if Claude had simply not used pirated books in its sources, the argument is that AI does not infringe on the intellectual property of the books it uses to generate written content. 

In Kadrey, the court argued that Meta can use pirated books in their training data, and that this does not infringe on the intellectual property of authors. In this way Kadrey was more lenient than Bartz. However, on the issue of market dilution, the court held a broader definition, while critiquing the plaintiff for not making use of that argument.

What does this all mean? Under the Copyright law, AI generative output is largely seen as transformative enough, and the burden of proof to show that these outputs break Copyright law is difficult to meet. What is also true is that court rulings have not shown a consistent framework to apply the four non-exclusive factors used to determine infringement of Copyright law with regard to AI.

From a legal perspective, the current law – and how it is applied – does not provide a clear answer to the question of whether AI is “stealing” from content creators whose data is used for training and output generation.

Relying On Courts Is A Fool’s Errand

In my opinion, relying on courts to match existing law, which was not made in the advent of AI, to interpret how fair use should be implemented with regards to AI, is a fool’s errand. Courts are given discretion to analyze whether fair use applies, based on a set of core factors and precedent. The problem? AI is a relatively new technology, with major LLMs like ChatGPT only becoming popular in 2022.

The law needs to be updated to further define, in particular, what market “dilution” looks like, particularly with regards to AI generation of content. Because of the specific nature of AI, being able to quickly take a piece of content and generate a variation, I would argue that this technology has the potential to disturb niches quite easily. 

Regulatory Framework – Compensation Versus Cost

\A more viable path is licensing copyrighted works for AI training, a framework proposed by The Author’s Guild. By ensuring that AI systems buy licenses from authors and artists, to then train on their data, the original creators of content get paid and AI systems are trained via rich datasets without “stealing” data. This will not solve market dilution, but it will create a system where original artists are paid for the contributions they make to improve or inspire AI outputs.

However, this framework will also contend with counterpoints, one being the affordability of AI products. Most LLM models like Claude or ChatGPT provide outputs for free, unless you wish to get more output per day, wherein you pay a monthly subscription. If the law changed to require AI companies to pay for licenses and use copyrighted works in their data, would this not increase the cost of the product? I think it will, and this is a trade-off worth considering.

Do we regulate AI to require payments to use copyrighted work, thereby helping original artists, or do we keep regulations the same, to ensure the cost of AI products remains manageable?

As this debate unfolds in America, we must also consider that whichever decision we make here, it will impact the AI race against China and other rivals. This is another area where regulatory consideration matters.

In my view, relying on the courts to weigh in with existing, old Copyright law, is not sufficient, given the recent rulings by courts show an inconsistent application. The current law provides too much discretion to the judicial branch on the application of AI at a time where Congress must take charge. Congress must narrow the definitions in a manner where AI’s position in Copyright law is more well defined, establishing metrics of fair use, specific to AI systems.

On the question of how far these new regulations should go, and where this balance can be found, that requires further discussion.