AI Training and Copyright: Is It Fair Use or Infringement?

Generative AI systems are built by analyzing massive datasets filled with books, articles, images, music, and other creative works—much of it protected by copyright. These models learn from human-made content to produce convincing text, visuals, and audio that often reflect the style, tone, or structure of the originals.

This practice has triggered a growing backlash. Writers, artists, musicians, and other creators argue their work is being used without permission or compensation. They say AI companies are benefiting from years of creative effort while offering little transparency and no real way to opt out. On the other hand, developers maintain that training is technically necessary, legally sound, and different from traditional copying.

What’s unfolding is more than a fight over technology. It’s a broader debate about how creative work is treated, who has control, and how credit and compensation are handled in the rise of AI-generated content.

Is Training AI on Copyrighted Content Considered Infringement?

For many creators, publishers, and rights holders, the problem isn’t just what AI systems generate—it’s how they learn. Using copyrighted works in training datasets without permission or payment, they say, amounts to unlicensed copying.

Training a generative AI model requires vast amounts of digital content so it can analyze patterns and learn how to mimic language, visuals, or sound. Critics point out that this process often involves ingesting entire works without consent. To them, it doesn’t matter whether the final output includes exact copies. The act of copying during training itself is a violation of rights.

That distinction is central to the debate. Even if an AI never reproduces a book word for word, many argue that duplicating it during training still crosses a legal line. From their perspective, copying for analysis is still copying, and it should require permission.

The issue raises a fundamental question: can AI developers lawfully use copyrighted content to teach their systems, even if the final outputs are new? For many in the creative community, the answer remains firmly no.

How Companies Justify Training on Copyrighted Material

AI companies argue that their use of copyrighted content for training is protected under U.S. fair use law, which allows limited use of such material when the purpose is transformative and doesn’t compete with the original. They claim that because AI training analyzes patterns rather than republishing or reproducing specific works, it qualifies as a new and lawful use.

OpenAI, for example, has stated that “training AI models using publicly available internet materials is fair use,” adding that the practice is “essential to developing models that are useful, accurate, and safe.” In its view, the outputs are original creations—distinct from the material used to train them—and restricting access to training data could harm innovation and public benefit.

Many in the tech industry echo this stance. They argue that broad, diverse datasets are necessary to build high-performing models that support applications ranging from translation tools to accessibility aids and scientific research.

Still, the legal basis for this defense remains uncertain. In Talkin’ ’Bout AI Generation, James Grimmelmann, a law professor at Cornell Tech, notes that “the training of generative models is a novel use, on a massive scale, that doesn’t fit neatly into any of the established fair use precedents.” While courts have previously allowed some machine learning applications under fair use, those decisions may not extend comfortably to generative AI—especially when outputs closely mimic the style or substance of the original training material.

The Four Fair Use Factors in the Context of Generative AI

While AI companies defend model training as transformative and legally permissible, U.S. courts evaluate fair use through a four-part framework. These factors are considered together, with outcomes depending on the specific details of each case. When applied to generative AI, each factor below raises unique and unresolved legal questions. 

1. Purpose and Character of the Use

This is often the most contested factor. Courts assess whether the new use adds something with a different purpose or character and whether it is considered “transformative.” AI companies argue that training qualifies because it involves extracting patterns to build new systems, not reproducing or distributing the original work. They frame this internal use as distinct from traditional copying.

Critics disagree. They point out that generative models can closely mimic the style of specific authors, artists, or musicians. Even if exact content isn’t reproduced, the ability to replicate a recognizable voice or aesthetic raises concerns about derivative use.  For example, comedian Sarah Silverman is one of several plaintiffs alleging that ChatGPT generated passages mimicking her tone and comedic style, based on her copyrighted writing. This issue becomes especially sensitive when living creators see their work echoed without consent or credit.

2. Nature of the Copyrighted Work

This factor looks at the type of work being used. Courts generally provide stronger protection for creative, expressive works—such as novels, songs, or photographs—compared to factual or functional materials like news articles or manuals.

Generative AI models are often trained on highly expressive content, including fiction, poetry, visual art, and film scripts. Because of this, the second factor typically weighs against fair use in the AI context, even if it’s not determinative on its own.

3. Amount and Substantiality of the Portion Used

Courts consider both the quantity and the qualitative importance of the material used. AI models typically ingest entire works rather than excerpts. Developers argue this is necessary to prevent bias, preserve context, and ensure model quality.

Still, large-scale ingestion of entire texts or datasets generally weighs against fair use. Courts have scrutinized bulk copying even when done for non-commercial or technical purposes. The scope of AI training magnifies this issue and raises questions about whether such widespread use can be justified as incidental.

4. Effect of the Use on the Market

This factor looks at whether the new use harms the market for the original work or its potential derivatives. Creators argue that AI-generated content could substitute for their work, especially when it imitates a distinctive style or is used commercially. Some have already seen their creative voice echoed by AI tools. Artist Eva Toorent, for example, described visiting a gallery and recognizing pieces that mimicked her visual style. “If I’m the owner, I should decide what happens to my art,” she said.

AI companies counter that outputs serve different purposes and audiences—such as casual users generating content for fun, not as a replacement for licensed work. They also claim there’s little evidence of widespread market harm.

This issue is still playing out in court. Several economic studies are underway, but the results are mixed so far. The outcome may depend on how clearly plaintiffs can show that AI tools are reducing demand for their work or displacing future revenue opportunities.

Legal Battles Over AI Training and Copyright Law

Courts haven’t reached a clear decision on whether it’s legal to use copyrighted material to train AI systems without permission. They’re also weighing whether the content AI generates qualifies for protection under fair use. Several major lawsuits are now testing these questions. Below, we break down some of the most significant cases shaping this legal debate.

1. Getty Images v. Stability AI

In one of the earliest and most visible lawsuits, Getty Images sued Stability AI, the company behind the image generator Stable Diffusion. Getty claims that Stability scraped more than 12 million of its copyrighted photos from the internet without a license. Some AI-generated images reportedly even include distorted versions of the Getty watermark, which the company says is proof that its content was copied directly.

Getty argues that even if Stable Diffusion doesn’t reproduce the exact images, the act of copying their content during the training phase violates copyright law. The lawsuit questions whether training on copyrighted content can be justified if the final results are visually distinct. This case is especially important because it could determine whether large-scale data scraping is protected under fair use or requires permission.

2. Authors Guild v. OpenAI

Another major case involves the Authors Guild, representing a group of well-known writers including George R.R. Martin, John Grisham, and others. The lawsuit accuses OpenAI of using their books without permission to train ChatGPT. The authors argue that even if the AI doesn’t reproduce their exact words, copying full texts during the training process still infringes on their exclusive rights.

This case reflects a broader concern across the publishing industry. Many authors believe their work is being used to fuel profitable AI tools without their knowledge, consent, or compensation. The lawsuit could set a key precedent on whether creators should be paid or credited when their work is used to train commercial AI models.

3. Disney, Universal, and Warner Bros. v. Fable Simulation

In a newer legal challenge, three of the biggest names in entertainment—Disney, Universal, and Warner Bros.—filed a joint lawsuit against Fable Simulation. Fable created an AI tool that can generate short animated videos using styles and characters that resemble existing shows and films. The studios allege that the company used copyrighted scripts, stills, and character designs to train its model, resulting in outputs that are visually and stylistically similar to their original works.

This case raises unique questions because the outputs are not static images or text—they’re videos that mimic the tone, pacing, and look of professional film and television. The studios argue that this kind of training threatens their intellectual property, especially if AI-generated content begins to compete with the originals in streaming or entertainment markets.

4. Authors v. Anthropic

In a recent decision, a federal judge ruled that Anthropic’s use of copyrighted books to train its Claude AI model could qualify as fair use—at least when those books were legally acquired. The case was brought by a group of authors who claimed their books were used without permission. The court sided with Anthropic in part, noting that training on purchased or digitized copies was “exceedingly transformative” and didn’t violate copyright law.

However, the ruling also revealed that Anthropic downloaded millions of books from pirate websites to build its training library. The judge allowed that part of the case to move forward, stating that using pirated material could still constitute infringement. That portion will go to trial later this year and could lead to major damages if willful copyright violations are proven. This case draws a line between legally obtained and pirated training data. This suggests that how content is sourced may matter just as much as how it’s used.

What’s at Stake for Creators, AI Companies, and the Public

The debate over AI training goes well beyond copyright law. It has serious implications for creators, tech companies, and everyday users of AI-generated content.

For creators, the key question is whether courts will recognize and uphold their rights in the training process. Many have already voiced concerns about their work being used without permission or compensation. The decisions ahead will determine whether creators can assert control or whether legal protections will remain limited. If courts require licensing, that shift could open new revenue streams and give creators a say in how companies use their work. But if courts interpret fair use too broadly, the system could leave creators out of the conversation entirely.

AI companies face a different set of risks. If licensing becomes the standard, development costs could increase significantly. This would be especially challenging for companies building large-scale models. It could also restrict access to the data needed to improve performance in areas like language generation, image creation, and other content-based tools. Depending on how the courts rule, companies may need to revise how they gather data, retrain models, or comply with new transparency requirements.

For the public, the outcome will shape how people experience and trust AI-generated content. Without clear rules, many wonder what counts as original, what’s derived, and what’s even allowed. This confusion already fuels deeper questions about authorship, misinformation, and digital trust. Clear legal guidance from courts or lawmakers can help rebuild confidence in how AI content is created, labeled, and shared. Ultimately, this isn’t just a copyright issue. It’s about who gets to shape our creative, cultural, and technological systems.

Practical Solutions to Support Both AI Innovation and Creator Rights

As legal battles unfold, a larger question is coming into focus: what should responsible AI training actually look like? There’s growing interest in finding solutions that don’t pit creators and companies against each other, but instead offer a path forward that’s fair, transparent, and sustainable.  A few leading proposals include:

1. Licensing Agreements for Training Data

One approach is licensing. Platforms could pay creators or rights holders to use their work in training datasets, similar to how streaming services pay musicians and publishers. Adobe, for instance, trains its Firefly model only on licensed or owned content and offers legal protection to users of its tools.  Future models might rely on creator marketplaces or collective licensing systems to manage permissions at scale. These approaches aim to ensure creators are compensated when their work supports commercial AI systems.

2. Transparency Requirements for Training Practices

Another key proposal is transparency. Today, most creators can’t tell whether AI models have used their work for training. New rules could compel companies to disclose the data they use, where they source it, and whether they obtained permission. Transparency can also highlight when certain voices or creative styles appear more frequently than others. This kind of visibility helps both creators and the public better understand how models are built and whose work they draw from.

3. Opt-Out Systems to Give Creators More Control

Opt-out mechanisms are also being explored as a short-term way to give creators more control. These tools allow artists to mark their work as off-limits for AI training. Early efforts include open-source initiatives like “Do Not Train” metadata tags and registries where creators can declare their preferences. While not a complete fix, opt-outs offer a starting point for consent-based systems while broader legal protections are still under debate.

4. Legislative and Policy Developments

There’s also movement on the policy front. The EU’s AI Act includes rules about data transparency and tracking where content comes from. In the UK, officials have released new guidance on AI ethics and creative rights. In the United States, Tennessee’s ELVIS Act was passed to protect people’s voices and likeness from being used without permission in AI-generated content. 

But federal action remains uncertain. In April 2025, former President Donald Trump publicly rejected proposals requiring AI companies to compensate creators for training on their work. His comments sparked backlash from artists and musicians pushing for stronger national protections, highlighting the political divide around how—or whether—creators should be compensated when their work powers commercial AI tools. These efforts show that lawmakers are beginning to take concerns seriously, but the path to comprehensive regulation remains politically contested.

Conclusion: The Future of Fair Use and the Boundaries of Consent

Whether AI can legally train on copyrighted content without permission is still an open question. The answer will help define not just how copyright law evolves, but what kind of role creators have in shaping the AI tools of tomorrow.

This conversation is about more than just legal arguments. It’s about fairness, consent, and recognizing the value of the people whose work makes these systems possible. That’s why efforts to give creators more control—like protecting their likeness, voice, or creative style—are starting to gain more traction.

The challenge ahead isn’t choosing between innovation and creator rights. It’s about finding a way to support both, through clearer rules, stronger protections, and shared responsibility for how AI develops from here.