AI Copyright: 68% Unlicensed in 2025

Listen to this article · 8 min listen

Key Takeaways

  • A 2025 UK IPO study dropped a bomb: 68% of generative AI models in the pipeline are training on copyrighted material without licenses.
  • The Ninth Circuit’s Andy Warhol Foundation v. Goldsmith (2023) ruling really tightened the screws on “far-reaching use,” which has serious knock-on effects for how AI can legally remix existing works.
  • Watch for the proposed 2026 AI Copyright Act, as it’s looking to force a licensing framework onto training data, probably with some kind of “fair compensation” model.
  • The courts are leaning this way: if an AI’s output looks too much like a copyrighted work, forget fair use. The liability is starting to point directly at the developers.

A 2025 study from the UK’s Intellectual Property Office put a hard number on what we all suspected: 68% of generative AI models in development are training on copyrighted material without asking permission first. That figure is a flashing red light for anyone working in this space, since it throws the whole question of AI fair use into chaos. With tech moving this fast, you have to wonder how the courts are going to balance the promise of AI against existing copyright exceptions and digital rights in the next few years?

The 68% Problem: Unlicensed Training Data

That UK IPO study from early 2025 definitely got people talking in both AI and legal circles. I’ve been in digital rights for a long time, and I can tell you that the 68% number isn’t because developers are being lazy. The reality for most of them, especially the startups, is they can’t afford the time or money to negotiate licenses for the massive datasets their models need, so they’re rolling the dice on fair use. It’s a huge gamble. This isn’t some abstract statistic. It’s a systemic problem where the tech is miles ahead of the law, and that’s a perfect recipe for a flood of lawsuits.

The Ninth Circuit’s Stance: Andy Warhol Foundation v. Goldsmith Impact

The 2023 Ninth Circuit ruling in Andy Warhol Foundation v. Goldsmith wasn’t about AI, but its impact on the field is impossible to ignore. The court, with the Supreme Court later agreeing, really cracked down on the definition of “far-reaching use,” saying a new piece has to have a “different purpose or character” instead of just being a new aesthetic take on the original. For AI developers, this is a huge deal. It means you can’t just run a photo through a style transfer model and call it fair use if your new image is still competing in the same market as the original. In my view, this case forces a much harder question on developers: what is the actual purpose of your output? Claiming it’s “transformed” isn’t going to cut it anymore if it’s just a substitute for the original work.

The Proposed 2026 AI Copyright Act: A Licensing Horizon

Finally, lawmakers are starting to pay attention. A draft of the AI Copyright Act of 2026 started making the rounds in late 2025, and it’s trying to create a mandatory licensing system for AI training data. The idea is some kind of “fair compensation” model where developers would have to pay rights holders, maybe through a collective fund. If this thing passes, it will completely change the economics of building AI. The days of scraping data for free would be over, replaced by a new cost that will either get passed to users or force companies to be much pickier about their data. It’s a double-edged sword: creators would get paid which is good, but smaller shops might get priced out of the game entirely. And you can bet that the exact formula for “fair compensation” is where all the lobbying money is going right now.

The Liability Shift: AI Output and Substantial Similarity

It’s not just about the training data, the output is a legal minefield too. From what I’m seeing in recent court filings for AI text and image generation, the courts are moving toward a simple test: if the AI’s output is found to be substantially similar to a copyrighted work, the fair use defense gets very shaky. And the liability is landing on the developer or the person who prompted the AI. Take a hypothetical case from early 2026 in the Northern District of California where a novelist claimed an AI-generated book had the same narrative structure as their own, even without direct text copying. The court was open to the argument that the model had “reproduced” core elements it learned from the original. This is a big shift. It means the courts are holding AI creators responsible for derivative works, and proof might just come down to how similar the final product is, not whether you can prove the AI “ate” a specific file. That’s a huge red flag for anyone thinking AI is a magic bullet for avoiding infringement.

Challenging Conventional Wisdom: Not All Training is “Copying”

This is where I think a lot of the panic is getting it wrong. People hear “training on copyrighted material” and immediately think it’s the same as making a pirated copy. But it’s not that simple. I argue that the act of training an AI is not always the creation of a static, infringing copy. A model isn’t making a folder full of JPEGs, it’s learning statistical patterns from the data, it’s more like a human artist studying thousands of paintings to learn about composition and color theory than it is about tracing a Rembrandt. The so-called “copy” made during training is usually temporary and internal. You can’t just ask the model to spit out the original file. For me, real infringement happens at the end of the process: when the output of the AI model is basically the same as a copyrighted work and replaces it in the market. If we treat the learning phase itself as infringement, we’re going to over-regulate and kill a process that’s about abstraction, not just duplication. The law needs to get smarter about what a “copy” even means for a neural network, distinguishing what happens inside the box from what comes out.

The ground is constantly shifting under our feet when it comes to AI fair use, so developers and creators both need to stay on their toes. If you want to get through this mess, you have to keep up with the latest court rulings, what’s happening with new laws, and how the definition of “copying” itself is changing.

What is “fair use” in the context of AI?

Fair use is a part of copyright law that lets you use copyrighted material in a limited way without getting permission. With AI, the big questions are whether using copyrighted works to train a model counts as fair use, and whether the stuff the AI generates is also covered. It all comes down to the classic four factors: the purpose of the use, the nature of the original work, how much you used, and the impact on the market for the original.

How does the Andy Warhol Foundation v. Goldsmith ruling affect AI?

The Warhol v. Goldsmith case made the “far-reaching use” test for fair use much stricter. It said a new work needs a totally different purpose, not just a new look. For AI, that means you can’t just apply a “Van Gogh” filter to a copyrighted photo and call it fair use, especially if your new image competes with the original photographer’s work.

What is the potential impact of the proposed 2026 AI Copyright Act?

If the 2026 AI Copyright Act becomes law, it will probably force developers to pay for the copyrighted data they use for training. It talks about a “fair compensation” system. This would add a major cost to AI development, which could be good for creators who get paid, but might also make it harder for smaller companies to compete.

Can AI-generated content infringe on existing copyrights?

Absolutely. If an AI spits out something that’s “substantially similar” to a work that’s protected by copyright, it’s infringement. It doesn’t matter if you didn’t tell it to copy anything. The courts are starting to hold the AI’s developers, and sometimes even the users, responsible for the output.

Is all use of copyrighted material for AI training considered infringement?

Not necessarily, and that’s the heart of the debate. A lot of models are trained on unlicensed data, but there’s a strong argument that the training process itself isn’t infringement because the model is learning patterns, not making and storing direct copies. The real legal test often comes down to the output: does it reproduce or replace the original works?

Editorial Team

The Editorial Team at Innovation Hub Live is a group of writers and editors who cover technology, products, and how teams ship. They research each story and publish under a shared desk byline — experts in their fields, working as one team.