fair use 2026: AI Copyright Law Update 2026: What Changed

For a few years, “AI and copyright” was mostly a debate – hot takes, panels, and a pile of lawsuits that hadn’t gone anywhere yet. In 2026 that changed. Real money moved, judges put real lines on paper, and the government kept refining what it will and won’t protect. If you create for a living – art, writing, music, video, photography – the rules you’re working under are firmer now than they were a year ago.

Here’s a plain-language tour of what actually shifted this year, and what each change means for you. (“IP” here means intellectual property, and this is general information, not legal advice.)

AI Copyright Law Update 2026 — context image

A gold timeline ribbon with milestone dots representing a year of AI copyright developments — 2026 turned years of AI copyright theory into concrete rulings, settlements, and bills.

The headline: a $1.5 billion payout finally got approved

The biggest news of the year is that the Bartz v. Anthropic settlement received final court approval. Anthropic – the company behind the Claude AI models – agreed to pay roughly $1.5 billion to authors whose books were used without permission. It’s the largest copyright settlement of its kind, and it turned an abstract fear (“they trained on my work”) into an actual check.

Track the cases yourself

Every current AI-copyright ruling — and the ones still being fought — lives in our AI Copyright Ruling Tracker. Filter by your creator type to see, in plain language, how each case affects your rights.

The rough math: about $3,000 per work after fees, across a class covering roughly 482,000 books. That number matters less than the precedent. For the first time, a major AI company paid creators directly for how it built its dataset.

But read the fine print, because it’s the part creators keep getting wrong. The settlement releases claims only for Anthropic’s past acquisition and copying of works – and only through August 25, 2025. It does not release claims about what the AI outputs, and it does not cover future conduct. Translation: this closed one chapter (how the books were obtained), not the whole book. The question of AI-generated outputs that resemble your work is still wide open.

The pattern courts settled into: training vs. how you got the data

To understand why Anthropic paid, you have to understand the ruling that set it up. In mid-2025, Judge William Alsup found that using books to train a large language model was “spectacularly” transformative – and therefore fair use – when the books were legally obtained. That was a big win for AI developers.

Then came the catch. The same judge flatly rejected the idea that downloading and storing millions of pirated books from shadow-library sites was fair use. Training could be fine; stealing the training material was not. That distinction – lawful use of the model, unlawful acquisition of the data – is exactly the exposure the $1.5 billion settlement paid to make go away.

A forking path showing legally acquired data leading to a checkmark and pirated data leading to a warning — Courts drew a sharp line: training can be fair use, but how the data was obtained is where the exposure lives.

Set that next to Thomson Reuters v. Ross Intelligence, decided in early 2025, where a court rejected a fair use defense. Ross had copied Westlaw’s legal headnotes to build a competing legal-research tool. The court leaned on two things: the use competed directly with the original’s market, and – crucially – Ross’s tool was not generative AI; it didn’t create new content, it retrieved. So it’s not a clean “AI loses” precedent, but it’s a loud reminder that copying someone’s work to build a product that competes with them is the fact pattern courts punish.

What this means for you: the emerging rule of thumb is that transformation can be defensible, but substituting for the original – or building on stolen material – is where liability lives. If you’re using AI tools, the provenance of what you feed them matters.

The human-authorship line held firm

On the registration side, 2026 mostly confirmed a rule the U.S. Copyright Office drew in its ongoing Copyright and Artificial Intelligence reports: human authorship is required, and it’s evaluated case by case.

The Office’s position, in everyday terms:

A work generated entirely by AI, with no meaningful human creative input, can’t be registered.
Prompts alone don’t cut it – even long, detailed, carefully engineered prompts. Describing what you want isn’t the same as authoring it, because you don’t control the specific expression the machine produces.
But when a human meaningfully shapes, selects, arranges, or edits the output, that human contribution can be protected. Registrations continued to be granted this year for works with genuine human authorship layered in.

What this means for you: if you use AI as one tool inside your own creative process, you may still own the parts you actually authored – your arrangement, your edits, your original elements. What you can’t do is press “generate” and claim a copyright on the raw result. When you register, disclose the AI-assisted portions and claim the human parts. Being straight about it protects your registration; hiding it can invalidate it.

Your voice and face got new legal muscle

Copyright isn’t the only front. 2026 saw real movement on likeness and voice – the deepfake problem. The NO FAKES Act, a bill that would create a federal right against unauthorized AI “digital replicas” of a person’s voice and visual likeness, advanced out of the Senate Judiciary Committee this year and moved toward the full Senate.

It isn’t law yet, and it may still stall. But the direction is unmistakable, and it’s backed by an unusual coalition – artists, labor unions, AI companies, and child-safety groups. Layer on the growing patchwork of state laws protecting name, image, voice, and likeness, and the practical upshot is this: cloning a real person’s voice or face without permission is getting legally riskier by the month.

What this means for you: if you perform, narrate, or appear on camera, your identity is becoming a protectable asset in its own right. And if you make content, using AI to mimic a recognizable real person is a fast-growing liability – get consent.

The international picture: not everyone agrees

Copyright is territorial, and 2026 made that obvious. In the UK, the High Court handed down Getty Images v. Stability AI, and Stability largely prevailed. Getty dropped its primary training-related claims during trial – partly because the training didn’t happen in the UK – and the court rejected the remaining “secondary infringement” theory, finding the AI model itself wasn’t an infringing copy under UK law. Only a narrow trademark issue stuck.

Meanwhile the EU AI Act pushed transparency the other way: providers of general-purpose AI models face obligations to publish summaries of the content used to train them. Different countries, different answers – which means where a case is fought can matter as much as the facts.

What this means for you: don’t assume a headline from one country describes your rights. If your audience or your infringer is abroad, the outcome can flip.

What this means for you: a 2026 checklist

Pulling it together, here’s how a working creator should respond to this year’s changes:

Keep proof you’re human. Save sketches, drafts, layer files, project files, and version history. In a world where authorship is the whole ballgame, being able to show your process is your strongest asset.
Be honest on registration. If AI touched your work, claim the human parts and disclose the rest. A clean, accurate registration beats an inflated one that can be voided.
Mind your inputs. If you train, fine-tune, or feed tools with other people’s work, provenance matters. “It was on the internet” is not permission.
Guard your likeness. If your voice or face is part of your brand, watch for unauthorized AI replicas – the law is moving to your side, and documenting misuse now sets you up to act.
Watch the output question. The settled fights this year were about training data. The next wave is about AI outputs that resemble real creators’ work – the part no settlement has closed.
Get advice when the stakes are real. For a commercial project, a licensing deal, or an infringement worth chasing, a qualified IP attorney is cheaper than guessing.

The through-line of 2026 is that AI copyright stopped being hypothetical. Creators who track how the rules are hardening – and who keep clean records of their own human work – are the ones who’ll be able to defend it. The ones still assuming “nobody can tell” are the ones about to find out that, increasingly, the law can.

This article is general information, not legal advice. AI copyright law is changing fast and varies by country and by the specific facts, so for anything high-stakes, get a qualified intellectual property attorney’s read before you act.

Sources & further reading:

AI Copyright Law Update 2026: What Changed This Year