Fair Use or Free Ride? AI Copyright Law Just Shifted
Generative AI is advancing fast but the laws that govern it are only just catching up. Recent court rulings in the U.S. and UK are redefining what counts as “fair use” in AI training, while regulators move to enforce new standards for transparency and licensing. We break down what’s changing, what it means for your organisation, and what to watch next.
Copyright Law Is Finally Catching Up With Generative AI
Generative AI is no longer moving in the shadows, it’s become a $250 billion market almost overnight. But as models scale and outputs flood the internet, a fundamental question has surfaced: who owns what? In just the past few weeks, judges and policymakers have taken major steps toward answering that and reshaping the legal ground AI companies, creators, and businesses now stand on.
Where the Law Just Moved
In a landmark decision, a U.S. judge ruled that using copyrighted books to train AI models can qualify as “fair use”. In Bartz v. Anthropic, the court found that training Claude on a broad range of texts was “exceedingly transformative,” tipping the balance in favour of innovation.
Just days later, Tremblay v. Meta saw a similar outcome. A judge dismissed authors’ copyright claims, noting they had failed to show economic harm which was a key requirement in fair use cases.
Across the Atlantic, Getty dropped its “output copying” claim in its case against Stability AI in the UK, keeping only narrower arguments around secondary infringement and passing off. This highlights the difficulty in proving that an AI-generated output is directly copied from a specific original work.
Taken together, these early rulings offer a signal training data may fall under fair use in some jurisdictions. But how that data is collected, stored, and monetised still carries real legal risk.
The Regulatory Front Line
Governments are beginning to act but they’re moving cautiously.
In the U.S., the Copyright Office’s May 2025 report outlined new licensing models and reaffirmed that some AI training may fall under fair use. Two bills are also making noise:
The NO FAKES Act (revived in April) aims to curb deepfakes using likeness or voice.
The Generative AI Copyright Disclosure Act would force companies to publicly list their training datasets 30 days before releasing a model.
Europe has taken a stricter route. The EU AI Act, passed in March, will require providers to publish detailed summaries of copyrighted data used for training. It’s a shift toward transparency by design.
In the UK, plans for a blanket copyright exception that would have allowed unrestricted AI training on copyrighted content were abandoned after major backlash from creators. By 2025, the government shifted towards a licensing-based approach, backing away from sweeping exceptions and instead exploring collective rights management to balance innovation with copyright protections.
Multilateral bodies are weighing in too. The OECD’s February 2025 paper outlines scraped data risks and proposes voluntary codes, while WIPO and GPAI are working toward future treaty-level rules.
What Scholars Are Arguing
Legal academics are split but trends are emerging.
Stanford’s Mark Lemley argues courts may treat AI training much like Google Books: transformative and ultimately pro-competitive. Harvard’s Sobel goes further, questioning whether an artist’s “style” can be copyrighted at all.
Elsewhere, the Columbia Law symposium makes the case for collective licensing at scale. Think blanket, opt-out systems for text and image rights, similar to how music royalties work today. It's a vision that may appeal to both big tech and publishers.
Is Regulation Keeping Up?
Not quite. The largest open-weight models doubled in size between April 2024 and 2025. Fine-tuning cycles that once took months now finish in weeks.
Meanwhile, major U.S. legislation remains stuck in committee. In the EU, full enforcement of the AI Act, including rules for high-risk systems, won’t begin until August 2026 or later depending on the application. With lawmakers still behind the curve, courts are left to fill the gaps, shaping AI policy one case at a time.
Why All This Matters
These rulings and proposals aren’t just legal footnotes, they define how innovation happens.
Litigation costs already exceed $300 million. Whether AI developers must license content or rely on fair use determines who can compete, small labs or just tech giants. Meanwhile, businesses are being nudged toward dataset transparency, with compliance tools like watermarking, opt-out dashboards, and dataset “nutrition labels” gaining traction.
This evolving two-track system, rapid innovation on one side, slower regulation and licensing on the other, is set to shape how generative AI expands across industries.
What Next?
Appeals in the Meta and Anthropic rulings could reach the U.S. Supreme Court.
The EU will issue new standards for dataset disclosures under the AI Act.
Collective licensing platforms and token-priced APIs are being trialled by major publishers and Microsoft.
Technical solutions such as model-side hashing and watermarking are emerging as compliance defaults.
And authors are fighting back: an open letter signed by over 70 creators calls for publishers to rethink how AI fits into their contracts and content pipelines.