Courts Are Turning AI Into a Product, Ruling by Ruling

In the last ten days of July, a single AI company received two legal outcomes that point in opposite directions. A federal judge in California ruled that Anthropic's practice of destroying physical books to train its models — scanning them, then pulping them — is fair use, a clean win for the proposition that training AI on legally acquired material is protected ^[1]. In the same window, the same company finalized the largest copyright settlement in U.S. history: $1.5 billion, plus a permanent obligation to delete every pirated dataset it had ever used ^[2]. The two outcomes do not contradict each other. Together they draw a line that is fast becoming the organizing principle of AI law: the act of training is protected, but the provenance of the training data is not.

We reached this settlement in 2025, after the court’s landmark ruling that training AI on books is fair use under copyright law — which remains the law today. — Anthropic

That line is one wall of a legal framework that courts across four jurisdictions have begun assembling in a matter of weeks. The second wall went up in Munich. In June, a German court ruled that Google is directly liable for defamatory claims produced by its AI Overviews. The company had argued that users must fact-check AI results — that the summaries were essentially speech, and Google was merely the platform. The court rejected that defense with a single doctrinal move that may prove to be the most consequential sentence in AI law this year.

a self-contained statement with independently comprehensible content. — Regional Court of Munich

The mechanism is plain. If an AI summary is a self-contained statement — not human speech, not a conduit for someone else's expression — then it is commercial activity. And commercial activity carries no free-speech shield. Google is answerable for what its AI produces the way a company is answerable for what it sells ^[3]. That is what converts a tool into a product in the eyes of the law. The UK's Competition and Markets Authority extended the framework in a different direction in early June. It ordered Google to let publishers opt out of having their content used to train or fine-tune AI models, and prohibited the company from retaliating by downranking those publishers in traditional search results ^[4]. The CMA chair framed it as a property question.

Today, we have introduced a world‑first requirement on Google's search services in the UK, enabling fair treatment, greater transparency and meaningful choice for businesses and consumers. — Competition and Markets Authority

The connection to the broader framework is that training data is being treated as a regulated input whose owners can set terms — not a commons anyone can draw from. The CMA's mandate does not use the word "product," but it reaches the same destination by a different route: if you want to use someone's content to build an AI system, you need their permission. India's Bombay High Court added another wall later in June, ruling that a Bollywood actress could sue Google and Meta for AI-generated deepfakes that used her likeness without consent ^[5]. The ruling extends the liability framework to identity itself — treating a person's image as an input that cannot be freely appropriated, the same way the other courts are treating copyrighted text and publisher content. Then, on the last day of July, a second Munich ruling landed. This one targeted Suno, the AI music generator, and it went further than any court had before. The court found that Suno's training process constituted "memorization rather than permissible analysis," rejecting both U.S.-style fair use defenses and the European Union's text-and-data-mining exceptions ^[6]. One Munich court has now found that training can constitute copying even where statutory carve-outs exist — a finding that, if it holds, removes a defense AI companies had been counting on. The framework is real, and it is accumulating. But it is also contested at every step. Google is appealing the Munich ruling that classified its AI Overviews as commercial speech ^[3]. In a separate case, the company is moving to dismiss a music copyright suit by arguing that YouTube's Terms of Service already grant it a "broad, royalty-free, and transferable license" for AI training ^[7].

Plaintiffs each granted YouTube, and Google—which provides the service—a broad license to use the uploaded content. — Google

The legal landscape is not monolithic. Anthropic won fair use for training on legally acquired books. Google is arguing its user agreements already cover AI training. Suno lost entirely. Different courts are ruling differently, and different companies are using different defenses. The framework is being built, but no one has accepted it as settled. And while the legal scaffolding goes up, the building itself keeps getting taller. Meta forecasts up to $145 billion in data-center spending for 2026 — nearly its entire projected EBITDA — and is so concerned about overbuilding that it has begun leasing excess compute to outside customers ^[8]. Nvidia projects that global data-center capital expenditure will reach $3 to $4 trillion annually by 2030, with the four largest AI hyperscalers spending over $650 billion in 2026 alone ^[9]. The legal framework and the infrastructure buildout are two construction projects on the same site, and neither crew has stopped for the other.