AI's legal identity crisis

If you replace every plank on the Ship of Theseus, is it still the same ship? Big Tech may be about to find out.

Meta got a $25 billion reminder last week when a New Mexico judge ordered the tech giant to pay claims tied to its Cambridge Analytica privacy scandal. The company plans to appeal, and experts expect the award could shrink. Meta also agreed to a $17 billion settlement with states this summer over alleged harms to teens involving products like Instagram.

Those cases aren’t necessarily about AI. But they do show how decisions tech companies make today can turn into enormous liabilities years later—and that has major implications for today’s AI developers and hyperscalers.

Platform, product, or something else?

For decades, companies like Facebook and Twitter benefited from Section 230, a law that generally protects online platforms from being treated as the publisher of content posted by their users.

Generative AI complicates that distinction. ChatGPT, Claude, and Grok don’t merely host somebody else’s words. Large language models are trained on enormous amounts of human-created material, respond to prompts written by users, and produce outputs shaped by how developers built and trained them. The result isn’t simply somebody else’s post being displayed on a platform. Now regulators are testing who shoulders that responsibility.

Eight states sued OpenAI this month over ChatGPT’s alleged harms to users, including teenagers, and are seeking not just damages but changes to how the LLM is designed. Other plaintiffs are testing whether AI systems can be treated under traditional product-liability theories when their design or outputs allegedly cause harm.

Anthropic is litigating a different boundary. In a new suit by the owners of Elvis Presley’s catalogue, music publishers argue that Claude was trained on copyrighted lyrics and can generate competing content. While Anthropic says using those works for training is transformative fair use, federal judges have already reached different conclusions in previous cases about when AI training qualifies for that protection.

Google, meanwhile, recently beat lawsuits from Chegg and Penske Media over its AI Overviews. While publishers argued Google was using their content to generate answers that siphoned away their rightful traffic, the case judge dismissed the antitrust claims—and acknowledged the broad problem of creative work becoming repurposed without compensation.

The legal tab gets operational

For investors, that question is becoming more than a law-school debate. As companies like Anthropic prepare to sell shares to the public, legal risk is joining compute costs, losses, and valuation on the list of things investors have to price in.

Broader liability around training data, product design, or AI-generated output could mean higher licensing and insurance costs. It could also force companies to redesign products or slow releases, hitting growth as well as margins.

Investors aren’t just betting on how good the models get: They’re also betting on how expensive the rules around them become.

Silicon Valley is replacing the old internet’s planks one by one. Now courts have to decide whether it’s still the same ship.—HC