This is where things get philosophically weird. If a human historian writes about Westphalia after reading twenty sources, nobody accuses them of plagiarism. But if an AI reads twenty sources and produces something that blends all of them, is that transformative or derivative?

Matthew Sag, a law professor at Emory University who studies AI and copyright, told me last year that courts will eventually need to decide whether training data and output are separate questions. Britannica's lawsuit forces that question. They're not arguing that training on their content is illegal—they're arguing that reproducing it in output is.

The practical implications are huge. If Britannica wins, every AI company will need to implement much more aggressive filters to prevent their models from producing anything that looks like training data. That's technically possible—OpenAI already has systems to detect and block certain kinds of memorization—but it makes models less useful. The whole point of ChatGPT is that it can answer questions with specific, detailed information. If it has to be vague enough to avoid copyright claims, it becomes less valuable.

There's also a question of selective enforcement. Britannica is suing because they're a legacy publisher with lawyers and a business model to protect. But ChatGPT was also trained on Reddit comments, blog posts, and Wikipedia articles. Those creators aren't suing—either because they don't know, don't care, or can't afford to.

The case will take years to resolve. In the meantime, OpenAI will keep operating, and Britannica will keep publishing. But this lawsuit is asking the right question: if an AI can do your job after reading your work, do you have any claim to the output?

The technology is impressive. The question is whether Britannica needs compensation for making it possible.

1 day ago