OpenAI fires back at authors over ChatGPT copyright claims

If someone reads your novel to understand how grammar works and then writes their own stories, did they steal your work?

That is essentially the legal battleground OpenAI just staked out this week in California federal court.

Earlier this summer, authors including Sarah Silverman and Paul Tremblay hit OpenAI with class-action lawsuits, alleging ChatGPT was trained on pirated copies of their books.

Now, OpenAI has filed its formal motion to dismiss, and it is going on the offensive.

Trimming the lawsuit down

OpenAI is not asking the judge to throw out the entire fight just yet.

Instead, it asked the court to toss almost every claim except for one: direct copyright infringement.

Claims like vicarious infringement, DMCA violations, unfair competition, negligence, and unjust enrichment need to be scrapped immediately, OpenAI argued.

The company wants to prune what it calls legally infirm theories before the lawsuit heads deeper into costly discovery, noting it plans to defeat the direct infringement claim at a later stage.

Learning the rules of language

OpenAI argued that the authors fundamentally misconceive how copyright law works.

Copyright exists to promote innovation and protect specific creative expressions, but not the underlying facts, ideas, or structural building blocks of language.

Even if these books made up a tiny fraction of the training data, OpenAI insisted using them in transformative ways falls squarely under fair use.

The company pointed out that it is not pirating books to resell them. The goal is simply to teach models how human language works so people can draft emails, save time, or entertain themselves.

To back this up, OpenAI cited the landmark Google Books case, reminding the court that statistical info like word frequencies, syntax patterns, and thematic markers cannot be copyrighted.

Why it matters

This dispute will shape the legal foundation of generative AI for years to come.

If the courts side with the authors and decide that pulling statistical patterns from text is infringement, building frontier models could become legally impossible or prohibitively expensive for startups and developers alike.

We think OpenAI played its hand smartly here by narrowing the scope early. Instead of rushing to fight every accusation at once, it stripped away the messier secondary claims to isolate the core argument around fair use.

Leaning on the Google Books precedent is clever, as large language models do operate on statistical probabilities rather than direct copy-pasting.

However, the risk remains. Unlike a search engine providing snippets, generative models can occasionally output text closely resembling original works, leaving a gray area that judges will eventually have to confront.