case-law

Doe v. GitHub: The Copilot Case Testing Open-Source Attribution

By Nora Whitfield · August 9, 2026

Doe v. GitHub started in November 2022 as one of the very first AI copyright lawsuits filed anywhere, months before ChatGPT even had its public moment. It's still going, and in February 2026 it landed in front of the Ninth Circuit on a question that could reshape how every AI coding tool handles the open-source code it was trained on: does an AI tool have to reproduce your code exactly for you to have a claim, or does stripping your name off a close paraphrase count too?

What Copilot allegedly does

GitHub Copilot was trained on public repositories across GitHub, a huge share of which carry open-source licenses like MIT, Apache, or GPL. Those licenses aren't "do whatever you want." Most of them require attribution: if you reuse someone's code, you have to credit them, and some require that derivative works carry the same license forward. That's the deal open-source licensing runs on, and it's the deal a group of anonymous developers, suing as "Doe" plaintiffs, say Copilot broke.

Their complaint, filed in the Northern District of California as case 4:22-cv-06823, alleges Copilot regenerates code learned from these repositories without the attribution or license terms attached, presenting it to users as if it were license-free, original output. If you're a developer whose MIT-licensed function ends up suggested to another developer through Copilot's autocomplete, with your name and license terms nowhere in sight, the plaintiffs argue that's a straightforward violation of the deal you struck when you open-sourced your work in the first place.

What survived, and what didn't

In May 2023, the court dismissed the core copyright infringement claims. The reason wasn't that copying open-source code through an AI is fine. It was narrower and more procedural: the plaintiffs hadn't pointed to specific instances where Copilot generated code substantially similar to their specific, identifiable code. General allegations that Copilot was trained on your repository aren't enough. You need to show the output.

What survived is arguably more interesting than what didn't. The DMCA Section 1202(b) claim, which covers stripping copyright management information (author names, license notices, that kind of metadata) from a work, made it through. So did breach-of-contract and open-source-license claims, built on the theory that using GitHub's platform means agreeing to honor the license terms attached to the code you pull from it, and Copilot's training and output process broke that agreement at scale. Those claims are the ones now working their way through discovery, and they're the ones the Ninth Circuit heard argument on in February 2026.

The question in front of the Ninth Circuit

The core legal question is deceptively narrow but has enormous practical reach: does DMCA 1202(b) liability require the AI's output to be an exact, identical copy of the original code, or does it also cover output that's a close paraphrase, functionally and structurally similar but not byte-for-byte identical, generated without the attribution the original license required?

If the Ninth Circuit reads 1202(b) narrowly, requiring near-identical copying, most AI-generated code escapes liability under this theory entirely, because generative models rarely output exact matches to any single training example. If the court reads it more broadly, covering output that's recognizably derived from licensed code even when reworded or restructured, that's a much bigger problem for every company building code-generation tools, not just GitHub. It would mean the attribution and license obligations that already bind a human developer copying and pasting code apply with real force to what an AI model regenerates too.

Why this is the test case for developers specifically

Most AI copyright litigation you've read about concerns books, images, music, or news articles. Doe v. GitHub is the one built specifically around code, and specifically around the norms of open-source licensing, which already has decades of established law about what attribution and derivative-work obligations mean. That existing body of law is exactly why this case is being watched so closely: it's not asking courts to invent new doctrine, just to apply doctrine that already exists to a genuinely new kind of intermediary, a model trained on licensed code that then generates new code on demand.

What this means for AI authors

If you write code and use Copilot, Cursor, or any similar AI coding assistant, day to day, this case doesn't create new obligations for you. But it's actively defining whether "the AI generated it" will hold up as a defense against license and attribution claims the way a human developer copy-pasting the same code would face. Right now, that question is genuinely unresolved, and the Ninth Circuit's ruling on the 1202(b) scope will be one of the clearest signals yet on which way it's headed.

If you maintain open-source projects, this case is the reason your license terms matter more than ever, not less, in an AI-assisted world. The theory Doe v. GitHub is built on assumes your license terms are still binding even when the thing copying your code is a model rather than a person. If that theory wins, the attribution obligations you attached to your repository years ago could turn out to be enforceable against AI-generated derivatives you never anticipated.

And if you're shipping AI-assisted code in a commercial product, the practical takeaway right now, while this is still unsettled, is that documenting what you wrote yourself versus what a tool suggested is worth doing before a dispute forces you to reconstruct it after the fact. A case like this one takes years to resolve. Having your own record of authorship doesn't depend on which way the Ninth Circuit eventually rules.