Doe v. GitHub: the Ninth Circuit Redraws the Risk Map for AI Coding Assistants

The filing of 16 September 2026

On 16 September 2026 the Court of Appeals for the Ninth Circuit filed its opinion in Doe v. GitHub, Inc. (No. 24-7700)^[1], affirming the partial dismissal ordered by the district judge.

The case reached the appellate court on interlocutory appeal under 28 U.S.C. § 1292(b). The trial-court docket bears the number 4:22-cv-06823-JST, before Judge Jon S. Tigar, Northern District of California.

The panel brings together Circuit Judges Sidney R. Thomas and Eric D. Miller, with District Judge Stanley Blumenfeld, Jr., sitting by designation. The opinion was authored by Judge Miller. The case had been argued on 11 February 2026 in San Francisco.

At issue are GitHub Copilot and Codex, tools that write code using language models trained on millions of software projects hosted on the platform.

Two questions that companies treat as one

The decision separates two questions that boards treat as a single one: who may bring suit, and which rule decides in favour of the party who does.

On the first question, the programmers win. The panel recognises Article III standing, because the plaintiffs plausibly alleged a substantial risk of harm.

On the second, the outcome flips. The court holds that those allegations do not make out a claim under § 1202(b) of the DMCA, the provision that protects copyright management information.

The distinction weighs on the risk ledger. An open courthouse door paired with a closed theory on the merits produces litigation that drags on, costs money and migrates towards other legal grounds. The governance signal is this: the threshold for getting before a judge stays low, the threshold for liability rises.

The "output" theory and the reason for dismissal

The plaintiffs argued that Copilot and Codex reproduced their code while stripping out the attribution.

Section 1202(b) targets the removal or alteration of copyright management information from a copy of an existing protected work. The panel applies the letter of the statute and concludes that the tools create new works that never carried that information to begin with.

The step is technical and decisive. A copy stripped of attribution falls within the provision; an output that is born without attribution stays outside its perimeter.

Anyone reading the ruling as a general acquittal of generative models is stretching the text. The panel decides a single DMCA provision, at a single procedural stage, within the perimeter of the allegations the plaintiffs brought. The other theories remain where they were.

The "input" theory is still on the table

The most consequential point for the next twelve months sits in a single line of procedure.

The panel declines to examine the "input" theory of DMCA liability and deems it forfeited. The issue stays procedural: it concerns how the argument was presented in the courts below, and it leaves the merits untouched.

For the General Counsel, the reading is straightforward. A theory deemed forfeited in this docket remains available to other plaintiffs, in other proceedings, with allegations built better from the trial court onwards. Exposure tied to the training phase therefore stays open, and a defence calibrated on the "output" outcome is off balance.

The question about output generation has been answered. A second question, about the ingestion of training data, has opened.

The exposure map for organisations already running assistants in production

For an organisation using coding assistants in production, the ruling redraws the exposure map rather than erasing it.

The dismissal of the § 1202(b) claim closes one route. Others remain open: classic copyright infringement, compliance with open source licences and their attribution requirements, the contract with the tool vendor, trade secret protection.

Many free licences tie use to preserving the attribution notice. An assistant that returns code without that notice shifts the problem from the DMCA to licensing and contract law. The perimeter changes, the duty to verify remains.

This is where a name is needed. Which role, in writing and before deployment, answers for the provenance and attribution of code generated inside the company? A framework that leaves the question open produces documentation instead of governance.

The European picture runs on a different rulebook

The decision binds the federal courts of the Ninth Circuit. Anyone operating in Europe answers to a different architecture, with different obligations and different arbiters.

Regulation (EU) 2024/1689, the AI Act, has been in force since 1 August 2024. Obligations for general-purpose models apply from 2 August 2025; the Article 50 transparency obligations from 2 August 2026.

The European framework therefore calls for model traceability and user notification, rather than a rule on the attribution of an individual code snippet.

On personal data, the authorities are already acting with consistency. The European Data Protection Board reported the €300,000 fine imposed by the CNIL on the company Extia for failure to respect data subjects' rights, as set out in the notice published on the EDPB website^[2]. The subject matter is different, and the shape of European enforcement comes through clearly: procedure, rights, documentary evidence.

Three decisions for the board

The decision turns three abstract questions into choices with a deadline.

1. General Counsel: map the contracts with coding assistant vendors and establish who takes on the intellectual property indemnity obligation.
2. Chief Risk Officer: update the risk register by separating the DMCA entry from the open source licensing entry, with two distinct assessments and two distinct owners.
3. Board Audit and Risk Committee: ask management for the name of the role that answers for the provenance of generated code, and the date of the next review.

The review must be carried out in writing, with dated and archived outcomes. The audit is still required; its scope has changed.

The CEO is left with a strategic decision that is already constrained: which projects depend on generated code, and what level of control the company funds before release. Compliance built now works as a competitive advantage rather than a cost.

Regulatory horizon

Current status: the opinion was filed on 16 September 2026 and is effective within the Ninth Circuit. The trial-court proceeding continues on the claims left alive before the Northern District of California.

Jurisdiction: US federal for the DMCA, with direct effect in the states of the Ninth Circuit, California included. For European deployers, the benchmark remains Regulation (EU) 2024/1689, applicable in phases.

Deadlines: the Article 50 transparency obligations apply from 2 August 2026, and the AI Office's implementing work continues. Organisations that name a role today, track provenance and archive outcomes will arrive prepared for the active enforcement phase.

One route is closed in the Ninth Circuit. The map, across the rest of the world, is still there to be walked.