GPL-2.0-only Apache-2.0 Compatibility for AI-Generated Code

Why GPL-2.0-only and Apache-2.0 conflict

Two features of GPL-2.0-only drive most of the concern. The first is the word "only". Some GPL-2.0 projects grant the option to use a later version of the licence. GPL-2.0-only does not, so the code cannot be relicensed under GPL-3.0 terms, and that closes the route most often used to combine GPL code with Apache-2.0 code.

The second is the interaction between licence terms. Apache-2.0 includes an express patent grant that can terminate for a licensee who brings certain patent litigation, and it carries attribution and notice conditions. GPL-2.0 forbids adding restrictions beyond its own terms to a combined work. The patent termination condition is commonly treated as such a restriction, which is why GPL-2.0-only code is generally considered incompatible with distributing a combined work under Apache-2.0.

The relationship is also directional. Apache-2.0 code is generally treated as usable in GPL-3.0 projects, but that does not make GPL-3.0 code usable in an Apache-2.0 project, and it does not extend to GPL-2.0-only. The practical consequence for an Apache-2.0 project is that GPL-2.0-only material in the tree is a distribution problem, not only a documentation problem. These are general descriptions, not legal advice. Specific combinations and remediation choices need review by qualified counsel.

How AI suggestions introduce the problem

NIST AI 600-1 notes that if a generative AI system's training data included copyrighted material, outputs displaying instances of training data memorization could infringe copyright. For code, that means an assistant may emit a snippet that matches or closely resembles licensed source. The assistant does not attach the licence, and the engineer accepting the suggestion often cannot tell where it came from.

This guide makes no claim about how often this happens, and no frequency figure should be assumed. The control question does not depend on the rate. A single unreviewed GPL-2.0-only snippet in an Apache-2.0 release creates the same distribution issue as a larger import.

The Copyright Office has listed liability for infringing works generated using AI systems as an open area of inquiry, including whether outputs can implicate the right of reproduction or the derivative work right. These are posed as questions, not findings. Until they are settled, the sensible position is to treat ambiguous assistant output as a licence question that a person should resolve.

Where dependency-level scanning stops working

Software composition analysis is built around declared components: package manifests, lockfiles and known library versions. A snippet pasted from an assistant has no manifest entry and no declared licence. It looks like first-party code, so a dependency scan can pass while the repository contains code under an incompatible licence.

NIST SP 800-218A does not distinguish between human-written and AI-generated source code, because all source code is assumed to need evaluation before use. That supports applying checks to assistant output at each stage where code moves, not only after merge.

At suggestion time, a check can assess origin and licence context before the engineer sees the snippet, and it is the only stage where a block prevents the code from entering the editor at all. At pre-commit and in CI, checks can flag changes that came from an assistant and route them to review. At release, checks can confirm that required licence statements and notices are present in the artefacts. Each stage catches what earlier ones miss, and matching on similarity has inherent limits: modified, reformatted or partial copies can be harder to identify than verbatim ones. Layered checks with human escalation are more dependable than any single detector.

Where Trussed fits

Trussed enforces policy in the execution path of coding assistants. Code provenance supplies the origin, licence and security context of code, and policy is enforced at runtime, not reviewed after the fact. For an individual interaction or tool call, the outcome can be allow, redact, block, rate limit or escalate for human approval. This maps to the suggestion-time step above and to the escalation path for ambiguous cases.

Each enforcement action also produces audit-ready evidence, so the record of what was checked and what was decided exists without a separate reconstruction exercise. Trussed is deployed self-managed in the customer cloud, with configurable retention and data residency, and has run in production in a regulated enterprise for more than 18 months.

This does not replace pre-commit, CI and release checks, nor does it replace counsel. It addresses the stage where dependency scanning has no visibility, and it gives legal and audit teams a decision record to review.