AI Coding Assistant GPL-3.0 Infringement Liability: Allocating Defense Costs When Copilot and Amazon Q Developer Both Contribute

The scenario: one file, two assistants, one allegation

Consider a realistic sequence. A developer accepts a Copilot suggestion that forms the core of a function. Weeks later, another developer uses Amazon Q Developer to extend it. A third colleague edits several lines by hand. The file ships, and a claim later alleges that it reproduces GPL-3.0 code.

Under the US copyright framework summarised by the Congressional Research Service, a plaintiff must show that the defendant actually copied the underlying work and that the new work is substantially similar to it. That framing shapes the evidence question. The enterprise first needs to establish what was generated, by which tool, and whether any matched public code was flagged at the time.

The analysis is harder for users than it looks. A user of an AI tool may have no direct access to, or awareness of, the work allegedly infringed. An AI company, meanwhile, could potentially face theories such as vicarious infringement, which applies to parties with the right and ability to supervise the activity and a direct financial interest in it. The Stable Diffusion litigation includes such an allegation, but it is an allegation, not a ruling.

How defense-cost responsibility is likely to divide

Each vendor sets its own indemnity conditions. These can include the customer's plan tier, whether duplicate-detection or reference-filtering features were enabled, and whether the customer modified the output. The conditions and exclusions should be read directly from the current contract terms and confirmed by counsel. The table shows the structural pattern, not a prediction of any court outcome.

• Enterprise: Distributes the file, owns the GPL-3.0 compliance analysis, and bears defense costs for any portion outside a vendor's indemnity, including portions it cannot attribute.
• GitHub/Microsoft: Any indemnity is tied to Copilot's contribution and to the customer meeting the conditions in Microsoft's terms, such as plan tier, filter settings and unmodified output.
• AWS: Any indemnity is tied to Amazon Q Developer's contribution and to the conditions in AWS's terms, which may differ from Microsoft's.
• Gap: mixed and modified lines: Where both tools and a human touched the same lines, neither vendor is likely to accept that its contribution caused the infringement.
• Gap: unprovable settings: If the enterprise cannot show which filters and plan were active at generation, it cannot show a condition was met.
• Gap: licence obligations: A defense-cost indemnity does not remove GPL-3.0 obligations that attach to the code itself.

What stays with the enterprise regardless of indemnity

GPL-3.0 is a copyleft licence. Its obligations follow the code and how the enterprise uses and distributes it, not which tool generated it. A vendor indemnity is a contractual arrangement about claim costs. It does not change the licence status of the file, and it does not remove the obligations the enterprise has if GPL-3.0 code is present in its product. Someone in the organisation needs to own that analysis, separately from the indemnity question.

The wider allocation of liability is also unsettled. The Copyright Office's 2023 notice of inquiry asked how liability should be apportioned between the user whose prompt produced an output and the developers of the system and its dataset. The FTC raised similar open questions, including whether training-data disclosures should affect liability. No particular allocation should be assumed.

NIST's Generative AI Profile (initial public draft) points to the contract side of this: vendor terms that may amplify or defer liability in unexpected ways, and the value of a clear assignment of liability and responsibility for incidents. For a risk leader, that means reviewing each assistant's terms for what they assign, what they condition and what they exclude, then recording which conditions the enterprise can actually evidence.

Attribution inside a multi-assistant file

Do not assume that either assistant records line-level origin in a form you can use. Where several assistants edit one file, separating their contributions needs per-line or per-hunk origin records created when the code is generated. No standard mechanism for this is established, so it is an enterprise design decision. A file-level note saying that AI was used will not separate Copilot's lines from Amazon Q Developer's.

Human modification deserves its own record. Copyright guidance indicates that authors may claim copyright only for their own contributions to works containing AI-generated material, which supports distinguishing human-authored and human-modified portions from AI-generated ones. Separately, vendor terms may condition coverage on unmodified output, so an edit can change the indemnity position for the lines it touches.

NIST guidance on provenance is framed around training data and content, not code. Its direction still applies: maintain provenance, verify that it complies with infringement and IP terms, and periodically monitor outputs for possible infringement with a defined process for responding to claims. Preserve the records in a form that can be produced during a dispute, and set retention to match the period in which a claim could arise.

Where runtime policy enforcement and provenance fit

Settings that are configured but never evidenced are weak support. The stronger position is that policy was enforced at the moment of generation and that the enforcement itself produced the record.

Trussed enforces policy in the execution path for coding assistants, with actions that include allow, redact, block and escalate for human approval. Enforcement happens at runtime, not through review after the fact, and it applies to individual interactions. Trussed code provenance covers the origin, licence and security context of code. Enforcement generates audit-ready evidence that shows which policy applied and what the outcome was for each interaction. Trussed deploys self-managed in the customer cloud, with configurable retention and data residency, which bears directly on how long generation-time records can be kept and produced.

This does not replace the legal review of vendor terms or the GPL-3.0 analysis. It supplies the factual record that both depend on.