OpenAI's indemnity for generated output has named exceptions

The process wall

Before delivering work with generated parts, a briefing has to answer questions a client will eventually ask anyway: what was the tool trained on, who owns the output, and what happens if a third party objects. Adobe's own Firefly FAQ states that its models were trained “on a dataset of licensed content, such as Adobe Stock, along with public domain content,” and that Adobe does not “train on any Creative Cloud subscribers' personal content,” positioning this as part of designing Firefly “to be commercially safe” — a vendor claim, stated as such in Adobe's own terms.

What the documents show

OpenAI's Service Terms take a more specific, contractual approach for API and Enterprise customers: OpenAI indemnifies a customer against “any third party claim that Customer's use or distribution of Output infringes a third party's intellectual property right,” but the terms name exceptions, including where “Customer or Customer's End Users knew or should have known the Output was infringing” or where the “Output was modified, transformed, or used in combination with products or services not provided by or on behalf of OpenAI.” That second exception matters directly for a briefing: an indemnity keyed to unmodified output does not automatically follow the work once a maker edits, composites, or combines it with other material, which is close to always. Separately, the Copyright Office's AI initiative page records that the Office does not treat AI-generated material as protectable by copyright on its own, and points to registration guidance requiring disclosure of that material rather than an assumption that a licence resolves the authorship question.

Authorship and limitation

These are three different kinds of promise: Adobe's training-data description is a factual claim about inputs, OpenAI's indemnity is a contractual risk transfer with stated carve-outs, and the registrar's guidance is a legal requirement independent of either vendor's terms. None of the three eliminates the others' relevance. A commercially-trained-data claim does not itself establish that a client owns copyrightable rights in the output, and a vendor indemnity does not itself satisfy a registration's disclosure duty.

What survives

This is an editorial checklist, not a substitute for the underlying documents: a briefing worth giving a client should separately cover which tool generated which part, what that tool's own terms say about training data and indemnity (naming any exceptions, not just the headline promise), whether the delivered work still qualifies for the indemnity after editing, and whether any AI-generated portion needs to be disclosed if the client intends to register the work.

• Ask the tool's own FAQ or terms what its training data is described as, and note whether that description is stated as a vendor claim.
• Read an indemnity's exceptions, not only its headline grant, especially any exception for modified or combined output.
• Flag any AI-generated portion that would need to be named and excluded on a registration application, per the Office's guidance.

A client briefing that only repeats a vendor's headline promise skips the exceptions, the modification carve-out, and the registration duty that sit right beside it in the same documents.