Green claims you can prove.

Analysis

When marketing copy is generated by AI, who checks the green claims?

Product descriptions written by AI multiply environmental wording at a speed no legal team can review. The failure modes, and the control points that scale with the volume.

Jeremy Fain · 3 min read

Answer in 40 words: the company publishing the copy is responsible, whoever or whatever drafted it. Generative tools change the volume and the failure mode, not the liability: they produce fluent, plausible environmental wording that no one selected deliberately and no one can trace afterwards.

Your own pages: the free scan reads up to five of your addresses and links the claims it finds to the rules that concern them, with their source.

Run a free scan

The three failure modes specific to generated copy

Fluent escalation: a factual input such as twenty-five per cent recycled content becomes eco-friendly packaging in the output, because the model writes what marketing copy usually sounds like.

Silent generalisation across a catalogue: one prompt applied to two thousand references propagates the same wording to products whose evidence differs.

Untraceable authorship: no one remembers which version of which prompt produced the sentence that is now printed, which makes the substantiation question unanswerable.

Inherited errors at scale: a wrong formulation approved once becomes the template for every product description generated afterwards.

Why sampling does not work here

Manual review of a sample was a reasonable control when a human wrote every description at human speed, because errors were sparse and independent. Generated copy fails systematically: if the prompt allows a phrase, it appears everywhere at once. A sample of twenty out of two thousand tells you almost nothing about the other one thousand nine hundred and eighty, and gives false comfort.

The control points that scale

Constrain the input: a list of authorised formulations and a list of prohibited terms, injected into the generation, is more effective than reviewing the output.

Check at publication, not at drafting: the sentence that matters is the one that reaches the customer, after every downstream transformation.

Keep the link between claim and evidence, so that a claim generated for a product can be traced to the file that supports it.

Re-check when the model or the prompt changes, because a wording that passed last month may not be the wording produced today.

Frequently asked questions

Can an AI assistant assess its own output? It can propose an assessment, which will vary between sessions. What it does not do is retain the assessment, its basis and its date.

Should generated copy be avoided for green claims? Avoiding it is rarely realistic. Constraining it and checking at publication is.

Who signs off? The same person as before. The tool changed, the accountability did not.

Does a human review of ten per cent suffice? Not when the error is systematic rather than sparse, which is precisely what generation produces.

ClaimTrace provides regulatory information: this article cites and dates texts, and draws no conclusion on any company's situation.
Portrait of Jeremy Fain
About the authorJeremy FainFounder of ClaimTrace

Has spent nearly twenty years building software that traces, governs and anticipates. Co-founded Verteego, a publisher of environmental traceability software, in 2007 and chaired it for eleven years.

More about Jeremy Fain
Stay informedOne email at each new stage of the service, never more than one a month.

You will confirm by a link sent by email, and can unsubscribe in one click. The form is loaded from Brevo, our email provider, only when you click.