AI Product Descriptions That Invent Claims: An Audit Workflow
Generated descriptions do not usually lie loudly. They add a material, a certification, or a benefit nobody supplied. A practical audit that catches invented claims before they publish.
The generated description read well, which was the problem. Somewhere in its second paragraph the jacket had become waterproof rather than water resistant, acquired a recycled fill that the spec sheet never mentioned, and been described as suitable for temperatures no one had tested it at. Nobody wrote those claims on purpose. The model filled gaps in the product data with plausible sentences, and plausible is exactly what makes them hard to catch in a quick read.
This is not an argument against generating descriptions. It is an argument for treating a generated description as untrusted copy until each factual sentence has been traced back to a source field. Voice, structure, and benefit framing can be generated freely. Facts cannot. A claim about material, weight, certification, compatibility, or performance has to come from data the merchant actually holds.
Audit by claim, not by paragraph
Read the draft and mark every sentence that asserts something checkable. Keeps drinks cold for 24 hours is a claim. A brushed finish is a claim. Machine washable is a claim with care consequences. Sentences that merely frame or connect can pass, but anything a customer could rely on, return against, or complain about needs a source: the supplier spec, the lab result, the care label, the manufacturer's own listing.
In practice this is faster as a side by side than as a read through. Put the source attributes on one side and the generated text on the other, and delete or correct every assertion with no counterpart. The deletions cluster in predictable places: durability, sustainability, health, and compatibility, because those are the benefits copy is expected to contain and the data is least likely to supply.
The data gap is the real finding
If a third of the catalogue is generating invented sustainability claims, the root cause is an empty field, not a badly behaved model. The generator was asked for a benefit led description of a product whose record contains a name, a price, and an image. It produced the description it was asked for. Fixing the prompt to say do not invent facts helps at the margin, and adding structured attributes for material, certification, and care helps a great deal more.
Measure it that way. Track the percentage of generated drafts that survive the claim audit unchanged, by category, before and after the attribute cleanup. That number tells you whether you have a writing problem or a catalogue data problem, and for most stores we have looked at in this kind of audit, it is the catalogue. Publishing the drafts anyway transfers the cost to returns and support.
High risk categories need a human gate
Some claim types should never publish on generation alone. Health and safety statements, children's product claims, allergen and ingredient statements, electrical and fire safety, and anything with a regulated wording need a named human approving the final text. That is not caution for its own sake. An invented hypoallergenic or non toxic claim is a regulatory and liability problem, not a copywriting slip.
Set the gate by claim type rather than by product, and log who approved what. When a claim is later challenged, by a customer, a platform, or a regulator, the useful question is not whether a model wrote the sentence. It is whether anyone checked it, and an approval log is the only evidence that anyone did.
Related reading: AI Commerce Watch, Answers, Glossary.
Frequently asked questions
What kinds of invented claims are most common?
Material and composition, performance figures such as waterproofing or insulation hours, sustainability and certification claims, and compatibility statements. They cluster where benefit led copy is expected but the product record has no supporting data.
Can a prompt instruction stop a model inventing facts?
It reduces it but does not eliminate it, especially when the product record is sparse. The reliable control is a claim level audit against source attributes, plus filling the data gaps that invite invention.
Which descriptions need human approval before publishing?
Any containing health, safety, allergen, children's product, or regulated claims, and any performance figure a customer would rely on. The approval should be by a named person and logged.
How should a store measure whether this is improving?
Track the share of generated drafts that pass a claim audit unchanged, broken down by category. A rising pass rate after attribute cleanup indicates a data fix. A flat rate indicates the audit is the only thing standing between the store and invented copy.