Review check

How to spot AI-generated reviews

More reviews now come out of a language model. A convincing one takes a minute: the right tone, a specific complaint, a tidy closing line. That reads well and judges badly — in a 2025 study, participants told real and generated product reviews apart with 50.8% accuracy on average, roughly what guessing would give you. So check the text, but check the reviewer and whether they ever used the product even more.

Direct answer: five checks on a review

1. Check the reviewer, not only the text. Has the account written other reviews, across different products and periods? An account that exists only for this one product is a known signal: the FTC names it explicitly as a reason to treat a review as suspicious.
2. Look for a burst in time. Many reviews within a few days, on a product that has existed for years, points to a campaign rather than customers. Compare the date distribution with similar products.
3. Read the text as text. Watch for the same sentence structures across reviews, repeated superlatives, a summary that describes no personal experience, and details that do not fit — a wrong model number, a wrong size.
4. Weigh negative reviews equally. Fake reviews are not always positive. A run of angry negative reviews sharing the same phrasing can be just as organised, for instance to damage a competitor.
5. Use the detector as triage. The text detector flags text carrying patterns common in model output. That is a reason to look closer — not proof that the review is fake.

Why this is harder now

Since the 2024 Consumer Review Rule, the US Federal Trade Commission prohibits creating, selling and buying fake reviews, and explicitly names reviews by someone who does not exist, such as AI-generated fake reviews (FTC, August 2024). Enforcement is not theoretical: in December 2025 the FTC sent warning letters to ten companies and pointed to civil penalties of up to $53,088 per violation (FTC business guidance, December 2025).

At the same time the text itself has become weak evidence. The study participants above performed barely better than chance, and language models given the same task did no better (arXiv 2506.13313, 2025). The FTC has long warned consumers that you often cannot tell from the text whether a review is real, and advises looking at the source and at the pattern over time (FTC, evaluating online reviews).

What a detector does and does not say

The text detector measures whether a passage carries patterns common in model text. A strong warning means: put this review beside the others and check the reviewer. A low score proves nothing, because edited, translated or formal human writing can pick up a signal too. Treat the result as triage, and read the measured limits as well.

For businesses and platforms

• Do not post reviews on behalf of customers, not even as an example or a draft.
• Never ask for a review in exchange for a discount, a free product or store credit; that is a violation, with or without AI.
• Remove reviews that did not come from a customer and record why you did.
• Reply professionally to a suspicious negative review and report organised campaigns to the platform.

More on judging sales material is in the guide on checking product photos before buying and on the marketplace checkers page.

Frequently asked questions

Does a high detector score prove a review is fake?

No. The score is a model-bound warning signal and says nothing about whether the reviewer actually used the product.

May I upload a competitor's review?

Only text you are permitted to use and without other people's personal data. Paste the review text rather than a screenshot showing a name and a photo.

What is the strongest sign of a fake review?

An account that exists only for this one product, a sudden burst of reviews in a short period, and wording that repeats verbatim across other reviews.