How do I spot an AI-generated product review before I buy?

Treat it like a verification job and look at the account, not the prose. In a newsroom we never decide on how a claim sounds. We check the source.

For a review that means: how old is the reviewer account, what else have they reviewed, and do they review wildly different product categories in the same week? If photos are attached, run a reverse image search. Stock or reused photos are a stronger signal than any sentence. Wording is the weakest evidence you have.

On detectors: I would not use one for this. Reviews are a few sentences long, and that is the regime where text classifiers have the least to go on. A short, positive, generic text can come from a person, a chatbot, or a person using a chatbot, and a score cannot separate those cases. If you take one thing from the accuracy thread, it is that short text is the hardest case.

Use the score as a reason to look harder at the account, not as a verdict on the review.

the tell i look for is perfect grammar. real buyers type things like "arrived tuesday box was crushed but the thing works". anything too polished and balanced with exactly three pros and one tiny con is a fake in my experience. i skip those and only read the messy ones.

@Jules and Aisha, both of you are partly right, and it is a useful split. Polished text is a weak signal in either direction, and I am going to add Aisha's point to the guidance we write for shoppers. Messy-looking reviews are also not safe, because fake ones can be made messy on purpose.

Please do not teach people that perfect grammar means fake. I am a non-native speaker and I write my reviews carefully, because I am afraid of sounding rude or confused in English. I often use a grammar tool as well. A careful review from me would look "too polished" by that rule. Messy does not equal honest, and polished does not equal fake.

I get offered work writing reviews. I say no. What gives them away is not the style, it is the pattern. Five reviews in two days, all in the same structure, all praising the same two features, all posted just after a bad one. A person does not do that. Look at the dates before you read the words.

Threat model it. Who benefits from this review being believed? The seller, a competitor, or a review service. Each has a different signature. Sellers push volume at launch. Competitors push one-star bursts. Services spread across many accounts so no single one looks odd.

For the buyer: read the three-star reviews, they are the least likely to be paid for. Look for details that cost effort to invent, like how the thing failed after two months, or how it compares with the previous model the reviewer owned. Specifics are expensive to fake at scale.

@Aisha R. oh, that is fair, i did not think about people who check their writing carefully on purpose. i take it back, "perfect grammar" is a bad tell. dates and account history are better, like marco said.