Can you tell if a review was written by AI?
By the Ingenious AI team · Published 14 November 2026 · Reviewed 14 November 2026
Run the free review audit See pricing · See a sample provenance report
Short answer
You cannot prove a review was AI-written from its text — and ReviewProof never claims to. What you can detect are patterns associated with generated text: stock phrases that generators reach for, and, in longer reviews, a high density of generic praise words combined with low vocabulary diversity. Those patterns are flagged as an AI-generation signal, weighted 0.6 in the risk score, and presented for human review. The FTC rule bans AI-generated fake reviews, which is why the pattern is worth surfacing — and worth labelling honestly.
What is an AI-generation signal?
An AI-generation signal is a textual pattern correlated with machine-generated writing: formulaic phrases such as 'exceeded my expectations in every way' or 'seamless experience from start to finish', or statistical flatness — many generic superlatives, few distinctive words. It is a signal about the writing, not a finding about the writer: enthusiastic humans use these phrases too, which is precisely why the honest product language is 'patterns associated with generated text'.
Why the wording gives generators away
Generated praise is optimised to sound like praise in general, so it reaches for the phrases a thousand other reviews used: the pleasure of the visit, the staff who went above and beyond, the wholehearted recommendation. Specificity is what it lacks — no table number, no rainy Tuesday, no name for the thing that went slightly wrong before it went right. ReviewProof matches a maintained list of these stock phrases and, for reviews of twenty-five words or more, also measures generic-word density above 6% together with vocabulary diversity below 0.45. Either pattern raises the signal; neither, alone, convicts anyone of anything.
ReviewProof computes every signal from the rows you import, never from a reputation guess about your business. The risk score is 100 × (gating × 1.0 + incentive × 1.0 + duplicate × 0.8 + AI-signal × 0.6) ÷ (2.6 × number of reviews), plus 15 points when a burst is detected, capped at 100. Each flagged review in the report names the imported row that triggered the signal, so a human can open that row, read it in context, and decide. The product labels these patterns signals for human review, not verdicts — and that restraint is deliberate, because a heuristic that overclaims would be its own kind of false evidence.
How to act on the signal without abusing it
Treat AI-flagged rows as a cluster question, not a per-reviewer accusation. One flagged row in four hundred is a human who writes formally. Thirty flagged rows arriving in a burst, all UNLINKED, all five stars, is a purchased package wearing generated text — and the finding is the cluster, documented as a whole. Never confront a named customer with an AI accusation based on wording; if the rows are yours to moderate, flag them internally with the pattern noted, and if they are a platform's, send the platform the cluster and the evidence, not an insult to a possibly innocent writer.
In practice the routine takes minutes. Export or copy the reviews for the location, import them into ReviewProof by CSV or paste, and read the report top-down: total rows, flagged count, risk score, then the per-signal breakdown. Open the flagged rows themselves — the report traces each signal to its row precisely so you can do this. Decide moderation openly (kept, published, or flagged internally; there is no delete), check which reviews carry a transaction link, and publish or refresh the verified widget so the public claim matches the audited record. On Pro, export the evidence pack while the file is fresh.
Mistakes that undo the work
The first mistake is marketing language that overclaims the signal ('our AI detects AI reviews with certainty') — it is false, and it will be quoted back at you. The second is deleting flagged rows, which punishes the occasional formal-writing human along with any cluster. The third is ignoring a rising signal count because the rating improved: the count trending across imports is the early-warning line, and it is free to watch.
Two design choices keep the detection honest. First, nothing is scored until you import it: there is no scraping of your public profiles and no third-party score blended in. Second, the weights reflect what the FTC fake-review rule (16 CFR Part 465) actually prohibits — how a review was solicited weighs heaviest (gating and incentive markers at 1.0 each), while wording patterns weigh less (duplicates 0.8, AI-generation patterns 0.6). A burst of same-day reviews adds a flat 15 points because timing is visible in the dates you supply. You can reproduce every number in the report from your own file, which is exactly the point.
Worked example: a home-services firm audits a subcontractor's promise
A fictional worked example, written to show the numbers and the decisions in practice.
A home-services firm imported 307 reviews after a marketing subcontractor reported a ratings jump. The report showed 44 rows with AI-generation signals — against a baseline of 3 in the firm's earlier imports — concentrated in six weeks, 39 of the 44 UNLINKED, with duplicate phrases recurring across supposedly different customers. The firm's conclusion, logged in the record, was narrow and provable: a cluster of generated-text reviews with no transaction backing had entered the profile during the subcontractor's tenure. That documented cluster, exported as an evidence pack, ended the contract discussion in one meeting — no individual reviewer was ever accused of anything.
Your five-action checklist
- Read the AI-signal count on every import and note its trend, not just its level
- Open flagged rows in clusters — judge concentration, dates and links, not single rows
- Check whether flagged rows are transaction-linked or UNLINKED
- Flag problem clusters internally with the pattern and your reasoning logged
- Never accuse an individual reviewer of using AI on wording evidence alone
Frequently asked questions
Does an AI signal prove a review was AI-written?
No, explicitly. It indicates patterns associated with generated text — stock phrases, generic density, low diversity — which humans also produce. The report labels it a signal for human review.
What exactly does ReviewProof measure?
Known generated-text phrases, and for reviews of 25+ words, generic-word density above 6% combined with vocabulary diversity under 0.45 (unique words divided by total words).
Are AI-generated reviews actually prohibited?
Yes, when they are fake reviews — the FTC rule (16 CFR Part 465) bans AI-generated reviews that misrepresent real experience. A genuine customer's own words remain a genuine review.
How much does the signal move the risk score?
At weight 0.6 per affected row — the lightest of the four wording signals, reflecting that it is the least conclusive — inside the standard formula.
Keep reading
Related guides from the ReviewProof library:
- How can you prove a review is real?
- How do you detect fake reviews in your own profile?
- What is review gating, and how do you stop doing it?
- What counts as an incentivised review?
- What does duplicate review text tell you?
- What is a review burst, and when is it innocent?
- How is the ReviewProof risk score calculated?
Start with the tool itself: open the free audit desk, see how ReviewProof works, compare Free, Pro and Agency plans, inspect the sample provenance report, or browse all ReviewProof guides.