AI-Generated Reviews Are Here: How Detection Is Changing

Google now runs Gemini's reasoning over Maps content to catch nuanced policy violations, review-extortion scams included, before they publish rather than after. The 2025 enforcement figure behind that effort is 292 million reviews blocked or removed. It is a genuine step up in platform-side moderation, and it aims at fraud in general rather than at the narrower question of whether a particular review was written by a machine, which is getting harder every year.

What follows: what Google announced, why generated text breaks one detection method and leaves the others intact, and what that changes for anyone judging a review's authenticity.

#What did Google actually announce about fighting review fraud?

On April 16, 2026, Google said it was applying Gemini's reasoning to catch policy-violating reviews and edits faster and with more context than keyword rules allow, including content whose problems are local or situational.[^1] Verified business owners also began getting proactive email alerts about suggested edits before those edits go live. The company said its systems now recognize specific scam patterns, among them review extortion, where someone demands payment to take down a bad review.

Our systems now better detect specific scam patterns, which lets us stop suspicious posts before they ever go live.

The same announcement carried the 2025 numbers: 292 million policy-violating reviews blocked or removed, a billion reviews Google counts as genuinely useful published, 79 million inaccurate or unverified edits blocked, 782,000 accounts restricted, 13 million fake business profiles removed.

#Is this the same thing as detecting AI-generated review text?

No, and the difference matters. Google described reasoning-based moderation for policy violations and scams: extortion, vandalism-style edits, coordinated campaigns. Nothing in it classifies text as human-written or model-written. Treating the announcement as an answer to the AI-text problem claims more than Google said.

That problem is narrower and nastier. A language model will write four hundred five-star reviews that share no phrase, each sounding specific, at a cost close to nothing. Detection built on matching wording across reviews has nothing left to match.

#Does the FTC's rule cover reviews written by AI?

Yes, in effect. The rule bans publishing a review that claims genuine experience the writer never had, and the provision does not ask who or what typed it.[^2] A company generating five-star reviews attributed to customers who never existed breaks the same clause as one paying a person to invent an experience. Penalties land identically either way, as the rule's five prohibitions make clear.

#If duplicate-text detection stops working, what still catches AI-written fake reviews?

The signals that never looked at wording. A campaign still has to post its reviews, and posting behavior does not improve just because the prose did. Twenty reviews inside an afternoon is still a spike. An account with no other reviews, no photos, and no Local Guide history is still an empty account, whoever or whatever wrote the sentences on it. The piece on burst days explains how that timing signal is measured.

Our own ordering already reflects this. As of August 2026, empty accounts account for 75,255 of the 100,021 flagged reviews across 450 analyzed places, ahead of spikes at 23,361 and repeated wording at 15,103. Account behavior and timing carry the weight, and generated text will only push the balance further that way.

#What should a business or consumer actually do differently because of this?

Stop reading fluent, specific prose as proof of a real visit. That inference is the one generative tools have destroyed, since text of that quality now costs nothing to produce in bulk. Judge the account instead: how much history sits behind it, what else it has reviewed, whether photos exist. Then judge the clock: does this review sit inside a cluster, or does it fit the pace the business normally sees?

Our analyzer weighs account history and timing alongside text patterns for exactly that reason. A detector built only to catch repeated phrasing has a shrinking useful life.

#Frequently asked questions

#Can I tell an AI-written review apart from a real one just by reading it?

Not reliably. Well-prompted output is specific, varied, and plausible. Reading quality stopped being evidence of authenticity, which is why account history and posting time carry the weight now.

#Does it matter, legally, whether a fake review was written by a person or an AI tool?

No. The rule prohibits reviews from people with no genuine experience, whatever produced the text.

#Has Google said how much of the fraud it blocks is AI-generated specifically?

No. The April 2026 announcement and the 2025 figures cover policy violations broadly, with no breakdown for generated text.

#Will duplicate-text detection become useless as AI writing improves?

Less reliable alone, not useless. Cheap campaigns still recycle templates rather than generating something new for every review, because recycling costs less to run.