In three separate 2026 experiments, researchers asked people to sort genuine product reviews from ones written by AI. The average accuracy was 50.8%. That is not a typo, and it is not a rounding error. It is, statistically, a coin flip. The AI systems brought in to judge the same reviews did no better, and in some tests, worse.
If you have ever prided yourself on spotting a fake review by its stiff wording or suspiciously excessive enthusiasm, this is the research that quietly retires that instinct.
1. The numbers behind the coin flip
The study, testing whether humans and machines could tell real reviews from large language model fakes, found something more unsettling than simple failure. It uncovered a “scepticism bias”: people were reasonably wary of suspiciously glowing fake reviews, but far more likely to wrongly accept a fake negative review as genuine. In plain terms, a bad review is now the easiest kind of fake to get believed.
That single finding should worry any UK small business more than the headline coin-flip statistic. It means a rival, a disgruntled ex-employee or a bot farm can post fabricated one star reviews and count on readers being more credulous, not less, than they would be about a fake five star one.
2. The scale of the problem sitting above your head
This isn’t a niche academic concern. Amazon’s own Trustworthy Shopping Experience report says the company blocked hundreds of millions of fake reviews before they ever reached a product page in 2025, and estimates that shoppers still spent around $770 billion on products influenced by fake reviews that slipped through. Trustpilot removed 4.5 million fake reviews in 2025, 7.4% of everything submitted, with 90% of those caught automatically by its own machine learning systems.
| Platform | Scale of the fake review problem | Source year |
|---|---|---|
| Amazon | Hundreds of millions of fake reviews blocked before publication in 2025 | 2026 report |
| Amazon shoppers | An estimated $770 billion spent on products influenced by fake reviews in 2025 | 2026 report |
| Trustpilot | 4.5 million fake reviews removed in 2025, 7.4% of all submissions, 90% caught automatically | 2026 report |
Those are the platforms with dedicated fraud teams and in-house AI. A hairdresser, plumber or café with three staff and a Google Business Profile has none of that infrastructure standing between them and a coordinated burst of fake reviews.
3. Why small businesses feel this more than Amazon does
Amazon can absorb a bad month of slipped-through fakes without anyone noticing. A local business with 40 reviews cannot. A burst of five fabricated one star reviews in a week, aided by the scepticism bias working in their favour, can knock a business below the 4.5 star line that increasingly acts as a hard cut-off for UK consumers deciding who even makes the shortlist.
4. What actually works now
Academic researchers, including a team at the University of East London, have been building hybrid detection systems that combine AI language analysis with behavioural signals, such as whether the emotional tone of a review matches its star rating, or how a reviewer’s posting pattern compares with genuine customers. In testing, their model reached 93% accuracy on Amazon data and 91% on Yelp, far ahead of older keyword-based filters. The signals that hold up are not the ones most people instinctively reach for.
| Old “tell” shoppers relied on | Still reliable in 2026? | Why |
|---|---|---|
| Stiff or overly formal wording | No | AI-written reviews now read as naturally as genuine ones |
| Excessive enthusiasm or exclamation marks | No | People trust upbeat fakes more, not less |
| Verified purchase or booking tag | Yes | Still the single strongest signal a review is real |
| Reviewer account history and posting pattern | Yes | Core input for current detection models |
| Star rating vs emotional tone of the text | Yes | The mismatch is what the newest hybrid detectors flag first |
5. What this means for how you collect reviews
None of this is a reason to give up on reviews; if anything, it is a reason to be more deliberate about how they are gathered. Reviews that are tied to a verified booking, purchase or job carry more weight than open invitations for anyone to post. Replying to reviews with specific, real details about the job or visit does more to signal authenticity to a sceptical reader than any disclaimer could. And a sudden, unexplained cluster of reviews, good or bad, arriving in a short window is worth investigating rather than simply accepting at face value.
Conclusion
You cannot out-eyeball a coin flip, and neither, on current evidence, can the machines built to help you. The only defence that holds up is not being able to spot a fake review after the fact. It is being able to prove, from the moment it lands, that yours came from a real customer in the first place.
Read next