A new kettle has 4.8 stars and nine hundred glowing write-ups. Fake reviews are as old as online shopping, but they used to cost effort: someone had to sit down and type them. SI (AI) text generators removed that cost. A convincing, varied, typo-sprinkled review now takes seconds to produce, in any language, in any quantity.
This guide explains what SI-generated reviews are, what the law in the United States, the European Union and the United Kingdom actually says about them, what the big review platforms report, and how a shopper or a shop owner can read reviews with a cooler head. It also explains why an SI text detector is a weak tool for this particular job.
The short version
- The problem is the lie, not the tool. A review is fake when it claims an experience that nobody had. SI just makes that lie cheap to mass-produce.
- It is illegal in the US, the EU and the UK for a business to write, buy or commission fake reviews, whoever or whatever typed them.
- People cannot spot them by reading. In a Yale experiment, readers judged SI-written restaurant reviews to be human more often than the real ones.
- Signals around the text are stronger than the text. Reviewer history, dates, the shape of the star ratings and verified-purchase labels tell you more than wording.
- Detectors struggle with short text. Most reviews are a few sentences long, which is too little for any statistical detector to judge.
What counts as an SI-generated review?
Two different things get called "SI reviews". Only the first is a problem.
- Fabricated reviews. Text generated to look like it comes from a customer who does not exist, or who never used the product. This is the fake review in the legal sense.
- Real experiences, polished by SI. A genuine customer asks a chatbot to tidy their grammar or translate their opinion. The experience is real, so the review is not fake, even if a detector might flag the wording.
The UK regulator's definition is the clearest one to keep in mind. In its guidance of 4 April 2025, the Competition and Markets Authority (CMA) writes that "a fake review is a consumer review that purports to be, but is not, based on a person's genuine experience". Nothing in that sentence mentions software: what SI changes is volume, speed and polish.
How common are they?
Nobody knows the true number, because the successful fakes are by definition the ones that were not caught. What we do have are the platforms' own transparency reports, which show the scale of what they remove.
| Platform (report) | Reviews in 2024 | Fake or fraudulent reviews removed | What it says about SI |
|---|---|---|---|
| Tripadvisor (Transparency Report, 18 March 2025) | 31.1 million submitted | 2.7 million rejected or removed | 214,000 reviews flagged and removed as SI-generated |
| Trustpilot (Trust Report, 29 May 2025) | 61 million posted | 4.5 million, or 7.4% of submissions | No count of SI-written reviews; 90% of fakes removed automatically |
Two details are worth noticing. First, on Tripadvisor the reviews identified as SI-written were a small slice of the 2.7 million fraudulent ones. Most review fraud is still defined by who posted it and why, not by how the words were produced. Tripadvisor says that "review boosting", where a business tries to lift its own ranking, made up 54% of the fraud it found in 2024. Second, Tripadvisor removes SI-written reviews as a matter of policy, to spare readers what it calls a "sea of sameness". Such a count reflects wording judged to be machine-written, which is not the same as an invented experience.
The CMA says around 90% of consumers use reviews, and it estimates that as much as £23 billion of UK consumer spending a year is potentially influenced by them.
Can people tell an SI review from a real one?
Not reliably. Balázs Kovács, a professor at the Yale School of Management, took 100 real Yelp restaurant reviews from 2019 and asked GPT-4 to write reviews of the same restaurants, at about the same length and in the same style, including human quirks such as typos and capitals for emphasis. He then showed mixed sets to 151 paid participants and promised a bonus to anyone who classified 16 out of 20 correctly.
Only 6 of the 151 earned the bonus. According to Yale's account of the study, published on 28 May 2024, participants recognised the human-written reviews about half the time, which is no better than a coin flip, and recognised the SI-written ones only about a third of the time. In other words, readers were more likely to believe the machine. "I wouldn't have expected it to be more human than human," Kovács said.
As with SI scams, voice cloning and fake profiles, the old tells, such as clumsy grammar, are gone. A review that "sounds real" proves nothing.
What the law says
Three large markets have moved in the same direction. This is a summary for general understanding, not legal advice.
United States: the FTC rule
The Federal Trade Commission announced its Rule on the Use of Consumer Reviews and Testimonials on 14 August 2024, after a 5-0 vote, and it went into effect on 21 October 2024. The announcement names the technology directly: the rule covers reviews "that misrepresent that they are by someone who does not exist, such as AI-generated fake reviews".
The rule text (16 CFR Part 465) makes it a violation for a business to "write, create, or sell" a consumer review that materially misrepresents that the reviewer exists, that the reviewer used the product or service, or what their experience was. Buying such reviews is covered when the business "knew or should have known" they were false. The rule also bans paying for reviews on condition that they express a particular sentiment, undisclosed reviews by insiders, company-controlled sites posing as independent, suppression of negative reviews through threats, and the sale or purchase of fake followers and views. Courts can impose civil penalties for knowing violations.
Asking customers for reviews is fine, and so is offering an incentive, as long as there is no requirement, stated or implied, that the review be positive. And a website that does nothing more than host reviews written by others falls under an exception: the FTC's own guidance says there is no requirement that a business hosting reviews investigate whether they are fake.
The rule targets conduct, not software. In 2024 the FTC had obtained a consent order against Rytr, an SI writing service whose tools, it alleged, let subscribers generate false and deceptive reviews. On 22 December 2025 the Commission set that order aside, stating that the facts alleged did not support a violation and that the order "unduly burdens innovation in the nascent AI industry". That decision concerned the tool maker; the reviews rule described above is a separate text.
European Union: the Omnibus Directive
Directive (EU) 2019/2161, often called the Omnibus Directive, amended the Unfair Commercial Practices Directive and has applied since 28 May 2022. It added two entries to the list of practices that are unfair in all circumstances:
- stating that reviews come from consumers who have actually used or purchased the product "without taking reasonable and proportionate steps to check that they originate from such consumers";
- "submitting or commissioning another legal or natural person to submit false consumer reviews or endorsements, or misrepresenting consumer reviews or social endorsements, in order to promote products".
It also makes it material information "whether and how the trader ensures that the published reviews originate from consumers who have actually used or purchased the product". In practice, a shop that shows reviews to EU consumers should say how it checks them. The European Commission's guidance notice of 17 December 2021 covers consumer reviews and endorsements.
United Kingdom: the DMCC Act
The Digital Markets, Competition and Consumers Act 2024 introduced a banned practice on reviews, and the CMA published its guidance on 4 April 2025. The ban covers submitting or commissioning fake reviews, concealed incentivised reviews, and publishing reviews or review information in a misleading way. Among the CMA's examples of what is prohibited: buying or selling "reviews which look like they have been written by individual consumers but have in fact been generated by software applications (such as bots)".
The UK goes further than the US rule on one point: publishers must take reasonable and proportionate steps to prevent and remove banned reviews. The guidance expects a published policy, risk assessments, checks and a way for users to report. The CMA is clear that publishers "should not rely on users (especially consumers) to police" the problem. On 6 June 2025 the CMA announced that Amazon had signed undertakings to detect and remove fake reviews and "catalogue abuse", where sellers move the reviews of a well-performing product onto a different one, and to sanction sellers and reviewers who break the rules. Google had signed undertakings in January of that year.
How to spot suspicious reviews
Combine weak clues from the text with stronger clues from everything around it. None proves anything on its own.
Signals in the text
- No checkable detail. Real customers mention the size they ordered, the delivery delay, the thing that broke in week three. Generated praise tends to restate the product description: "sleek design", "exceeded my expectations", "highly recommend".
- Sameness across reviews. One polished review means nothing. Twenty reviews with the same length, the same three-part structure and the same closing recommendation are a pattern. This is the "sea of sameness" Tripadvisor describes.
Our guide on how to spot SI-written text covers these habits in more depth. Remember the Yale result, though: a model told to imitate human quirks hides them.
Signals outside the text
These are the patterns that regulators and platforms themselves rely on. The CMA's guidance lists, as examples of what automated checks look for, reviews written by different reviewers from the same email or IP address, networks of reviewers reviewing the same products or businesses, average review length, and "a spike in highly positive or negative reviews over a short period of time". You cannot see IP addresses, but you can see a lot:
- Dates. Sort by newest. A burst of five-star reviews in a few days, after months of silence, deserves suspicion.
- The reviewer's profile. A history of one review, or of dozens of five-star reviews for unrelated products and distant cities in the same week, is a warning. The CMA notes that letting users see a reviewer's public history helps them decide what to rely on.
- The shape of the ratings. A wall of five-star ratings with almost nothing in between is worth a second look. Read the middling reviews too: they often contain the concrete detail that the glowing ones lack.
- Does the review match the product? Reviews praising a phone case on the page of a pair of headphones point to the catalogue abuse the CMA described.
- Photos. A customer photo can itself be generated or lifted from elsewhere. If an image matters to your decision, you can run it through an SI image detector and a reverse image search.
Can an SI detector catch fake reviews?
Only partly, and it is better to be clear about that. Text detectors work on statistics: sentence rhythm, vocabulary variety, typical model phrasing. As we explain in how SI detectors work, those statistics need enough text to mean anything. Many reviews are only a sentence or two. Our own SI text detector asks for at least 80 words for that reason, and most reviews never reach it.
A detector result on a review answers "does this wording look machine-written?", not "did this person buy the product?". A genuine customer who used a chatbot to translate or tidy a review may be flagged. A fake written by a paid human, or by a model told to write casually, may pass. Never accuse a named reviewer on the strength of a score alone.
Where a detector can help is with longer material, such as a 300-word testimonial on a landing page or a batch of long reviews that arrived in one week. In those cases a free SI checker gives you one more clue to weigh alongside the dates and profiles above. For the general limits of these tools, including false positives, see are SI detectors accurate?
Our own test: one long fake review, one short one
We asked an SI model (Claude, by Anthropic) to write a glowing review of a cordless vacuum cleaner, the kind a seller might order in bulk. It produced 132 words, beginning:
"I recently purchased this wireless vacuum cleaner and I am absolutely thrilled with my decision. [...] The suction power is truly impressive, effortlessly picking up dust, crumbs and pet hair from both carpets and hard floors. [...] Overall, this vacuum cleaner offers exceptional value for money, and I would highly recommend it [...]"
Result: 72% SI, "Likely SI-generated"
- Sentence rhythm: length variation 0.18, average 18.9 words per sentence, which is very even.
- Uniformity: a single block of text, one list of three, three transition openers.
- Human marks, pulling the other way: ten first-person words. A review is written in the first person by nature, so this signal lowers the score even on machine text.
The verdict is correct, but only just: 72% sits barely above the 70% line where our text tool says "likely SI". A slightly more casual version could land in the uncertain zone.
Then we pasted a 20-word review ("Great vacuum, strong suction and the battery lasts a long time. Very happy with it, would buy again. Five stars."). Our tool declined to score it: it asks for at least 80 words, because shorter texts do not carry enough signal. Most real reviews are that short, so for a typical review a text detector, ours included, has nothing to say. The signals around the review matter more than the words in it.
If you run a shop or a review page
- Never generate, buy or swap reviews. It is unlawful in the US, the EU and the UK, and "an agency did it" is not a defence when you knew or should have known.
- Ask every customer, not only the happy ones, and never tie a discount or gift to a positive rating. If you offer an incentive, say so clearly on the review.
- Say how you check reviews. EU law treats that information as material. If you claim reviews come from real buyers, you need reasonable steps behind the claim, such as linking reviews to orders.
What to remember
SI did not invent the fake review. It made it cheaper, faster and better written. In the US, the EU and the UK, a review that pretends to reflect an experience nobody had is unlawful, however it was typed. Detection by reading is the weak link, because reviews are short. So look less at the adjectives and more at who wrote them, when, and how the ratings are distributed.
