TripAdvisor’s own 2025 transparency report puts a number on something a lot of Bristol travellers have half suspected for a while. The platform removed 214,000 reviews it believed were written by AI in a single year, part of 2.7 million fraudulent submissions taken down in total, according to the 2025 transparency report covered by CNBC. That works out to roughly 4,000 AI-flagged reviews a week disappearing from a site millions of people still use to pick a hotel. Some of what didn’t get caught is sitting on Bristol hotel and attraction pages right now, dressed up as somebody’s genuine weekend in the city.
My take after going through the research for this piece: chasing a single suspicious review is a waste of time. The useful skill is reading the pattern across ten or twenty reviews at once, and knowing which handful of signals actually correlate with fakery rather than just feeling a bit off.
Why this is suddenly Bristol’s problem too
Independent research firm Originality.ai analysed TripAdvisor review text going back to 2019 and found the share flagged as likely AI-generated rose from 4.49% to 10.7% by 2024, a 137% increase in five years. Reporting from the Seattle Times in September 2026 pushed the number higher still for the luxury end of the market: more than 20% of luxury hotel reviews in 2025 were judged AI-generated, double the 2022 rate, and 57% of those fakes carried five stars. A WalletHub survey cited in the same piece found 83% of Americans are already worried about being manipulated by AI content when they shop or book travel.
Bristol doesn’t need imported statistics to make the point, though. Back in 2019, before generative AI was writing reviews for anyone, the owner of the Urban Kohinoor in Clifton found his restaurant had dropped 38 places in the city’s TripAdvisor rankings after a one-star review accusing staff of sexism. He noticed the same reviewer, using the same wording, had posted an almost identical complaint about the unrelated Bristol curry house Brunel Raj on the same day. TripAdvisor investigated and ultimately sided with the reviewer, which is its own useful lesson: a pattern that looks damning to a business owner isn’t always what a platform’s internal data shows. Bristol has form for contested reviews. It just didn’t need AI to produce them seven years ago.
What actually gives an AI-written review away
Scott Dylan, who runs an AI venture fund and studies these patterns for a living, told the Seattle Times that AI text has a particular texture. “Real people mention staff names or specific, messy details,” he said. “AI prefers generic praise like ‘exceptionally attentive.'” That distinction holds up across the academic literature too. A 2025 machine-learning study in the journal Information Technology & Tourism found AI-generated hotel reviews are shorter, more repetitive in vocabulary, and heavier on generic sentiment words, where genuine reviews contain more specific nouns: cost-effectiveness, particular room numbers, named amenities.
One of the sharper examples in the reporting came from Nicky Zhu, a product manager who went looking for exactly this kind of tell. She found a beachfront resort with 47 consecutive five-star reviews that all praised a restaurant three blocks away. That restaurant had been closed for two years. “Real people could not have written these reviews,” she said. AI review bots repeat outdated or wrong details because they’re pattern-matching against old training data or a lazy template, not an actual stay.
Academic researchers who built the MAiDE-UP dataset, a set of 10,000 real and 10,000 AI-generated hotel reviews across ten languages, found something similar: AI text tends to be more descriptive per sentence but less readable overall, using more adjectives and adverbs while real reviews use more numerals and specific nouns. Their human evaluators correctly identified AI-written reviews only 71.5% of the time, and the failure mode ran one direction. People were far more likely to mistake a fake review for a real one (a 60% false-negative rate) than the reverse. That asymmetry is worth sitting with: your gut is worse at catching fakes than it is at wrongly accusing genuine reviewers.
| Signal | What a genuine review usually looks like | What raises suspicion |
|---|---|---|
| Specific detail | Names a staff member, a room number, a specific noise or view | Generic praise like “exceptionally attentive staff” with nothing verifiable |
| Posting timing | Trickles in over weeks, roughly matching guest turnover | Dozens of reviews arrive within hours or a single day |
| Star rating spread | A mix of 3, 4, and 5 star reviews with specific complaints mixed in | Almost everything clusters at 5 star or 1 star, little middle ground |
| Reviewer history | The account has reviewed other unrelated places over time | Brand new account, or one that has only ever reviewed this single property |
| Sentence rhythm | Uneven length, informal asides, the occasional typo | Uniform sentence length and unusually polished, superlative-heavy phrasing |
Sources: Seattle Times reporting (Sept 2026), MAiDE-UP multilingual hotel review dataset (NAACL 2025), and TripAdvisor’s own published detection methodology as described by industry outlet Review Sell (April 2026).
Check the clock, not just the words
Timing is the tell that AI struggles hardest to fake, because a model can produce flawless sentences but it can’t manufacture a plausible arrival schedule on its own. TripAdvisor’s fraud detection stack, described in detail by the trade outlet Review Sell, runs a “velocity anomaly” model against a learned baseline for how fast reviews normally arrive at a given property. Anything more than two or three standard deviations outside that baseline triggers a manual look, whether the spike is suspiciously positive or a sudden flood of one-star reviews timed to hit right after bad press.
This isn’t a new trick platforms only just discovered. Consumer group Which? analysed almost 250,000 TripAdvisor reviews back in 2019 and found that at two Travelodge branches, 40% and 48% of five-star reviews came from accounts that had never reviewed anywhere else. Premier Inn, reviewed for comparison in the same investigation, showed nothing unusual. One in seven of the top-rated hotels Which? checked worldwide had what it called “blatant hallmarks” of fake reviews. The method still works today. If you open a hotel’s review page and see a cluster of five-star reviews from accounts with a join date the same week as the review, and no other reviews anywhere else on their profile, that’s the single most reliable free tell available to an ordinary traveller.
Running suspect text through an AI detection tool
If a review still looks off after the timing check, copying the text into an AI detection tool is a reasonable next step, though it’s worth knowing what these tools were actually built for. Most detectors, zerogpt included, were trained to flag AI-written essays and articles, which run to hundreds of words. A hotel review is often three sentences long, and short text gives any detector far less to work with, so treat the output as one data point rather than a verdict.
There’s a second complication that’s easy to miss. TripAdvisor’s own head of trust and safety, Becky Foley, told CNBC in 2025 that “most reviews written by AI are not fake,” calling that “one of the myths I love to bust.” Plenty of real travellers run their own honest review through ChatGPT to tidy up the grammar before posting. That review reflects a genuine stay. A detection tool will still flag it as AI-written, because it is, technically, AI-polished. What it isn’t is fake. The distinction TripAdvisor draws, and the one the UK’s regulator now writes into law, is between AI-assisted writing and a review that never corresponds to a real stay at all. An academic study published in Knowledge-Based Systems in 2025 adds a further wrinkle: simple paraphrasing of an AI-generated review can knock detection accuracy down significantly, meaning a moderately careful fraudster can slip past most consumer-facing detectors anyway. Use the tool, but don’t let a clean result talk you out of a suspicious timing pattern, and don’t let a flagged result alone convince you a review is fraudulent.
What UK law changed under your feet in 2025
Since 6 April 2025, submitting or commissioning a fake consumer review has been an automatically unfair and illegal practice in the UK, under Schedule 20 of the Digital Markets, Competition and Consumers Act 2024. The law defines a fake review plainly: one that “purports to be, but is not, based on a person’s genuine experience.” Platforms and hotels that publish reviews now carry a positive legal duty to take reasonable steps to detect and remove them, spelled out in the CMA’s fake reviews guidance. Breach it, and the Competition and Markets Authority can fine a business up to 10% of global turnover.
The CMA isn’t just holding the rulebook and looking the other way. By March 2026 it had opened investigations into five more businesses over fake and misleading reviews, spanning funeral services, food delivery, and car sales, bringing the total under active review to 14. Fake reviews and hidden fees together were estimated to cost UK consumers £2.2 billion a year before the ban. None of the current CMA cases are Bristol hospitality businesses specifically, at least not yet, but the legal exposure for any hotel or attraction that buys reviews, or simply fails to monitor for suspicious clusters, is real and now enforceable in a way it wasn’t when Which? was raising the alarm back in 2019.
My three-minute routine before I book anything in Bristol
I sort by most recent first, not by “most helpful,” because helpfulness rankings can themselves be gamed by exactly the kind of coordinated posting this article is about. Then I scan the star spread rather than the overall score. A property sitting at 4.6 built from a normal mix of 3s, 4s, and 5s reads completely differently to me than a 4.6 built almost entirely from 5s with a scattering of 1s and nothing in between.
Then I go looking for the three-star reviews specifically. Priyanka Bitra, quoted in the Seattle Times piece, put it better than I could: “I ignore 5-star and 1-star reviews. The 3-star review is usually written by a rational human being who actually stayed there.” That’s become my rule too. A three-star review has no obvious incentive to lie in either direction, which makes it the review most likely to mention the thing that actually matters to you, whether that’s a noisy ice machine down the corridor or a slow check-in on a Friday night.
So which review would you actually trust more before booking a room in the Old City: the fifth nearly identical five-star review this week praising “exceptionally attentive” staff, or the single three-star grumble about a lift that’s been broken since Tuesday? I know which one I’d read twice.
Frequently asked questions
What percentage of hotel reviews are AI-generated right now?
Estimates from 2025 research put the figure at roughly 10.7% across TripAdvisor overall, rising above 20% for luxury hotel listings specifically, according to Originality.ai and Seattle Times reporting on the same data trends.
Can I trust a platform’s fake review filter completely?
No. TripAdvisor auto-rejects around 7% of submissions and flags another 5% for human review, which means the majority of what’s posted is never independently checked before it goes live, and its own transparency report shows hundreds of thousands still slip through each year.
Is it illegal for a Bristol hotel to buy fake reviews?
Yes, since 6 April 2025, under Schedule 20 of the Digital Markets, Competition and Consumers Act 2024. Businesses that submit, commission, or fail to reasonably police fake reviews can face CMA fines of up to 10% of global turnover.
Do AI detection tools always work on short review text?
Not reliably. Most detectors were built and trained on long-form essays, so a two or three sentence review gives the model very little to analyse, and academic testing shows simple paraphrasing can reduce detection accuracy substantially.
What’s the single fastest way to spot a fake AI review?
Check the posting timestamps. Genuine reviews trickle in over weeks matching normal guest turnover, while fake campaigns tend to arrive in tight bursts, often from accounts with no review history anywhere else.
Should I ignore five-star reviews altogether?
No, but weight them less heavily on their own. Read the three-star reviews first, since they carry the least incentive to exaggerate in either direction and often contain the most specific, useful detail.
How this article was put together
This piece drew on TripAdvisor’s 2025 transparency report as reported by CNBC, Originality.ai’s published TripAdvisor analysis, Seattle Times reporting from September 2026, the peer-reviewed MAiDE-UP dataset study (NAACL Findings 2025), a 2025 machine-learning paper in Information Technology & Tourism, the UK government’s CMA guidance on the Digital Markets, Competition and Consumers Act 2024, and contemporaneous Bristol Live and Which? reporting on real, named review disputes. All figures were checked against their original source between 15 and 22 September 2026. Where a statistic came from a vendor with a commercial interest in AI detection, such as Originality.ai, that’s noted in the text rather than presented as neutral. Review detection technology and UK enforcement activity both move quickly, so the legal and platform-specific figures here are worth rechecking after mid-2027.