"Spotless." That was the word Tripadvisor's AI chose to summarize cleanliness at the Riu Palace Santa Maria in Cape Verde.
One scroll down the same page, a guest described "dead little roasted mice by the sitting area" on what she called her "nightmare" holiday.
Another guest was served raw chicken. Others posted photographs of flies and birds landing on the buffet food. When Which? counted in March, there were 102 mentions of food poisoning on that single listing.
The AI's reported verdict on the restaurants associated with these complaints? "Diverse restaurants" earning "rave reviews."
This is not a rounding error. The five-star resort is now the subject of a group legal action representing at least 412 holidaymakers who say they fell ill after staying there, with seven deaths reported since 2023. One guest reportedly died this year, according to Which?.
Between December 2025 and April 2026 alone, the hotel collected 32 one- and two-star reviews. Fourteen of them said at least one member of the party fell seriously ill. Many were hospitalized. Some flew home early.
None of this appeared in the reassuring paragraph that Tripadvisor placed at the very top of the page — the first thing you read, positioned precisely where your eye lands and your trust settles.
That placement is the whole story.
Tripadvisor built a tool it knows sands down safety warnings, placed it at the top of the page, and then declared that it's your job to distrust it and scroll past.
It gets more damning when you ask directly. Tripadvisor also runs an interactive trip-planning bot called Ollie. Which? asked Ollie point-blank about the risk of food poisoning at the Riu Palace.
Ollie said it was "quite unlikely," and that the resort had "a strong reputation for high hygiene standards." This was reportedly said about a hotel with seven reported deaths since 2023 and a 412-person legal action attached to it, according to Which?.
The instinct here is to blame the model — to assume the AI simply made a mistake, the way AI makes mistakes. That instinct is wrong, and understanding why is the point.
Duncan Brumby, a professor of human-computer interaction at UCL, reportedly offers the mechanical explanation, per Digital Trends: AI models are trained on data that skews bland and polite, so they tend to sand down harsh criticism. When a guest writes a negative review, the AI sometimes treats it like a minor inconvenience, averaged into the pleasant mean.
In other words, softening the alarm is not a glitch. It is what the technology does by design when you feed it a page full of human warnings. The one-star review — the one written by someone who thought they were going to die — is exactly the input the summary is built to smooth over.
Now watch how Tripadvisor defends itself, because the defense is the confession.
The company says its summaries "explicitly are not intended to replace individual reviews." It says "users can easily click to see the traveller quotes behind each review element." It says its community "understands that AI technology is still developing" and "has the common sense to check any AI advice."
Read that sequence again. Tripadvisor built a tool it knows sands down safety warnings, placed it at the top of the page so you'd see it first, and then declared that it's your job to distrust it and scroll past.
That is the accountability gap in a single move: maximum prominence, zero responsibility. They engineered the position of trust and disclaimed the consequences of it.
Rory Boland, editor of Which? Travel, named it exactly. Tripadvisor, he said, "may insist users can still fact-check its summaries against real reviews, but this ignores the fact that it made the decision to push these summaries to the very top of the page."
"This failure to surface critical safety information is unacceptable and potentially life-threatening."
Life-threatening is not hyperbole here. It is the literal claim, backed by seven deaths.
Tripadvisor insists it "fundamentally disagrees" with the investigation, and that its systems "automatically suppress AI Summaries for listings that feature warnings from travellers about serious safety incidents such as death, drugging or sexual assault." Which is a striking thing to assert about a listing where a guest died this year and the summary called the place spotless.
The pattern repeats beyond one hotel. At the Occidental Caribe in the Dominican Republic, the AI noted merely "inconsistent" cleanliness and "maintenance issues" — while reviews described no running water and sewage smells.
At a hotel on Turkey's Antalya coast, guests reported repeated sexual harassment from male staff, including a restaurant worker who followed a woman up the stairs to her room. The AI summarized the service as "friendly," with "lapses noted by a few."
Sexual harassment, rendered as a service lapse. This is the sanding-down doing its job.
Here is why you should care even if you never book a hotel on Tripadvisor: this exact pattern is arriving in your pocket, dressed as convenience.
Apple's iOS 27 is, according to a single forward-looking report on glencoehighlandholidays.com, reportedly set to ship a feature called Call Context. When you phone an airline to change a flight, that report claims it uses on-device AI to anticipate and surface the information it decides is relevant — pulling your confirmation code from Mail before you ask.
Useful, genuinely. But look at the shape of it. An AI is now deciding, on your behalf, which pieces of information mediate a high-stakes transaction. It reportedly surfaces what it judges relevant, which arguably leaves out the rest.
That is the same architecture as the review summary: a machine standing between you and the raw facts, presenting a curated version at the exact moment you're making a decision. When it works, it saves you thirty seconds. When it silently omits the wrong thing, you never know what you didn't see.
The convenience and the danger are the same feature. You cannot have the summary that saves you time without the summary that can quietly drop the detail that mattered.
So here is the protocol, and it is simple enough to use every time money or safety is on the line.
Treat any AI summary as marketing copy, not a verdict. It is the blurb on the back of the book, not the review section. It was placed first to be believed; believe it accordingly, which is to say, not at all where it counts.
Scroll straight to the one-star reviews. That is where the raw chicken, the sewage smell, the followed-up-the-stairs incidents live — the specific, checkable, unflattering detail the AI is structurally built to average away.
Cross-check on a second site. If a hotel is genuinely dangerous, the warning will exist in more than one place; if the alarm only appears once, you can weigh it, but the absence of a warning in an AI summary is not evidence of safety.
And never let a summary be the last word. Not because AI is always wrong — a recent study reportedly put Google's AI Overviews at accurate roughly 91 percent of the time — but because that remaining sliver, at internet scale, is tens of millions of answers an hour that are erroneous or at least questionable.
The Riu Palace summary has since been quietly taken down. The 102 mentions of food poisoning were there the whole time, one scroll below the word "spotless."
The information that could have saved a booking was never hidden. It was simply placed second — and someone decided second was good enough.