Hotel reviews exist because brochures omit the raw chicken, dry taps, and staff who made guests feel unsafe. Tripadvisor then placed an AI-generated summary above those reviews and, according to a Which? investigation published in July 2026, managed to put some of the brochure voice back in.

The system compressed severe complaints into language so mild that a traveler scanning the top of the page could receive a substantially different account from someone reading the reviews below. Information about illness, hygiene, water failures, and alleged harassment did not disappear from Tripadvisor entirely. It became harder to see in the part of the page specifically designed to save people from reading everything.

What the summaries left out

Which? examined Tripadvisor's AI summaries for hotels with serious negative reviews. The clearest example was the Riu Palace Santa Maria in Cape Verde. Tripadvisor's summary described spacious rooms, diverse restaurants earning "rave reviews," and "spotless" cleanliness.

The underlying reviews were rather less soothing. Guests reported raw chicken, flies and birds around the buffet, illness affecting whole families, and what one reviewer described as dead roasted mice near a sitting area. The hotel chain was also facing a High Court claim brought by hundreds of guests alleging illnesses linked to hygiene and food-safety failures. Those are allegations being litigated, not findings this site can declare proven. They are still material context for anyone being told that cleanliness was spotless.

RIU Hotels & Resorts told The Guardian that it operated with high professional and service standards and treated hygienic and sanitary safety as a top priority. The AI summary for that property was no longer available by the time of the reporting.

Another summary praised a Dominican Republic hotel for its abundant amenities, mentioning only inconsistent cleanliness and maintenance issues. Reviews beneath it described taps running dry, guests showering with bottled water, and widespread sickness in a large wedding party. A summary cannot include every complaint, but "maintenance issues" is doing heroic work when guests say basic water service failed.

A safety warning became friendly service

The Turkish example exposed the same failure with higher stakes. Guests wrote that repeated sexual harassment by male hotel staff made them feel unsafe, including requests to connect through social media. Tripadvisor's AI praised the hotel's "friendly service" and reduced the negative reports to service "lapses noted by a few."

The compression changed the practical meaning of the source material. Summarization always decides what survives, what gets grouped together, and how strongly it is described. Turning reports of feeling unsafe into a minor service lapse told the hurried reader how seriously to take facts that remained somewhere below.

Tripadvisor said its systems automatically suppress AI summaries when travelers warn about severe safety incidents such as death, drugging, or sexual assault. That makes the Turkish result particularly awkward. Either the reports did not trigger the suppression rule, the system failed to identify them correctly, or the process had a gap between what the company intended to catch and what the page actually displayed. The public sources do not establish which technical path failed. They establish that the intended protection did not prevent this output.

Summarization is an editorial decision

AI summaries are often presented as mechanical convenience: millions of reviews go in, a digestible paragraph comes out, and nobody has to scroll through forty complaints about breakfast. In practice, a generated summary is an editorial layer. It chooses representative details and assigns tone. Calling a property "spotless" is not data compression; it is a conclusion.

This matters because review platforms sell access to accumulated human experience. The useful part is not the average adjective. It is the detail that changes a decision. A traveler may tolerate dated furniture or a slow check-in. Reports of unsafe staff behavior, contaminated food, or unreliable water belong in a different category. Blending them into a warm average defeats the reason people consult reviews before handing over money and arriving in another country.

University College London professor Duncan Brumby told The Guardian that AI tends to sanitize sharper criticism, possibly because bland observations dominate its training material. His description fits the output Which? found: shortening the reviews also rubbed off the edges that made the warnings useful.

Tripadvisor's safeguards did not catch these examples

Tripadvisor said it was monitoring and refining the tool and investigating examples where reviews did not match the intended property. The company also said it remained confident that the features were doing what they were designed to do: helping travelers understand the breadth of feedback while leaving the original reviews available.

Leaving the reviews available is necessary, but it does not excuse a misleading summary above them. The summary exists because Tripadvisor expects many people not to read a billion contributions individually. Telling users they can scroll down and verify the AI shifts the quality-control burden back onto the people the feature was supposed to help.

There is no documented case in these sources of a traveler booking one of the hotels because of the summary and then being harmed. That missing evidence limits what can be claimed about customer consequences. It does not change the observed product failure. The generated text described real, deployed pages in ways that softened information directly relevant to safety and hygiene.

Reviews only help if the summary preserves the warning

Tripadvisor had the difficult part already: people had written detailed accounts of what happened. The platform did not need an AI system to discover hidden hazards or infer them from thin evidence. It needed the system not to launder blunt warnings into pleasant copy.

A useful review summary should preserve severity even when a complaint is uncommon. It can say that most guests praised a hotel while a smaller number reported serious safety concerns. That is more informative than averaging every opinion into one agreeable paragraph. Frequency matters, but so does consequence.

If a generated summary sits between customers and safety information, test whether it preserves the warnings people most need to see. Do not measure quality only by whether the paragraph sounds balanced or resembles the average review. A summary that leaves the danger buried at the bottom has not really summarized the evidence. It has curated it.