Key Takeaways
- An investigation by UK consumer group Which? found that Tripadvisor's AI review summaries glossed over serious complaints such as food poisoning and harassment, describing a hotel facing group litigation as 'spotless'
- Reviews are turning into commerce data that AI reads on shoppers' behalf, and the case shows how a single summarization design choice, treating every review equally regardless of rating, directly shapes the outcome of AI-mediated discovery, comparison, and transactions
- Any business that lets AI summarize reviews or UGC needs to build the detection and weighting of severe negative signals into the design, and to monitor how external AI describes its products
What the Which? Investigation Found
AI-generated overview found to gloss over allegations of sexual harassment and describes hotel being sued over hygiene as 'spotless'
www.theguardian.comOn July 2, 2026, UK consumer group Which? published the results of an investigation into the AI review summaries displayed on Tripadvisor hotel pages. The summaries treated serious complaints in guest reviews lightly, and in some cases portrayed hotels facing group litigation in a favorable light. This article reads the episode not as a quality lapse at one travel site, but as a design question that now confronts every merchant: how do you let AI handle review data, which has quietly become commerce data.
The most striking example is the Riu Palace Santa Maria, a five-star resort in Cape Verde. The AI summary introduced the hotel as 'popular with many travellers', described its cleanliness as 'spotless', and said its restaurants earned 'rave reviews'. Actual reviews told a different story: guests reported being served raw chicken and seeing flies and birds in the buffet. According to Which?'s tally, there were 102 mentions of food poisoning as of March 2026, and of the 32 one-to-two-star reviews posted between December 2025 and April 2026, 14 described members of the party falling seriously ill. The hotel is also the subject of a group legal action in the UK brought by at least 412 guests who say they fell ill.
The problems go beyond hygiene. At a hotel in Antalya, Turkey, multiple reviews said guests felt unsafe due to harassment from male staff, yet the AI summary described the service as 'friendly' and only mentioned 'lapses noted by a few'. At a hotel in the Dominican Republic, guests reported showering with bottled water when the mains ran dry, and half of a 68-person wedding party falling ill, while the summary softened all of this into 'inconsistent' cleanliness.
What should not be missed is that different AIs reading the same reviews reached opposite verdicts. For London's Britannia International, Google's AI summary reported that the property is 'frequently rated as one of the worst hotel chains in the UK', while Tripadvisor's AI claimed guests 'often praise the clean rooms' and described a 'charming' atmosphere. And when Which? asked the company's AI chatbot Ollie about the risk of food poisoning at the Riu Palace, it reportedly answered that it was 'quite unlikely' and that the resort had a 'strong reputation for high hygiene standards'.
Tripadvisor's Rebuttal Reveals How the Summaries Are Built
Tripadvisor responded in strong terms. The company said it 'fundamentally disagrees' with the premise of the investigation, explaining that AI summaries are snapshots based on high volumes of user-generated content and are not intended to replace individual reviews. Because users can click through from each element of a summary to the underlying traveller quotes, there is, in the company's words, no need to 'blindly trust AI-generated content'. On pages carrying traveller warnings about serious safety incidents such as death, drugging or sexual assault, it says a safeguard automatically suppresses the AI summary, and it insisted that 'no review content has been suppressed or hidden by the introduction of these tools'. The company described the chatbot Ollie as a product in development and said it would examine the flagged examples. RIU Hotels & Resorts, for its part, responded that it operates to international standards certified by external specialist firms.
The most useful outcome of the exchange is that the mechanics of the summaries became concrete. By the company's own account, the system uses large language models and natural language processing to read recent reviews, identify the most common themes, and turn them into short plain-English overviews. Each update draws on the previous 12 months of reviews, refreshes monthly, and treats reviews equally regardless of rating, highlighting what reviewers mention most often. Under that structure, a serious allegation buried in a crowd of favorable comments gets pushed out of the summary the moment it loses on frequency.
This equal-weight, frequency-first aggregation is a natural engineering choice. But even 102 mentions of food poisoning can be a minority against the full base of reviews posted over 12 months. Whether safety-related information should be handled by severity rather than frequency is a value judgment, and it either gets built into the summarization pipeline or it does not. The gap Which? documented is better explained as the consequence of that aggregation design than as a raw capability failure of the AI.
Why AI Sands Down Criticism
At the same time, the episode reflects a tendency of generative AI that goes beyond any one company's design. Duncan Brumby, a professor of human-computer interaction at University College London, told the Guardian that he has observed the same phenomenon in his research into academics using AI during peer review.
Here you have guests describing a really negative experience, but the AI has decided to tone it down. It's as if it's being polite.
Because the bulk of training data consists of bland, measured writing, Brumby argues, AI tends to sanitize sharp criticism and rub off its edges. The tendency of opinion summarization systems to reduce the richness of consumer feedback to shallow sentiment has also been documented in ACL research. And in January 2026, Google removed some of its AI health summaries over misleading information, showing that summary quality is a cross-platform problem. Left to its defaults, AI will round off criticism. The weighting and safeguards that stop it from doing so exist only if a business deliberately designs them.
Reviews Are Shifting from Reading Material to Purchase Data AI Consumes
The reason this investigation carries weight is that the way reviews get read has changed. Consumers increasingly compare and narrow down options starting from AI chat and AI summaries, for hotels and physical goods alike. The first impression an AI presents already shapes purchase decisions. Reviews are shifting from something humans read at length into machine-readable data that AI consumes as input for buying decisions.
In agentic commerce, the stage where AI agents run comparison and purchase on the user's behalf, the problem sharpens further. In a world where an agent selects the hotel or the product and executes payment, the last line of defense disappears: no human scrolls down to the one-star reviews. If the summaries and review data an agent consumes are biased, that bias transfers directly into the outcome of the transaction. Rory Boland, editor of Which? Travel, called the failure to surface critical safety information 'unacceptable and potentially life-threatening', a judgment that resonates most strongly in exactly this context.
A second lesson lies in the fact that Google and Tripadvisor described the same hotel in opposite terms. In AI-mediated discovery, it is not only your own summaries that matter; external AI reinterprets the same body of reviews on its own. No merchant controls the single lens through which its products or venues get described anymore. As we argued in our piece on brand reputation risk in AI answers, knowing how you appear across multiple AI channels is the starting point of reputation management in this era.
For the platform itself, the episode is a threat to its core asset. Tripadvisor's competitive position rests on an accumulation of more than a billion reviews and contributions, pre-purchase data at scale. If trust in the summaries erodes, the value of that asset erodes with it. With analysts arguing that closing the trust gap is the precondition for agent-mediated transactions to spread, the reliability of review data is emerging as a competitive axis for OTAs (online travel agencies) and marketplaces alike.
What Retail and E-commerce Operators Should Take Away
First, any business that runs AI summarization over reviews or UGC (user-generated content) on its own site needs to build the detection and weighting of negative information into the design. Equal-weight frequency aggregation is easy to implement, but it structurally buries low-frequency, high-severity signals such as health harm or harassment. Tripadvisor's automatic suppression of summaries for serious incidents had its limits exposed this time, but the underlying idea of holding safeguards in a separate layer is worth borrowing.
Second, the machine-readability of review data shapes how you appear on external AI channels. Bazaarvoice now offers a review discovery API for AI agents, and Trustpilot is positioning itself as a supplier of trust data for the agent era; the infrastructure for handing structured reviews to AI is being commercialized. Preserving raw signals, rating distributions and severe complaints included, in a form agents can reference, rather than only the rounded-off prose of a summary, belongs on the same list as product data readiness.
Third, the more traffic arrives via AI, the more summary errors rebound as a brand trust problem rather than isolated mishaps. This time the design under scrutiny was Tripadvisor's, but if the review summaries on your own storefront behave the same way, the criticism lands on your brand. Alongside that, regularly checking whether external AI describes your products inaccurately, and having a procedure to request corrections, deserves a place in ongoing operations.
Conclusion
The Which? investigation documented concrete cases in which Tripadvisor's AI review summaries obscured serious complaints ranging from food poisoning to harassment. The exchange that followed also surfaced the design behind the summaries: a 12-month window, monthly refreshes, and no weighting by rating. The gap was driven less by a shortfall in AI capability than by an aggregation design decision about how to treat signals that are rare but severe.
As reviews shift from human reading material to purchase data that AI consumes, and agents move toward executing purchases end to end, summarization design is becoming part of the trust infrastructure of commerce. How you let AI summarize your own UGC, and how you track the way external AI describes you, are the two design choices that will separate reputations and revenues in the AI commerce era.
