Tripadvisor AI summaries downplay hotel safety issues

UK consumer group Which? found that Tripadvisor's AI-generated hotel summaries and its Ollie chatbot downplay or omit guest reports of food poisoning, poor hygiene, and sexual harassment, publishing its investigation on July 2, 2026 (also reported by The Guardian, The Times, Metro, and CommsTrader). At the five-star Riu Palace Santa Maria in Cape Verde, Tripadvisor's AI described the hotel as "popular," with "spotless" cleanliness and "rave reviews" for dining, while Which? found 102 mentions of food poisoning in reviews as of March 2026 and 14 of 32 recent one- and two-star reviews describing serious illness. The hotel is now the subject of a group legal action involving more than 412 holidaymakers, with seven deaths reported since 2023. Tripadvisor's chatbot Ollie, when asked directly about food-poisoning risk at the hotel, called it "quite unlikely" and cited the resort's "strong reputation for high hygiene standards." Which? found similar gaps at Garza Blanca in Cancun, the Occidental Caribe in the Dominican Republic, and Kaia Coracesium in Turkey, where the AI summarized reports of sexual harassment by staff as "lapses noted by a few." For comparison, Google's AI overviews for the same properties surfaced the safety warnings directly, including "potential for illness" and "outbreaks of illness" at the Riu Palace. Tripadvisor told Which? it "fundamentally disagrees with the premise" of the investigation, that its summaries "are not intended to replace individual reviews," and that its systems "automatically suppress AI Summaries for listings that feature warnings from travelers about serious safety incidents such as death, drugging or sexual assault." The OECD's AI Incidents Monitor (OECD.ai) has classified the case as an active AI harm affecting consumers.

The pattern is consistent with a known failure mode in review summarization: models optimized for aggregate sentiment or majority themes tend to underweight low-frequency, high-severity signals, since standard evaluation metrics (ROUGE-style overlap, crowdworker preference scores) reward representativeness over recall on rare hazard classes. Tripadvisor said its summaries draw on the prior 12 months of reviews and are refreshed monthly, which can further dilute recent severe incidents inside a larger volume of routine feedback. Which?'s finding that Tripadvisor's system did sometimes reference minority complaints, but in softened language ("lapses noted by a few" for reported sexual harassment), suggests the gap is not pure omission but a systematic softening of risk language during abstraction.

Teams building or evaluating consumer-facing summarizers over user-generated content should measure recall on safety-labeled complaint classes (illness, injury, harassment, structural hazards) as a distinct metric from overall summary quality, and should audit whether abstractive rewriting is quietly downgrading explicit hazard language into vague descriptors like "maintenance issues" or "inconsistent cleanliness." Surfacing direct links to representative negative reviews alongside any AI summary, rather than relying on the summary alone, is a low-cost mitigation Google's comparable feature already appears to apply more consistently.

Whether Tripadvisor's stated auto-suppression safeguards for "death, drugging or sexual assault" warnings actually catch cases like Kaia Coracesium's harassment reports and the Riu Palace's illness reports going forward, and whether Which? or regulators press for independent testing of the update. Coverage has not yet established whether other travel platforms using similar AI summarization exhibit the same softening pattern; that comparison is a natural next test given the OECD.ai monitor now tracking this as a live incident.