A hotel where guests report being served raw chicken, spotting dead mice by the seating area, and getting hospitalized with food poisoning. Tripadvisor’s AI summary calls it ‘spotless.’ That gap between the machine-written blurb at the top of the page and the human reviews below it is the story, according to a Hacker News investigation sourced from UK consumer group Which?. And it’s a warning for anyone trusting AI to summarize high-stakes information.
The stakes here aren’t abstract. They’re life and death.
What the AI hid
Which? examined the AI summaries Tripadvisor now places at the top of hotel pages. At the five-star Riu Palace Santa Maria in Cape Verde, the AI described ‘spacious rooms,’ restaurants earning ‘rave reviews,’ and cleanliness that was ‘spotless.’
Scroll down to the actual reviews and the picture flips. Guests reported ‘exceptionally poor hygiene,’ food that was ‘awful, bland, unsafe and inedible,’ and one family that fell ill wrote the place ‘has the potential to take lives.’ When Which? checked in March, there were 102 mentions of food poisoning. The hotel now faces a group legal action from at least 412 holidaymakers, with seven deaths reported since 2023.
Tripadvisor’s chatbot Ollie went further. Asked directly about food poisoning risk, it said illness was ‘quite unlikely’ and praised the resort’s ‘strong reputation for high hygiene standards.’
The pattern repeated elsewhere:
- Garza Blanca, Cancun: guests reported illness, including a wedding party. AI called it ‘immaculate cleanliness.’
- Occidental Caribe, Dominican Republic: reviews cited sewage smells, mould, and no running water. AI mentioned only ‘inconsistent’ cleanliness.
- Kaia Coracesium, Turkey: reviewers described repeated sexual harassment by male staff, including following guests’ daughters. The AI called service ‘friendly,’ with ‘lapses noted by a few.’
That last one is the tell. Describing sexual harassment as a service ‘lapse’ isn’t a data gap. It’s a system that found the complaint and then sanded it down.
Why this matters now
AI summaries are being bolted onto everything: search results, product pages, medical portals, travel sites. The pitch is always convenience, turning messy human input into something ‘easy to digest,’ as Tripadvisor put it. The problem is that digestion strips out the outliers, and sometimes the outlier is the one fact you needed.
A summary optimized for ‘most common themes’ will bury a minority signal by design. But a handful of food-poisoning deaths is exactly the kind of low-frequency, high-severity signal a averaging model flattens. The math that makes summaries readable is the same math that makes them dangerous.
Here’s the sharpest detail: Google’s AI overview for the same Riu Palace warned of ‘potential for illness’ and flagged ‘outbreaks.’ Same reviews, different result. This isn’t an unsolvable AI limitation. It’s a design and priority choice, and Tripadvisor chose differently.
Tripadvisor told Which? it prioritizes ‘transparency and impartiality’ and that summaries ‘surface a range of both positive and negative community feedback.’ It called Ollie ‘a product in development’ and said it’s now investigating the flagged examples.
Takeaways for practitioners
If you’re shipping AI summarization, the lesson is concrete:
- Severity should override frequency. Safety, health, and harassment signals need a separate path that can’t be averaged out. Weighting every review ‘equally regardless of rating’ is how a death report gets outvoted by breakfast praise.
- Don’t launch summaries as glossy defaults. A ‘product in development’ shouldn’t be the first thing millions of users see presented as fact.
- Benchmark against a competitor before you ship. Google’s version proves the safer output was achievable. If a rival’s summary flags a risk yours misses, that’s a bug, not a style difference.
- As a user, treat AI summaries as a headline, not the article. Scroll to the primary source, especially when money, safety, or health is on the line.
The convenience layer is only as trustworthy as what it chooses to leave out. Right now, some of these systems are leaving out the part that could ruin, or end, a trip. More details are available at the original source.