Tripadvisor's AI Summaries Called a Hotel "Spotless." 102 Guests Reported Food Poisoning.

So here's what actually happened. A consumer group in the UK called Which? dug into Tripadvisor's AI-generated review summaries... the ones that sit at the top of a hotel's page and give you the "quick take" so you don't have to read 200 individual reviews. They found a resort where 102 guests mentioned food poisoning. Thirty-two one- and two-star reviews between December 2025 and April 2026, fourteen of which described serious illness. Seven deaths reported among guests since 2023. Over 400 people are part of a group legal action. The AI summary? "Spotless."

Let that land for a second. Not "mixed reviews about food safety." Not "some guests reported illness." Spotless.

And it gets worse. Another property had multiple reviews mentioning sexual harassment by staff. The AI summary described the service as "friendly." This isn't a quirky bug. This is a fundamental architectural problem with how large language models handle sentiment. A professor at University College London nailed it... AI trained on massive text datasets tends to "sanitise and rub off the edges" of negative content. The model averages everything. It rounds toward pleasant. Which is fine if you're summarizing restaurant reviews about slow service. It is genuinely dangerous when the negative reviews describe people getting sick and dying. Tripadvisor says their systems "automatically suppress summaries for serious safety incidents." Clearly, 102 mentions of food poisoning and seven deaths didn't meet that threshold. That should tell you everything about how well those systems actually work.

Here's the part that matters for operators. This cuts both ways, and neither direction is good. If your property has a real problem... a mold issue, a pest problem, a safety concern you're working to fix... the AI might be papering over it in ways that bring more guests into a situation you haven't resolved yet. That's liability you didn't ask for. But the other side is just as bad. If you're a property that FIXED a problem... spent real money, retrained staff, replaced equipment... the AI summary is still averaging in those old one-star reviews. The 150-word summary at the top of your page doesn't know you replaced the kitchen hood six months ago. It doesn't know you fired the sous chef. It's still averaging the sentiment from reviews written before the fix. Your $80,000 renovation just got erased by an algorithm that treats a review from 2024 the same as one from last week.

Look, I've been watching AI get bolted onto hospitality platforms for years now, and the pattern is always the same. The vendor builds the tool to optimize engagement (Tripadvisor has said users interacting with their AI tools generate 2-3x more revenue), ships it fast because the competitive pressure is real (Google's AI Overviews are eating Tripadvisor's organic traffic and they know it), and the edge cases... the ones where the AI does something actively harmful... get discovered by someone outside the company, not inside it. Tripadvisor didn't catch this. A consumer advocacy group caught it. That's not a technology failure. That's a priorities failure. And by the way, AI-generated reviews on Tripadvisor increased 137% from 2019 to 2024, making up 10.7% of all reviews. So now you've got AI writing the reviews AND AI summarizing them. At what point does any of this still qualify as "user-generated content"?

The question nobody's asking is whether we should be using generative AI to summarize safety-critical information at all. Not whether the AI can be "improved" or "fine-tuned"... whether this is an appropriate use case. I wouldn't build a system that averages sentiment across reviews containing reports of death and illness. Not because I can't. Because the failure mode is someone booking a hotel room that gets them sick. Or worse. The Dale Test question here is simple: when this system fails, what's the consequence? If the answer is "someone might die," maybe don't ship it until you've solved that.