expectedwrong hindsight

The Itinerary Is Fake But the Links Are Real

A travel app that uses ChatGPT to hallucinate your vacation, then links to the hallucinations.

2 min read 296 words #llm #chatgpt #product #travel #hallucination
hindsight — evolved

This was the hallucination problem before anyone called it that. The specific pattern — AI confidently generating plausible-but-fake content that gets treated as real — became the defining issue of the next two years. Grounded search and tool use mostly solved the travel case, but the underlying dynamic is still everywhere.

Someone built a travel app where ChatGPT writes your itinerary and then ChatGPT extracts the points of interest from the itinerary it just made up, and then those get turned into hyperlinks.

The pipeline is two prompts. First: what is an ideal itinerary for N days in [city]? Then, fed the output: extract the points of interest out of this text, with no additional words, separated by commas. The second prompt parses the first prompt's fiction into a structured list, which gets hyperlinked inline. The whole thing runs on text the model invented.

Nobody verified anything. The restaurant might be closed. The museum might not exist. The hyperlinks go somewhere, probably, because someone else built something on top of the same hallucination at some point.

The interesting part — and this is what makes it worth thinking about — is that the approach scales. Same chain, more steps: validate the POI against something real, suggest a better one if it fails, slot in hotel accommodations, price things out. You could build a full travel planning product this way, with ChatGPT doing all the work and you doing none of the verification. The output is confident. It reads like it was written by someone who's been to Lisbon.

The cost is what caps it. Every step in the chain is an API call, and API calls in February 2023 are not free — which means this is probably better as a micro-site factory than a live consumer product. Run the chain once per destination, cache the output, publish it. Thirty cities, thirty pages, done. The hallucinations become a feature: they're consistent, they're linkable, they look like content.

It's a content mill that outsourced the lying to a language model. The language model is very good at this.