Guidera, the iPhone tour guide I build, doesn't let an AI narrate from memory. Before it tells a story about a place, it researches that place from Wikipedia, Wikidata, Wikivoyage, OpenStreetMap and official city websites, and a verifier then removes anything those sources don't support. The rule is "no evidence, no claim", so for little-known places the guide says less instead of making things up.
Why do AI chatbots make up history about places?
A language model writes the most plausible next sentence, not the most verified one. For a famous landmark there is a lot of written material, so the plausible sentence is often also true. For a small chapel, a side street or a local monument there is very little, but the model can still produce a confident, fluent paragraph about it. The result sounds exactly as convincing as the true stories.
For a tour guide, that is the worst kind of mistake. You are standing in front of the building and have no easy way to tell which parts are real. That is the problem Guidera's design starts from.
How does Guidera research a place before telling its story?
When you tap a place, Guidera first looks it up in public sources: Wikipedia, Wikidata, Wikivoyage, OpenStreetMap and official city websites. What it finds is collected into a set of short factual claims, and each claim is linked to the sources that support it. The research also records gaps: questions the sources don't answer.
For Big Ben in London, for example, the research holds nine claims from Wikipedia and Wikivoyage, from the origin of the name to the weight of the Great Bell, and one gap: the sources don't explain how the clock mechanism keeps time. Nearby places come from map data, with their distances.
What happens between the research and the story?
- Research. The place is looked up in the sources above.
- Claims with sources. The findings become short claims, each tied to its sources, plus a list of gaps.
- Narration. The story is written in the guide you chose (Arthur the Historian, Max the Local Insider or Luna the Artist), in your tone and length.
- Verification. A verifier checks the story against the claims and removes anything the sources don't support.
- The checked story. The result is what you read or hear.
What does "no evidence, no claim" look like in a real story?
Here are four sentences from a real Guidera story about Big Ben, each with the source behind it. The full story, with every sentence traced, is on the Guidera home page.
| Sentence in the story | Based on |
|---|---|
| “The name Big Ben actually belongs only to the Great Bell swinging in the dark of the belfry, not the stone tower itself.” | Wikipedia, Wikivoyage |
| “The tower was officially called the Clock Tower until 2012, when it was renamed the Elizabeth Tower for Queen Elizabeth II's Diamond Jubilee.” | Wikipedia, Wikivoyage |
| “Completed in 1859, the structure was designed in the Perpendicular Gothic style by Augustus Pugin and Charles Barry.” | Wikipedia |
| “Across the street, about one hundred and eighty metres to the south-west, lies St. Margaret's Church.” | map data |
Sources: Wikipedia: Big Ben, Wikivoyage: London/Westminster
Just as important is what the story leaves out. The sources had no details on how the clock mechanism works, so the story doesn't mention it at all, even though a chatbot could easily have produced a paragraph about it.
What does Guidera do with little-known places?
It says less. Guidera works on any place Apple Maps knows, including places with only a line or two of public information. When the sources are thin, the story is shorter and sticks to what is supported. A short, true story is the intended result, not a failure.
Why do two people get the same story for the same place?
Once a place's story has been researched and checked, it is stored and shared. The next visitor who taps the same place gets the same checked story instantly, instead of a new story being written each time.
What data does Guidera send when you tap a place?
To create a story, Guidera sends only the coordinates of the place you tapped, never your own location. There are no user accounts. The full details are in the privacy policy.
What are the limits of checking stories against sources?
Checking against sources reduces made-up facts, but it can't fix a source that is wrong, thin or inconsistent. A real example: while researching our Munich walking guide, I found that Guidera's research notes and the English Wikipedia article on the Hofbräuhaus name different dukes as the founder of the brewery in 1589. On the blog, I only state the year that both agree on.
Stories can also be brief or incomplete, especially for small places. That is why the app shows this note with each story: "AI-generated content may not be accurate. Always verify important information."
Frequently asked questions
Does Guidera make things up?+
Guidera is built to avoid it: every story is based on researched claims with sources, and a verifier removes anything the sources don't support. Sources themselves can still be wrong or incomplete, so the app asks you to verify important information.
Which sources does Guidera use?+
Wikipedia, Wikidata, Wikivoyage, OpenStreetMap and official city websites.
Why is the story for some places so short?+
Because there is little reliable information about them. For little-known places Guidera says less rather than filling the gap with invented details.
Does Guidera know where I am?+
No. To create a story, Guidera sends only the coordinates of the place you tapped, never your own location, and there are no user accounts.