Hallucinated Events

Timeline Maker AI, the Preceden tool that generates timelines for you, sometimes makes up events that never happened. Those inventions are called hallucinations, and every large language model (LLM) has this problem. Eliminating them is an active area of research.

For some kinds of timelines, like project plans, invented events are not a problem. They are the point, because the plan is meant to be fictional.

On a timeline about real-world historical events, though, it matters: you can come away believing something happened that never did.

What a hallucinated timeline looks like

Timeline Maker AI works out for itself how many events a topic needs, and it usually (but not always) does a pretty good job on well-documented subjects. Obscure topics are where made-up events show up most, because the model has less to work from and no way to signal that it is guessing.

For example, here is a timeline generated for "History of Preceden Timeline Maker":

Preceden was conceived in 2009, not 2011, and launched in 2010, not 2012. Every event on that timeline is invented. If you didn't know much about Preceden, or if only some of the events were made up, you could easily take the whole thing as accurate.

Reducing hallucinations

The narrower and better documented your topic is, the less room the model has to invent. "History of the Roman Empire" will come back in far better shape than the history of a small company or a niche local event.

Whatever you generate, check the events against a reliable source before you rely on them, particularly when the exact dates matter. Timeline Maker AI is best treated as a fast first draft: once the timeline is open in the editor you can correct any date, rewrite any event name, and delete anything that never happened.