Hallucinated Events

Timeline Maker AI, the name of the tool on Preceden that generates timelines for you, may sometimes make up (aka "hallucinate") events that did not occur in the real world. All Large Language Models (LLMs) like ChatGPT (which Timeline Maker AI uses behind the scenes) have this problem and it's an active area of research to figure out how to eliminate these hallucinations.

For certain types of timelines like project plans, this isn't a problem and is actually desired because we want the tool to generate a fictional project plan.

However, when you're using Timeline Maker AI to generate a timeline about real-world historical events, this can be a problem because you may be led to believe that certain historical events took place that did not in reality.

Reducing hallucinations

When you use Timeline Maker AI to generate a historical timeline, we recommend you leave the default Automatic event option selected in the Advanced Settings. This will tell ChatGPT to use its judgement on how many events to include and it usually (but not always) does a pretty good job of generating a timeline with events that actually took place.

If you choose to select a custom event count, Timeline Maker AI will attempt to generate a timeline for you with that number of events, and will happily hallucinate events that never actually occurred.

For example, if we try to create a timeline for "History of Preceden Timeline Maker" and request 50 events, it will create a timeline where the large % of the events never actually happened:

Preceden was conceived in 2009, not 2011, and launched in 2010, not 2012. In this case, all of the events on this timeline are hallucinated. But if you didn't know much about Preceden, or if some percentage of the events were true, you could easily believe the events are accurate.

In short, we recommend keeping the "Automatic" event count option selected when generating timelines about historical events. Otherwise, feel free to experiment with different event counts.