Family trees and AI: when historical evidence becomes fiction

For EditieNL, I explored reporter Eise Pulles's family tree with AI, and discovered why a convincing answer is not the same as historical evidence.

For an EditieNL segment, I used AI together with reporter Eise Pulles to investigate his family tree. It was a fascinating and successful experiment, but it immediately exposed one crucial lesson: generative artificial intelligence only produces truly valuable results when you understand exactly how to direct the model.

Ask a lighter language model to simply create a family tree and it can produce an impressive, complete-looking overview within seconds. A closer look at the archives may reveal that names, dates of birth and family relationships were fabricated. Lighter AI models are susceptible to sycophancy, a kind of digital people-pleasing. They want to give the user a complete answer, even when that means filling gaps in history with invented facts.

Generative AI as an analytical engine

How can AI support reliable genealogical research? Generative AI is not a magic box that produces truth from nothing. It is, however, very good at rapidly collecting, structuring and interpreting large amounts of complex information.

AI can search large sources and recognise patterns, provided it operates within a strict framework. This analytical work requires a deliberate combination of the chosen model, its instructions and the user.

1. Choose a capable reasoning model

For complex investigation, model choice matters. A capable reasoning model has more capacity to keep many sources in view and compare them systematically. It can, for example, check whether a date of birth is consistent with the age stated on a death certificate. That accuracy is essential: one incorrect link caused by a misinterpreted family name can undermine an entire family tree.

2. Give the model strict instructions

A capable model alone is not enough. Require the AI to use only verifiable, certified historical archive sources and to report uncertainty explicitly. This prevents it from closing gaps with assumptions. Instead of asking for a polished story, ask the AI to behave like a critical archivist who draws a conclusion only when the evidence allows it.

3. Keep the user in control

The user must retain control. Good prompting is the bridge between an unhelpful generation and a high-quality analysis. Define clear conditions in advance, require objective and traceable information, and tell the model to flag missing documents or the risk of confusing two people with the same name.

AI only becomes a reliable research assistant when source verification, uncertainty and human control are part of the assignment.

From storyteller to research assistant

Apply these three principles consistently and AI changes from an unreliable storyteller into a powerful assistant for historical research. AI's speed is valuable, but only when every conclusion can be traced back to real evidence.