Can a machine live one day of your ancestors' lives?

AncestorBench asks AI models to step into the life of a real, documented person from the past and render their world as they saw it. Historians check the facts. You decide what moves you.

Place · Date · Point of view

A real person who once lived, on a real day. What they could see, hear, smell, buy and say. What they knew, and just as important, what they could not yet know.

Attested from a source Inferred from the period Invented and marked as such

Every simulation ships with its footnotes: each concrete claim, tagged by how we know it.

How it works

In 2003 the philosopher Nick Bostrom imagined that a future civilization might run detailed simulations of its ancestors, for research and for entertainment. We are starting small, with words.

A real person

Every simulation centers on one specific individual with a paper trail: records, letters, censuses, court files. Never a composite, and never someone who died less than fifty years ago.

Their eyes only

Models write in the first person, limited to what that person could perceive and know. No hindsight, no borrowed words from a later century.

Show your sources

Alongside the story, the model lists its claims and marks each one as attested, inferred or invented, the way a historian's footnotes do.

Graded twice

Once for truth, once for delight. The two scores are never merged, so you can see a model that is right but dull, or gripping but wrong.

Two scores, kept apart

Fidelity

Was it true to the record?

AI judges audit every claim for dates, people, prices, objects, language and social norms, and a panel of historians and genealogists checks the judges. Factual errors weigh heavily; honest differences of interpretation weigh lightly.

Delight

Did it bring them back to life?

Readers compare two simulations side by side and vote, producing an Elo rating like a chess ladder. A model trained on those votes gives fast preliminary ratings, and fresh votes keep it honest.

What we believe

An ancestor simulation is only worth building if it is honest about the past and humble about what we cannot know.

The whole world, from day one. People from every region, era and walk of life, sampled branch by branch like a family tree. We publish who the set covers, and who it misses.
The past as it was, not as we wish it had been. People held the beliefs of their time and spoke in its language. We do not sanitize them, and when a model softens or refuses period-accurate material, we measure that and report it on its own.
Invention is allowed, but it must be labeled. The record is full of silences, especially for the poor, the illiterate and anyone who never came before a clerk. Filling them is fine; hiding that you filled them is not.
Built to be measured. A public feed for everyone, a hidden held-out set for serious evaluation, published harnesses, fixed sampling settings and error bars, so AI labs can track real progress.

Roadmap

We are building in the open and shipping as we go. The first benchmarks are on the way.

First

Text

First-person scenes with footnotes, short enough to read in a minute and chained into serialized lives.

Then

Voting and leaderboards

Side-by-side reading, public Elo, and the first fidelity results.

Later

Video

Point-of-view clips rendered from the same scene cards.

Someday

VR

Step inside the room. The highest fidelity, and the most headroom for AI.