How AI recreations of historical figures work

The engine: a large language model

At the core is a large language model: a program trained on a very large collection of text to predict which words come next. Trained this way, it learns grammar, facts, styles of writing and patterns of argument. It has read a great deal about famous people, including their own works where those are public, and also a great deal of what others have said about them, myths and misquotations included.

On its own, such a model can imitate a famous person, but loosely. It mixes eras, reaches for well-known lines whether or not the person said them, and fills gaps with plausible inventions. The rest of the system exists to narrow that down.

The research file on each figure

A careful recreation starts with a written record of the person, prepared by people, not generated on the fly. In Yuga, each figure's file holds a summary of the life with dates, their documented opinions with a source for each, sample lines from their own writing with sources, the people in their life, notes on how contemporaries described their voice and manner, and sensitivity notes on subjects that need care, such as slavery, war or colonialism.

For Abraham Lincoln, for example, the file records that he wrote to Albert G. Hodges on 4 April 1864 that if slavery is not wrong, nothing is wrong, and also that in his debate at Charleston, Illinois, on 18 September 1858 he said he did not favour social and political equality of the races. Both are documented, and a recreation that only gave the first would be misleading.

The persona instructions

Before each conversation the model receives instructions built from that file. In plain terms they say:

• Who you are, and the year and place the visitor has chosen.
• How you speak: vocabulary, pace, habits of speech, temperament.
• What you believe, with the documented positions and their sources.
• The edge of your time: you know nothing after the chosen year, including the rest of your own life.
• Honesty about the record: hold the views you really held, even uncomfortable ones; quote only from the passages supplied; say so when you are reasoning beyond the record.
• If someone sincerely asks if they are talking to a real person, say plainly that you are an AI recreation.

The figure's own writing, retrieved as you talk

Models do not memorise texts reliably, so a good system does not ask the model to quote from memory. When you ask a question, the system searches the figure's writings (letters, speeches, books in the public domain) for relevant passages and gives them to the model alongside your question. That is how a recreation can quote accurately: it is reading, not remembering.

This is also why figures who left a lot of writing are recreated more faithfully than those who left little. Socrates wrote nothing; every word comes through Plato, Xenophon and others. Harriet Tubman could not write, and most of her recorded words were written down by others, especially Sarah Bradford, who shaped them. William Shakespeare left almost no statements of personal opinion outside his plays and poems, and a character's speech is not his belief.

How Yuga marks what is quoted and what is inferred

Because a recreation mixes the record with reasoning, it should show which is which. In a Yuga conversation, a reply drawn from the figure's writing carries a label, From their own writing, which opens the passage it used, with the title of the work, its licence and a link to the source. A reply where the AI reasons beyond the record is labelled In their spirit, not their recorded words, with a short explanation that it is not quoted. On contested subjects, a historian's context card can follow the reply.

Labels help, but they do not make the system infallible. A passage can be applied to the wrong question, and reasoning in a figure's spirit is still a guess. Check anything important against the sources.

Why the voices are designed, not cloned

For most people in history no recording exists. The first machine that could record and play back sound, Edison's phonograph, dates from 1877. So no one knows exactly how Lincoln, Shakespeare or Socrates sounded. A voice for them has to be designed, from what contemporaries wrote about their speech: accent, pitch, pace, age.

A few figures did live into the recording age. Florence Nightingale recorded a short message in 1890, Leo Tolstoy made recordings in his last years, and Thomas Edison recorded himself. Even for them, Yuga's voices are designed to suit the person rather than copied from the recordings. Old recordings are short and rough, a recording can have its own copyright separate from the speaker, and copying a real person's voice to say words they never said is a line Yuga does not cross. The app says on each figure's page whether recordings survive.

What AI figures get right, and what they get wrong

There are deeper limits too. The record favours people who could write and whose papers were kept, so the voices of enslaved people, women and the poor are thinner. A figure who spoke Greek or Bengali speaking to you in English is already a translation. And any account of what a person would have thought about something new is a guess, however well informed.

How to use one well

• Read about the figure first, so you notice when something sounds wrong.
• Ask about things they wrote about, and ask for the source.
• Try a question with a false premise; a good recreation corrects it.
• Treat every answer as a lead. Find it in a primary source before you use it.
• Cite the document, never the chatbot.

Yuga is free to start (15 voice minutes, then 5 a week, plus text every day), in 24 languages, on iPhone and Android, for ages 13 and up. Its recreations are built from the figures' own writings and the historical record, and like every AI they can get things wrong. For classroom use, see AI historical figures in the classroom; for students, using AI to study history.