How Witness uses AI · 22 October 2026

The record writes the story. A model only puts it into sentences.

Witness uses language models to tell you about the people in your own family file, and a recorded voice to read a household aloud. This page says what those models are given, what they are kept from, where you meet them, and what to do when they get something wrong.

Three rules hold everywhere. The model writes only from documented facts, and the record outranks the story. Nothing about a living person is ever written. Nothing read from FamilySearch Family Tree ever reaches a model.

The models behind it

The writing is done by Anthropic's Claude, through its API: Claude Sonnet for most of the work and Claude Opus for the overview of a whole tree. The recorded narrator in the Hearth is a Microsoft Azure neural voice. Both are used under commercial terms that forbid training on what we send; Anthropic deletes API inputs and outputs on its side within thirty days. Nothing Witness sends is used to train or tune any model, by us or by them.

Witness keeps what a model wrote for your tree, so a story is not written twice, until the story is rewritten or the tree is deleted. Delete the tree and the stories go with it.

What it sees

Only the deceased people in your own file, and the public record around them.

Given

The names, dates, places, events, relationships and citations of the deceased people in your tree, from the file you brought in. The reading of a photograph you uploaded to Witness. Public sources for the world around them: Wikidata's dated events, and newspaper pages from the Library of Congress.

Kept back

Anyone who appears to be living: never named, never described, never sent. Any other reader's tree. The notes you write for yourself on a person's page: the writer never sees them. Your account, your identity, your location.

And every fact read from FamilySearch Family Tree, which the next section explains.

Each piece of writing is made for one bounded subject, one person, one line of descent, or one household, never from a free question over the whole tree.

Where you meet it

Where a model wrote something, this page names it.

Their story, on a person's page, and the paragraph of the world around them beneath it. The panel says so: written by AI from the documented facts of this record; the record, not the story, is the authority.
The daily note on Home, a few sentences about one person.
A story arc: one founder's line, told across generations.
The overview of your tree on the Tree tab.
A research brief: suggested lines of enquiry for one person's gaps.
The Hearth's narration, when it is told by a named narrator in a recorded voice. The narrator is a persona, not an ancestor. The plainer telling, read by your device's own voice, is composed by code with no model in the path.
Reading a headstone you photographed, and reading a photograph you uploaded: a transcription, and a match offered for you to judge.
Obituary leads: whether a newspaper page is about your person, offered under Records to Check.

Not AI, so there is no doubt: the reading of your file, the maps, the timelines, the census households, Tree Health, the record registers, and FamilySearch's own record hints, which Witness shows with FamilySearch's rating, unranked by any model.

The rules the writer keeps

Facts, not fiction.

It is given a person's recorded facts and nothing else about them, and is told not to add an occupation, a relationship or an event that is not in the record, and never to guess at a cause behind one.
Where a note in your file is the only source for something, the story says so in plain words rather than presenting it as documented.
What happened around your ancestors is told as history, never as something they did, unless the record says they did it.
No story of any kind is written about a living person. Living relatives are counted where a count matters, and never named.
Nothing a model writes is ever sent anywhere outside Witness by Witness. You can save your own story as a text file; that is your choice and your copy.

FamilySearch Family Tree

Nothing read from Family Tree reaches a model.

A tree read from FamilySearch Family Tree is marked as such the moment it is read, and a wall in the code and in the database keeps it out of every model and the recorded voice. Each function that would call a model checks the mark first and refuses; a database rule refuses to keep any model's output for such a tree even if a function forgot. This is enforced, not merely promised.

So a tree read from Family Tree gets the record, the maps, the timelines, the census households, the Hearth rooms in the plainer telling, Tree Health, the registers, and FamilySearch's record hints. It does not get the written stories, the world around them, the daily note, the overview, the briefs, the recorded narrator, or the readings. The page says so where a story would be.

Nothing a model writes is ever sent to Family Tree: not as a Memory, a note, a source or a change to any person. The one thing Witness writes to Family Tree is a headstone photograph as a Memory, with a caption you typed, after you accept FamilySearch's content submission agreement.

Witness is a FamilySearch Compatible Solution. The disclosure behind this page is on file with FamilySearch and changes with the product.

When it gets something wrong

The record outranks the story.

Models make mistakes, and so do the records they read. Every story sits beneath the facts it was written from, so you can check it against them on the same page, and Witness's terms say what the app says on the panel: a story is a reading of the record, not a verified fact, and anything you rely on should be checked against the source. If a story says something the record does not support, or you know to be wrong, write to support@witnesslives.com with the person's name. We read every one, and we fix what you find.

Your family is yours

Never sold, never pooled, never trained on.

Your tree is yours alone. Trees do not share data with each other, nothing about your family is sold or shared for any model's training, and you can delete a tree, or your whole account, from the app at any time. What we keep and who else is involved is set out in full in the privacy policy.