AI Chat lets you ask questions about the data you have gathered in plain English — “who are my closest matches on my mother's side?”, “who do Elsie and I both match?” — and get an answer worked out from your own database. It reads your data; it never changes it.
The important part is what the AI is not given. Before anything leaves your computer, every
name, surname and place is replaced with a token such as [M1234]. The AI works with the
tokens, and the Client puts the real names back on your screen. See
How Your Privacy Works for the detail.
You need three things: a Gold or Angel subscription, some gathered data, and an AI to talk to.
anthropic, openai,
gemini and others). These give the best answers. You need your own API key
from that company, pasted into the key box. You pay that company for what you use, at
its own prices; the status line shows the size of each exchange.ollama or lmstudio) running on your
own computer. Nothing leaves your machine and there is no cost, but answers are slower
and noticeably less capable.Your API key is stored on your computer, encrypted the same way the Client stores your DNA-site passwords. It is sent only to the AI company it belongs to.
When you first connect, the chat shows a list of starter questions. Click one to put it in the message box, replace anything in (parentheses), and press Enter. Where a question needs a person, type the name or pick it with @. These are the kinds of question it handles well:
A good first question. You get a table of your kits with the company and match count for each, plus what kinds of data you have — shared matches, segments, trees.
You can filter by shared cM, by company, and (for Ancestry) by maternal or paternal side.
Shared cM, number of segments, largest segment, which of your kits she matches, her side, and whether her tree is in your database.
Shared matches (also called in common with, or ICW) come back ranked by how much DNA each one shares with you. Where the testing company provides it, you also get how much each shares with the other person — Ancestry does, FTDNA does not.
Segment questions work for FTDNA, 23andMe, MyHeritage and GEDmatch kits. Ancestry does not give out segment data, so there is nothing to show for Ancestry matches.
These are where the AI earns its keep: it runs several lookups, compares the results and explains its reasoning. Follow-up questions work too — “now do the same for her brother”.
The surest way to say who you mean is to pick them from a list. In the message box, type @ and the first few letters of a name:
A list opens above the box showing the people whose names start that way, in two groups:
| Group | Who they are | Shown with |
|---|---|---|
| Match | People who took a DNA test and match one of your kits | Shared cM and company, strongest first — Elsie Marlow · 85 cM · Ancestry |
| In trees | Ancestors found in your matches' family trees | Birth year and how many trees they appear in — Josiah Hartley · b. 1820 · in 11 trees |
Use the arrow keys and Enter (or Tab, or a click) to choose. The name goes into your message — @Elsie Marlow (Ancestry) — and you carry on typing. Press Esc to close the list without choosing. You can pick more than one name in the same message.
@mar el finds Elsie
Marlow. Each part you type has to be the start of a word in the name.@hart shows every match and tree
ancestor with a name starting that way.A match is one person with one set of DNA results. Pick a match when your question is about DNA: how much you share, who you both match, which segments.
A tree person is an ancestor, and the same ancestor usually appears in several of your matches' trees — each match has their own copy. Pick a tree person when your question is about who descends from them:
The AI finds that ancestor in every tree they appear in and tells you which matches those are and where the ancestor sits in each tree. Tree people are matched on name and birth year, so the same man entered as “Josiah Hartley, b. 1820” in one tree and “J. Hartley, b. 1821” in another appears as two entries in the list. Pick the one in the most trees, then ask about the other.
You rarely remember exactly how a match's name is written, and you don't have to use the @ list. When you simply type a name, the Client finds the person you mean on your computer, before the AI is involved:
| You type | It finds | Why |
|---|---|---|
| Marlow, Elsie | Elsie Marlow | Word order doesn't matter |
| Elsie Marlowe | Elsie Marlow | A one-letter slip |
| Jim Thornbury | Mr. James Albert Thornbury | Nicknames; titles and middle names are ignored |
| Pam Ostrow | Pamela Ostrow | A short form of the first name |
| Zoe Muller | Zoë Müller | Accents are ignored |
| hartley84 | hartley84family | The start of a username |
What happens next depends on how many people fit:
get_shared_matches. Only tokens appear in them, which is your evidence of what the
AI actually received.[M1234] and click Reveal.The AI is told to quote figures from your data, and in our testing with a strong cloud model its cM values, segment positions and counts matched the database. Its interpretations — “this looks like a second cousin” — are informed suggestions, not findings. Shared cM is consistent with a range of relationships and never proves one.
AI Chat reads what the Client has gathered. That sets some limits worth knowing:
| Topic | What to expect |
|---|---|
| Family trees | Trees are available for Ancestry matches only. An FTDNA or 23andMe match showing “no tree” here may well have one on that site. |
| Common ancestors | Found by spotting the same name and birth year in two Ancestry trees. Treat them as leads to check. A spelling or date difference will hide a real common ancestor, so “none found” means little. |
| Segments | Not available for Ancestry kits (Ancestry does not provide them). |
| The same person at two companies | Appears as two separate matches. The AI will usually notice the shared name and tell you. |
| Notes and free text | Your research notes are not sent to the AI. |
| Anything not gathered | If you gathered shared matches only above a certain cM, or skipped trees, the AI cannot see what isn't there. A fuller gather gives better answers. |
| Ancestral hubs, clusters | Not available in AI Chat yet. Use the Autosomal tools for clustering. |
The AI Chat button shows for every member, so everyone can see it is there; using it needs a Gold or Angel subscription. Without one, pressing it shows a message instead of opening the chat. If you are Gold and still get that message, log out of the Client and log back in so it picks up your current subscription.
AI Chat can be switched off for everyone while a problem is being fixed; the button comes back once it is switched on again. Also check that you are on the latest version of the Client.
A last safety check found something that looked like a real name in what was about to be sent, and stopped it. Nothing was sent. Rephrase the question, or refer to the person by their token — it appears in the tan tool lines and the pink Privacy lines.
For a cloud AI, check the API key and that your account with that company has credit. For a local AI, check that Ollama or LM Studio is running and that the Endpoint address is right.
The pink Privacy line shows who it picked. Ask again with the fuller name, or add something that narrows it down: “the Elsie Marlow who shares about 85 cM on Ancestry”.
York, Washington and Paris are places, and also surnames in somebody's tree. Place names such as “New York” or “North Carolina” are recognized as places. For a single word, say which you mean and it will be read that way: “born in York” is the place; “the surname York” or “the York line” is the family.
The list needs at least two letters after the @, and the chat must be connected. If it stays empty, nobody in your gathered data has a name starting that way — try the surname, or a different spelling.
Smaller and local models make more mistakes; a stronger cloud model is the usual fix. If a figure looks wrong on a strong model, please tell us the question you asked — you can leave the names out.