Why AI invents ancestors, and how to catch it.

A family story can sound complete long before the evidence is complete. A name fits. The dates look possible. Someone supplies a reason for a move, and the whole account starts to feel settled.

AI can make that account read smoothly while getting its facts wrong. The Coalition for Responsible AI in Genealogy warns about false and incorrect content. It asks researchers to check accuracy against other records. Read the coalition’s principles.

The danger is a statement that loses its doubt as it passes between people. Suppose an old note says a man was “perhaps a cousin”. A later account calls him a cousin. A tree then makes the relationship look established. This is how a guess hardens with each retelling: repetition removes the question while adding no evidence.

The example above is invented to explain the problem. The case further down comes from Ed’s family archive. It shows why a useful assistant needs a method for correcting its answers.

Before you type a family name into any AI, turn off training on your account; the privacy guide shows you where.

What a convincing answer leaves out

Consider the sentence “Harold moved north to take a better job”. It contains several statements. Harold moved. The new address lay north of the old one. A job changed. The job improved his position. The job caused the move.

Two dated addresses might support only part of that sentence. A new occupation might support another part. The motive still needs its own evidence. Treating the sentence as one statement hides those differences.

One sentence, five statements

Suppose you have found Harold in two censuses, ten years apart. They give different addresses and different jobs. A family story says the rest. Here is that sentence split into statements.

Statement	How we know	How sure	What would settle it
Harold moved house	Official record	High	Already settled: the two census entries give two different addresses
The new house was north of the old one	Later copy or index	High	Already settled: both addresses appear on a published map
Harold changed jobs	Official record	High	Already settled: the two census entries give two different jobs
The new job was a better one	Worked out	Low	Worked out from the job titles alone. A wage book, a trade directory or a letter would settle it
He moved to take that job	Told second-hand	Low	A story passed down the family. A letter, a newspaper notice or an employer’s record from around the move would settle it

Every row also rests on one hidden statement: that both census entries are the same Harold. That statement is a guess, marked low, until his age and birthplace match in both. The whole sentence sounded settled. Only three of its five parts are.

Ask an assistant to separate the statements before asking it to write a life story. Keep its suggested sources on a search list until you have opened them. A plausible title or archive reference is not evidence that a record exists.

If the assistant cannot show a source, treat its answer as a guess, marked low. Asking the same question again does not change that. Nor does a confident explanation of why the answer sounds right.

Read two marks on every statement

How we know says where a statement came from. An official record made at the time is the strongest kind. A memory passed on second-hand is weaker. A conclusion worked out from other statements is the weakest kind of all. The mark says how the statement reached you. It cannot prove it true: a diary can lie.

How sure is your judgement: low, medium or high. Base it on how we know, on how many independent sources agree, and on how many times the story was retold. Keep the statement narrow. “The entry names Harold as a weaver” says what the source records. It does not settle every question about Harold’s identity.

A worked-out statement can still be high. Record the reasoning, the alternatives you considered and the evidence that favours your conclusion. Readers need its basis.

The marks describe single statements. They do not certify an entire person or branch. A birth date may be supported while the parent link remains open. A correctly copied record can preserve a mistake made by its original author.

Keep the source wording beside your conclusion. Add the date of your review. If new evidence changes the conclusion, record the change and its reason. A mark without that trail can make a weak statement look better researched than it is.

The older standard behind the method

The Genealogical Proof Standard is the set of rules professional genealogists work to. It asks researchers to do five things: research the question widely enough, cite sources accurately, compare evidence, resolve conflicts and write a reasoned conclusion. FamilySearch’s explanation of the standard describes these requirements.

These marks help you organise the work. They do not replace the standard. In particular, attaching a source does not resolve a contradiction. You must explain why the competing evidence differs, or leave the conclusion open.

Nor does a free search route promise to complete every investigation. If an essential record remains inaccessible, record that limit. “I have checked the sources available to me” is different from claiming that every relevant possibility has been examined.

The coalition adds principles for AI use: accuracy, disclosure, privacy, education and compliance with relevant rules. Its published principles explain each. A useful disclosure tells readers what the assistant did and what the researcher checked. It does not hand responsibility to the software.

The photograph that did not show Ed’s grandfather

Names in these examples are changed. The records and the reasoning are real.

Ed found a cropped RAF training photograph that might show one of his grandfathers, here called Arthur. The resemblance was tempting. The archive stored the proposed match as a guess, marked low.

A second copy preserved the full name key, reading from left to right. It accounted for all ten men. None was Arthur. The photograph also showed pupils, whereas Arthur was an instructor.

The evidence trail names the cropped photograph and the full copy. It also records the evidence that rejected the identification.

The name key did the decisive work in Ed’s account. The instructor detail supported the rejection, but would not by itself exclude someone from every group photograph. State only what this evidence settles: this picture does not identify Arthur among its ten named subjects.

Preserve the cropped version and the reason it once looked promising. Attach the fuller copy and the correction. Add a new statement that proves the match wrong, citing the full name key. Keep the old statement on the record. Do not leave it circulating with a softer caption such as “probably Arthur”.

This correction matters because the first guess belonged to the researcher. The method has to test the family’s preferred answer as carefully as a stranger’s suggestion.

What this case establishes

The story above is drawn from Ed’s cleared archive notes. The evidence trail remains in his archive. The original records were not independently reopened for this guide.

It shows evidence rejecting an identification the researcher hoped was true. The useful result depended on keeping the question separate from the answer we would like.

A correction is part of the research record. Keep enough detail for your future self to understand why you changed your mind. Otherwise an old note can revive the same mistake years later.

Step 1. Give the assistant one small job

After turning off training and removing living-person details, start with a short passage you are entitled to share. Use this instruction:

Instruction to paste

Split this passage into single statements. Name the source given for each. Say how we know each statement and how sure it is. Mark any statement with no source as a guess, low, and explain it. List contradictions and questions separately. Do not add people or invent citations.

Read the result against your passage. Check for a changed date, an added relationship or a confident interpretation of an unclear word. Correct the working note before asking for a narrative.

Step 2. Check the statement before publishing it

Open the cited source yourself. Compare the exact wording with the statement. If you cannot inspect it, preserve the limitation. An assistant’s report of what a source says is not your own inspection.

For a disputed identity, write down what would disprove the match. Then search for that evidence. A conflicting household, a complete photograph caption or a differently named parent can change the conclusion.

Use the verification checklist before adding the result to a shared tree. The kit in development will teach this method; it cannot remove the need for your judgement.

Want to hear when a new guide is published? Get an occasional email.