The Obituary That Never Was: How to Catch AI Hallucinations Before They Enter Your Family Tree
Photo: Yourusernamewillbepublic2, CC0, via Wikimedia Commons
A genealogy friend recently asked an AI chatbot to draft an obituary for her great-grandfather. She gave it a name, a rough date of death, and a city. The tool produced a warm, polished paragraph: a loving father of six, a church deacon, a man who worked for the railroad for thirty-two years and was buried beside his wife, Eleanor. There was just one problem. He was a father of four. He was not a deacon. He never worked for the railroad. And his wife’s name was Helen.
The story is funny until it is not. AI tools are now woven into everything from record-search platforms to genealogy forums to the chatbot window open in the next tab. They can summarize dense probate files, translate old newspaper clippings, suggest research paths, and draft narrative sections of a family history. Used well, they are useful assistants. Used carelessly, they are confident-sounding fiction machines that can slide invented ancestors into a shared tree before anyone checks the source.
This is not a call to ban AI from genealogy. It is a call to treat it like a very fast, very eager intern who has never actually been to a courthouse.
The Problem Is Confidence, Not Creativity
The unsettling thing about AI-generated family history is not that it makes mistakes. Humans make mistakes too. The unsettling thing is the tone. A chatbot can invent a birth year, a spouse, a migration route, and a cause of death in the same smooth paragraph it uses to say good morning. There are no furrowed brows, no “I’m not sure,” no margin note in pencil. It just writes.
That is why the first rule is simple: AI output is a lead, not a source. It can point you toward a record. It can help you phrase a question. It cannot stand in for the record itself. If a sentence cannot be traced to a document, a photograph, a headstone, an interview, or another verifiable source, it belongs in a draft folder — not in the family tree.
Rule One: Make the Machine Show Its Work
Ask for citations. Then check them. If the tool says your ancestor was born in 1912 in Ohio, ask which record says so. If it cannot name a source, treat the detail as unverified. If it names a source, look it up yourself. AI tools can misread indexes, blend two people with similar names, or cite a book that does not exist. The only citation that counts is one you can open.
Rule Two: Separate Fact, Inference, and Story
Family history has three layers. Fact is what the records say. Inference is what you reasonably conclude from the records. Story is the narrative you build around both. AI is good at story. It is not reliable at fact, and it is often too smooth about inference.
A practical habit is to keep three columns or three colors in your notes: documented, probable, and narrative. “Born in 1889 according to the county birth register” is documented. “Probably moved to Chicago after the 1910 census” is inference. “A restless young man chasing opportunity in the big city” is narrative. None of these are bad. They just should not be mixed together without labels.
Rule Three: Date and Place Discipline
AI tools love to round dates and smooth places. A birth in “the 1880s” becomes 1885. A death in “rural Appalachia” becomes a specific county with a confident zip code. Make the machine preserve uncertainty. Ask it to flag unknown values as unknown, not to guess. Then check every date and place against an original record or a reliable index.
Be especially careful with obituaries. They are already tricky sources: written by grieving families, based on secondhand information, and sometimes wrong about the details that matter most. An AI-generated obituary is not a source at all. It is a writing exercise. Label it as such.
Rule Four: Never Paste Raw AI Into a Shared Tree
This is the rule that saves friendships. If you are working in a collaborative tree, do not paste AI-generated text, dates, or relationships directly into a shared profile. Put it in a private research note, mark it clearly as AI-assisted, and verify it first. Shared trees are built on trust. One unverified paragraph can travel through a family network faster than a rumor at a reunion.
Rule Five: Use a Red-Team Cousin
Every research group needs at least one skeptic. That is the cousin who asks, “Where did this come from?” and does not accept “the computer said so” as an answer. If you do not have one, be one. Before you add a new ancestor, ask a second person to look for the weakest link in your evidence. AI can generate ten plausible details in seconds. A good research buddy can spot the one that does not fit.
A Sensible Workflow
Use AI for the boring parts: summarizing a long record, suggesting search terms, translating a foreign-language newspaper clipping, or drafting a list of questions for a relative. Then do the human parts: verify, cite, compare, and decide. Keep a research log. Note what the tool suggested and what you confirmed. If the tool invents something, write that down too. Future you will appreciate the warning.
The goal is not to keep AI out of family history. The goal is to keep fiction out of the family tree. A chatbot can help you tell a better story. It cannot tell you what happened. That part still belongs to the records, the relatives, and the researcher with the patience to check.