AI Explores the Family Tree

From transcribing parish records to hallucinating ancestors, AI has the ability to impact every stage of family history research.

Millions of people research their family histories every year, sifting through census records, deciphering centuries-old handwriting, and following paper trails across borders and languages. The work is slow by nature, governed by whatever records survived and whoever transcribed them. AI has begun reshaping every stage of that process.

Processing the archives

Genealogy runs on historical records: census forms, ship manifests, parish registers, military draft cards. Archives hold millions of these documents in handwritten form, and until recently only human volunteers could make them searchable, transcribing one document at a time. That bottleneck left the vast majority invisible to researchers.

FamilySearch deployed AI handwriting recognition to read those documents at scale, extracting names, dates, and genealogically relevant information from records that volunteers would have taken decades to process. Ancestry invested $450 million over 10–15 years in digitizing records and applying AI to them, an effort that CEO Howard Hochhauser described as the engine behind the company’s return to growth.

Compiling life stories

A searchable record remains a data point, a name matched to a date and a port of arrival. The genealogy platforms began using AI to synthesize those fragments into something that reads like a life. FamilySearch product engineer Ryan Parker described the ambition at RootsTech as “reasoning applied to human connection.”

Ancestry launched AI-narrated audio stories that turn ship manifests and draft cards into spoken narratives of an ancestor’s journey and acquired iMemories to pull old family photos and home videos into the same pipeline. Ancestry also introduced Ancestry Networks, which groups people by shared communities, eras, and events beyond the traditional family tree.

MyHeritage built its own suite of narrative tools. Its AI Biographer generates summaries from the records attached to an ancestor’s profile, and its photo tools colorize, enhance, and animate old family photographs.

Fabricated histories

General-purpose AI models fabricate genealogical records with the same confidence with which they transcribe real ones. Earlier this year, a British family historian asked Google’s Gemini to help trace a will from 1590. Gemini fabricated the record, generated fake screenshots of The National Archives website (complete with misspellings), and admitted that its methodology had been “a series of errors” papered over with “more artificial information.” Steve Little, director of the National Genealogical Society’s AI program, has documented the broader pattern in his own testing. An AI transcribes one German parish record brilliantly, then confidently mistranslates the next, invents a date, or cites a source that does not exist.

Ancestry’s CTO Sriram Thiagarajan acknowledged that hallucinations in genealogy AI will never fully disappear. Ancestry built specific guardrails in response, training its AI not to assume that someone served in a war just because it found a draft card in their records.

At RootsTech 2026, a coalition of researchers published five principles for responsible AI use in genealogy, centered on accuracy, disclosure, privacy, education, and compliance. James Tanner, chairman of the Family History Guide Association, warned that once AI exhausts verifiable information, it starts trying to please people by telling them what it thinks they want to hear.

Augmented, not replaced

Genealogists bring a built-in advantage to the AI era. The hobby has always required cross-referencing sources and questioning inherited narratives, and these habits produce practitioners who scrutinize a claim regardless of its origin.

AI has expanded what genealogists can find and how quickly they can find it. Hochhauser told Semafor this week that genealogy enthusiasts’ joy lies “in the doing and discovery.” The tools that serve that process have changed dramatically, but the process itself has not.