Genealogy AI Tools

MyHeritage AI Time Machine, Storied AI, Familio, ChatGPT/Claude/Gemini for translation, OCR, summarization workflows.

The 2024–2026 LLM wave reshaped genealogy in three places: handwriting recognition (Transkribus + vision LLMs), translation of foreign-language records, and biographical narrative generation. For HTR specifically see Genealogy Document Scanning & HTR; for photo restoration AI see Genealogy Photo Restoration AI; for translation see Genealogy Translation & Paleography; for self-hosted LLMs see Self-Hosted AI / LLM.

Hosted commercial AI features

MyHeritage AI Time Machine — paid; reimagine ancestors in different historical eras (Roman togas, Renaissance dress, etc.). Gimmicky-but-popular.
MyHeritage Deep Story / AI Biographer — paid; auto-generated biographical narratives from tree facts. Quality variable; useful starting point.
MyHeritage AI Photo Color / In Color / Photo Enhancer — paid; restoration / colorization. See Genealogy Photo Restoration AI.
MyHeritage Deep Nostalgia — paid; "make grandma blink." Culturally controversial — flag.
Ancestry LifeStory / Hint generation — paid; LLM-flavored narrative + record-suggestion improvements rolled out 2024–2025.
Storied AI photo description — paid + free tier; AI captions for old photos.
Familio — paid + free; AI-prompted tree storytelling, narrative generation.
Storyline AI / Remento — paid; AI-narrated family interview compilation.
FindMyPast AI — paid; experiments with semantic search of records.

★ Claude (Anthropic) — paid + free tier; strong for translation of old-script records, paleography, biographical narrative drafts, citation formatting, research planning. Long-context (200K) lets you paste an entire census page or will.
★ GPT-4o / GPT-5 (OpenAI) — paid + free tier; comparable; strong vision (read scanned records).
★ Gemini 2.5 / Gemini Pro (Google) — paid + free tier; strong vision; cheap; long context.
Local LLMs — see Self-Hosted AI / LLM. Llama 3, Qwen, Mistral; free; privacy-respecting; quality below frontier.
Open-source vision LLMs — Qwen2.5-VL, InternVL, MiniCPM-V; free; OCR-capable; usable for genealogy on a modest GPU.

Read this old document for me — paste image, get transcription. Vision LLMs surprisingly good on legible 18th–19th c. hands; weaker on Sütterlin, secretary hand, faded ink. Pair with Transkribus for tough cases.
Translate this Latin/German/Polish/Russian/Hebrew parish record — vision LLM extracts text + translates; verify with reference dictionaries. The biggest practical AI-genealogy win in 2024–2026.
Summarize this obituary — extract structured fields (parents, siblings, spouse, children, residence history).
Format this citation per Mills' Evidence Explained — useful template generation; verify accuracy.
Plan a research strategy — "I'm stuck on John Smith born ~1820 in Kentucky; what records should I check?" — usable but generic; pair with FamilySearch Wiki locality pages.
Draft a relationship-proof argument — feeds reasoning + sources; require verification.
Cluster DNA matches — paste shared-cM data, let LLM hypothesize relationships. Use Genealogy DNA Third-Party Analysis tools as the rigorous version.
Normalize place strings — "Kraków, Galicia, Austria-Hungary" → modern coordinates + period-correct administrative name. See Genealogy Maps & Migration.

OpenAI / Anthropic / Google APIs — paid per token; build batch pipelines.
Replicate / Fal.ai / Together AI — paid per call; for vision / restoration models.
LangChain / LlamaIndex / Haystack — see AI Agent Frameworks; orchestrate "OCR → translate → extract → cite" pipelines.
Ollama / LM Studio — free; local LLM hosting. See Self-Hosted AI / LLM.
n8n / Zapier — paid + free; automate "new Paperless-ngx document → Claude OCR + extract → email summary."
Genealogy-AI Discord / Reddit communities — share prompts + workflows.

Hallucination is real — LLMs invent plausible names, dates, places. Always verify against the original document.
Date arithmetic — LLMs miscount days/months across calendar systems regularly. Use a calendar converter (see Genealogy Calendar Conversion).
Translation ambiguity — old place names and personal names have multiple plausible modern forms; LLMs pick one. Verify with gazetteers.
Citation format errors — LLMs approximate Evidence Explained without nailing the comma/colon punctuation rules.
Provenance discipline — annotate which fields in your tree came from LLM output and need verification.

Sending family documents to a cloud LLM — be aware your records may be used to train models depending on the API + policy. Use API endpoints with no-train commitments (Claude API, OpenAI API with data-control settings) over consumer chat UIs.
Living-person data — never send identifying information about living people to a third-party LLM without consent.
Local LLMs — see Self-Hosted AI / LLM. Run Llama 3 / Qwen on a Mac or modest GPU for full privacy. Quality below cloud frontier but adequate for many tasks.

Transkribus + vision LLM hybrid workflows — Transkribus for the bulk page-by-page; LLMs to clean up errors and extract structured facts.
Local vision LLMs improving fast — Qwen2.5-VL on a single GPU now reads many old scripts adequately.
MyHeritage / Ancestry / FindMyPast all rolling AI hint quality upgrades; semantic search ("find articles about a 1908 mining accident") emerging.
Long-context LLMs (Claude 200K, Gemini 2M) — paste entire ledger books, get name extractions in one shot.
AI-skepticism backlash — academic-genealogy community emphasizing "verify, verify, verify"; LLM hallucinations have led to bogus relationship claims in shared trees.

Default LLM for translation / paleography / drafting: Claude or GPT-4o/5.
Cheapest cloud vision (high-volume): Gemini Flash.
Fully private / no cloud: Ollama + Llama 3 / Qwen2.5-VL locally.
Best HTR for handwriting: Transkribus (specialty model) > vision LLM (general).
Hosted "do it for me" experience: MyHeritage AI features (paid).
Pipelines / batch processing: API + LangChain + n8n.
Verification-first workflow: treat every LLM output as a "lead," confirm with primary records.