Using AI and ChatGPT for Genealogy: What Works and What Does Not

AI is a research assistant, not a researcher

Large language models like ChatGPT, Claude, and Gemini have changed how genealogists work. Used carefully, they can save hours on tedious tasks. Used carelessly, they introduce errors that take years to undo.

The central rule: AI is an assistant, not a source. It cannot verify facts about your ancestors. It cannot access genealogy databases. It will sometimes generate fluent, plausible, completely fabricated information ("hallucinations") in response to your questions. Every AI output must be verified against a real primary source before it enters your tree.

With that caveat in mind, here are the tasks AI does well.

Tasks AI does well

1. Transcribing handwriting Upload a photo of a 19th-century parish register entry, will, or census page and ask the AI to transcribe it. Modern multimodal AI handles secretary hand, copperplate, and even some Gothic scripts with reasonable accuracy. Always check ambiguous letters against context.

2. Translating documents AI translates Latin baptismal entries, German Kirchenbücher, French état civil, and Italian stato civile fluently. It also handles older variants (Church Slavonic, medieval Latin) better than most online translators.

3. Summarising long documents Feed it a 12-page probate inventory and ask for a list of beneficiaries and bequests. Excellent for triaging large record sets.

4. Generating research hypotheses "Given that John Smith was born in 1810 in Cork and emigrated to Boston before 1840, what record sources should I check for his parents?" produces useful, structured next steps.

5. Reformatting data Converting place names to standard form, normalising date formats, suggesting GEDCOM tag fixes. AI is fast and accurate at format manipulation.

Tasks AI does badly

1. Genealogical "facts" about specific people Asking "who were the parents of John Smith born 1810 in Cork?" will often produce a confident answer with names, dates, and places — none of which is real. The model invents plausible-looking ancestors. Never accept this output without an independent primary source.

2. Citing sources AI will fabricate citations to books, archives, and websites that do not exist. Always click through any URL it provides, and verify any cited book in a real library catalogue.

3. DNA interpretation While AI can explain shared cM concepts well, it cannot reliably analyse your specific match list. Use established DNA tools and GEDminer's DNA Planner for actual analysis.

4. Solving brick walls AI cannot do research. It can only manipulate what you give it. A brick wall requires new evidence, not new prose.

Privacy: keep your tree off public AI

Be careful what you upload to AI services. Most consumer AI tools store conversations and may use them for training:

• Never upload your full GEDCOM file to a public AI service — it contains personal data on living relatives
• Anonymise before uploading: replace names of living people with placeholders, strip emails and addresses
• Use offline AI for sensitive data: tools like Ollama or LM Studio run open-source models locally, with no data leaving your machine
• Read the privacy terms of any AI service before pasting personal information

GEDminer itself parses your GEDCOM entirely in the browser and never sends data to any AI service. For analysis of your tree, that is the safer route.

Effective prompts for genealogy

Better prompts give better answers. Some patterns that work:

For transcription: "This is a page from an 1851 English census. Transcribe every line as plain text, preserving the column structure. Mark unclear words with [?]."

For translation: "Translate this Latin baptismal record into English. Identify the child, parents, godparents, and date. Note any abbreviations you expanded."

For research hypotheses: "My ancestor John Smith was born about 1810 in County Cork, Ireland, and appeared in the 1850 US census in Boston. List the top 10 record sources I should check to identify his parents, in priority order, with the rationale for each."

For naming pattern analysis: "In Scottish tradition, what does the naming pattern of these eight children tell me about likely grandparent names? [list children]"

Avoid open-ended questions like "tell me about my Smith ancestors" — they invite hallucination.

Verification workflow

For every piece of AI output that affects your tree, follow this workflow:

1. Source check: What primary record supports this claim? If none, it does not enter the tree.
2. Independent search: Re-find the cited record yourself in the original archive or database
3. Cross-reference: Does this fact agree with what you already know? Contradictions mean one of them is wrong
4. Cite properly: Record the actual source you verified, not the AI conversation
5. Re-analyse in GEDminer: After adding new facts, re-export your GEDCOM and re-run the analysis. The Errors tab will catch any inconsistencies

This workflow takes 5-10 minutes per AI claim. It feels slow, but it is much faster than rebuilding a corrupted tree later.

Where AI fits in a modern workflow

A practical genealogist's workflow with AI in 2026:

• Documents in foreign languages or old scripts → AI transcription and translation, then verification
• Long probate or land records → AI summarisation, then read the original
• Research planning → AI brainstorming combined with GEDminer's Research Plan Generator
• Tree analysis → GEDminer (no AI hallucination risk, runs locally)
• DNA analysis → established DNA tools and reference data, not AI
• Source citation → manual, always

AI accelerates the boring parts of research without replacing the rigour that makes genealogy reliable.