--- title: "How to Use ChatGPT for Genealogy, and How to Give It Your Real Records" description: "ChatGPT is good at reading a hard document and planning a search, and bad at knowing anything about your ancestor. Here is what to ask it, what never to trust it with, and how to connect it to your..." canonical: "https://kleiobase.com/blog/how-to-use-chatgpt-for-genealogy" last_updated: "2026-11-02" --- # How to Use ChatGPT for Genealogy, and How to Give It Your Real Records ChatGPT is good at reading a hard document and planning a search, and bad at knowing anything about your ancestor. Here is what to ask it, what never to trust it with, and how to connect it to your own family archive so it answers from your records instead of from memory. Claims about ChatGPT and KleioBase last verified 2026-10-30. You paste a baptism entry into ChatGPT and ask what it says. The answer comes back fast and fluent: a child's name, the parents, the godparents, a date. Then you ask a follow-up, "where would her marriage be recorded?", and get a confident answer that names a register, a parish and a year range. One of those answers is a reading of a document you supplied. The other is text that looks like knowledge. Most of the trouble people have with ChatGPT in family history comes from not telling those two apart. It reads, translates, organises and drafts well. It does not know who your ancestors were, and when you ask it anyway it seldom says so. **The short answer.** Use ChatGPT for genealogy as a reader and a planner, not as a source. Give it the document and tell it what kind of document it is, ask it to mark every word it is unsure of, and check each name against the image. Ask it where a kind of record would be kept, never whether a particular record exists. Do not copy anything it says about a specific ancestor into your tree until you have found it in a record. If you want it to answer from your own research rather than from memory, connect it to your archive with a custom MCP server, choose read-only access first, and ask it which record each answer came from. ## ChatGPT can read your documents but does not know your family ChatGPT works on what is in front of it and invents what is not. **Reading a document you supply is its strength.** Give it a scan of an 1880 census page, a Latin baptism from an Irish Catholic register or a Spanish *bautismo* from a parish in Jalisco, and it can usually tell you what kind of entry it is, what the columns or formulas mean, and what most of the words say. A team led by the historian Mark Humphries at Wilfrid Laurier University tested multimodal models from OpenAI, Anthropic and Google on 18th- and 19th-century English handwriting, posting the results on arXiv in November 2024. They measured character error rates of 5.7 to 7 per cent, and as low as 1.8 per cent when the models corrected an existing transcription. That is a good first draft. It also means several wrong letters on a typical page, and in a register the wrong letters are often in the surnames. **Planning and drafting are close behind.** Describe what you know about a person, what you have searched and where you are stuck, and it will suggest record types you have not tried, spelling variants and nearby places. It can turn your notes into a research log or a readable family story. In both cases the facts come from you. ChatGPT rearranges them. **Recalling facts about your family is its weakness.** ChatGPT has no access to the parish registers, census indexes or vital records that hold your ancestors, unless a tool gives it that access. Asked about a specific person, it produces text with the shape of an answer. ## A confident citation is not evidence that the record exists The best measurement of how often this happens comes from outside genealogy. William Walters and Esther Wilder, writing in *Scientific Reports* in 2023, had ChatGPT write short literature reviews on 42 topics and checked all 636 citations in the 84 resulting papers. 55 per cent of the citations from GPT-3.5 were fabricated, and 18 per cent of those from GPT-4. Even among the real ones, 43 per cent of GPT-3.5's and 24 per cent of GPT-4's contained substantive errors. **One thing not to claim, because nobody has measured it.** We searched again in October 2026 for a study of chatbots inventing genealogical record citations, such as a register volume or a film number, and found warnings on blogs and forums but no measured rate. Walters and Wilder tested scholarly references with older models, and a newer model with web search switched on is a different case. The direction is clear, but the rate for your kind of question is not known. Our comparison of [which AI tools can actually find your ancestors' records](/blog/which-ai-can-find-your-ancestors-records) goes through what search changes and what it does not. **"Where would this kind of record be kept?" is a fair question.** Questions like where Ohio kept death certificates in 1910, which archive holds a Mexican diocese's registers, or what an Irish civil registration district was called have published answers, and ChatGPT with search switched on can usually find the page that says so. Open the page it cites and read it. **"Does this record exist?" is a question for the record itself.** Whether your great-grandmother's baptism is in a particular volume can only be answered by the volume. If ChatGPT names one, treat it as a place to look, and record what you find there, including nothing, as [negative evidence](/blog/negative-evidence-in-genealogy). ## Tell it what the document is before you ask what it says Start every transcription request with context. ChatGPT reads a page better when it knows what the page is supposed to contain. **Name the record, the place, the date and the language.** "This is a Latin baptism from a Catholic parish register in County Cork, about 1840. Transcribe it, then translate it" gets a better result than "What does this say?". It tells the model to expect Latin, a fixed baptismal formula, the parents' names and the godparents. **Ask for doubt to be marked.** Add: "Put any word you cannot read with confidence in square brackets with a question mark, and do not guess at surnames." Without that instruction, a misread name looks the same as a correct one, which is the problem our guide to [how accurate AI handwriting transcription is](/blog/how-accurate-is-ai-handwriting-transcription) describes in detail. **Keep the columns.** For a census page or a register laid out in a grid, ask for a table with the original column headings, one row per person. You can then compare row by row against the image instead of rereading a paragraph. **Ask for the transcription and the interpretation separately.** "First give me exactly what is written, including abbreviations. Then, separately, what it means." That stops a helpful expansion of *fil. leg.* to "legitimate son" from replacing what the clerk wrote. ## Give it facts and ask it to arrange them When you want a research plan, a timeline or a story, the rule is the same: supply the facts, and tell it not to add any. A prompt that works: "Here are the records I have for Ann Delaney, with dates and places. Using only these, list the gaps in her life between 1850 and 1900, and for each gap name the kind of record that usually covers it in Ireland and in Massachusetts. Do not mention any record I have not listed as something I have." The last sentence matters, because without it ChatGPT will happily write "her 1855 emigration record shows" about a record nobody has seen. The trouble with this method is the pasting. Every conversation starts from whatever you copied in, and the copy goes out of date the moment you confirm another record, which is the problem a connection solves. ## Connecting ChatGPT to your family archive This section describes our product, so read it as disclosure: we make KleioBase, an AI research companion that turns scattered historical records into one connected family archive. The same method works with any family history service that offers an MCP server. MCP, the Model Context Protocol, is an open standard that lets an assistant ask another service for information while it works. Connect ChatGPT to a KleioBase archive and, instead of reading a pasted profile, it can search your people and records, read a family tree or a person's timeline, open the record an extracted fact came from and look at the scan itself. **The steps, on ChatGPT on the web.** KleioBase was not in ChatGPT's plugin directory when we searched for it there on 30 October 2026, so you add it as a custom MCP server: 1. Open **Plugins** in the ChatGPT sidebar, select **Add**, then **Add custom MCP server**. 2. Name it KleioBase and paste the address `https://kleiobase.com/api/mcp`. 3. Choose **OAuth** as the sign-in method, tick the box saying you understand the risk, and create it. 4. ChatGPT opens a KleioBase page. Sign in, choose **Read only** or **Read and write**, and select **Allow access**. In a new conversation, type `@` and pick KleioBase, so ChatGPT knows to use it for that message. The [Connecting AI assistants guide](/docs/ai-assistants) keeps the current steps, including the older Settings > Apps route some accounts still show. **Which plans include it can vary.** Whether **Add custom MCP server** appears depends on your ChatGPT plan and, in a work or school account, on what the workspace's administrator allows. If you do not see it, check your plan or ask your administrator. Do the setup on ChatGPT on the web, because OpenAI's help article says custom MCP servers cannot be connected from the mobile app. ## Decide what ChatGPT may change before you let it A read-only connection can answer questions. A connection with write access can also edit your tree, and that needs supervising. **Start with Read only.** KleioBase asks when you connect. A Read only connection can look up anything in your archive and change nothing; if ChatGPT tries to change something, it is told the connection is read-only. **ChatGPT has its own approval setting.** OpenAI's help article on connected apps lists the choices as Always ask, Allow read actions, Allow low-risk actions and Allow all actions. We would pick Allow read actions, so that lookups run straight away and anything that changes your research waits for you. **Changes are labelled.** Of KleioBase's 88 tools, 35 only read. Every tool that deletes, overwrites or merges is marked as destructive even when it can be undone, and each request carries a one-sentence description of what it is about to do, which ChatGPT can show you when it asks for approval. Deleted people and records go to the Trash, a merge can be undone, and emptying the Trash cannot. **You can cut it off.** Settings > Developer in KleioBase lists every connected assistant, its permission and when it was last used. Select **Revoke** and it stops working at once. **Mind what leaves your archive.** Whatever ChatGPT reads from your research is handled under OpenAI's terms, not ours. OpenAI's help article says that for Free, Plus, Go and Pro users it may use information accessed from apps to train its models if the "Improve the model for everyone" setting is on, so check that setting before you connect a tree that includes living relatives. Our guide to [what not to publish about living relatives](/blog/genealogy-privacy-ethics-publishing-family-data) covers why that matters. ## Five prompts that work once ChatGPT can see your records Each of these is a question a pasted profile could not answer. **Find the gaps across the whole tree.** "@KleioBase Who in my Delaney workspace has a birth but no death recorded, and when were they last seen in a record?" ChatGPT searches every person in the workspace, including the ones you would have forgotten to paste. **Check an extraction against the scan.** "Open the 1900 census record for the Rossi household and compare the names that were extracted with the image. List any that look different." It reads the extracted text and the scan side by side and gives you a list of differences to check yourself. **Build a timeline from the records, with sources.** "Give me Ann Delaney's timeline and name the record behind each event." Because every event comes from a record in your archive, you can open the record and check it. **Turn the gaps into a plan.** "What records does my archive expect for Giuseppe Rossi that I do not have yet? Order them by which is easiest to get." KleioBase keeps a list of the records each person should appear in but does not have yet, based on where and when they lived, and ChatGPT can read it. **Review suspected duplicates without merging.** "List the possible duplicate people in my archive, and for each pair compare what their records say. Do not merge anything." The instruction at the end costs nothing, and a Read only connection would refuse the merge anyway. ## What KleioBase does with ChatGPT, and what it does not We are a tool in this space, so read this as disclosure. KleioBase is an AI research companion that turns scattered historical records into one connected family archive. You upload record photos, review what was read from them, and confirm people, who are then linked across every record that names them, with duplicates and research gaps flagged for you. Connecting ChatGPT gives it the same archive you see. It works on every KleioBase plan, including the free Explorer plan, and your plan's limits apply as they do in the app: uploading is free, processing a record uses credits. Research findings, the suggestions KleioBase generates about gaps and inconsistencies, are available from the Researcher plan up. The [Research Companion](/docs/research-companion) inside KleioBase does the same job without a second account. In our view the connection is for the times you are already working in ChatGPT and want your research in the same window. We described the reasons for building it in [the announcement](/blog/use-claude-chatgpt-with-your-family-tree). There are three honest limits. **ChatGPT's conclusions are still its own.** It gets the same records you see, but what it infers from them can be wrong. Ask which record a claim came from and check it before you accept a merge or a correction. **Photos only, one image per record.** Uploads must be JPG or PNG. For photos on your computer, ask ChatGPT to create an upload link: KleioBase makes a page, valid for 15 minutes, where you drop the images in. A multi-page PDF has to be saved as separate page images first. **It cannot search other sites for you.** A connection gives ChatGPT your archive, not FamilySearch's or Ancestry's. Finding a record you do not have yet is still a search you, or ChatGPT's own web search, have to run. ## Where to start on your first connected session Connect with Read only, and keep ChatGPT's approval setting on Allow read actions. Pick one person you know well, with five or more confirmed records. Ask ChatGPT for their timeline with the record behind each event, and check two of the events against the records yourself. Then ask it what records the archive expects for that person and does not have. Choose one, search for it, and if you find it, upload it and confirm it. When you trust what it does with Read only, reconnect with Read and write and let it file the next one. The parish clerks and census enumerators who wrote these records could not have imagined a machine reading them, and they did not write for one. What they wrote is still the evidence. A connected ChatGPT is useful because you can make it cite the entry behind an answer, and then check what the clerk wrote. Canonical: https://kleiobase.com/blog/how-to-use-chatgpt-for-genealogy