An informal introduction to a powerful research assistant
Responding to members’ interest in artificial intelligence, Steve Miller led an informal and highly interactive Zoom session designed to remove some of the mystery surrounding AI. There were no slides and no attempt to turn the evening into a technical computer lesson. Instead, members were invited to ask questions, share experiences and watch a series of live demonstrations showing how conversational AI might assist family historians.
Steve’s central message was reassuringly simple: modern AI need not be intimidating. Unlike a conventional computer program full of menus and buttons, a large language model can be approached through an ordinary conversation. The challenge is not learning a new piece of software so much as learning to explain clearly what help is required.
How a large language model works
Steve began by explaining that tools such as ChatGPT, Claude and Gemini are large language models, usually shortened to LLMs. They have been trained on very large quantities of written material and generate answers by recognising patterns in language. They do not think, remember experiences or understand physical objects in the human sense.
To illustrate the point, members were asked to complete familiar phrases such as “Mary had a little…” and “fish and chips with salt and…”. Most supplied the expected words because they recognised familiar linguistic patterns. AI operates on a vastly larger version of that principle: it predicts how words are likely to fit together. As Steve observed, it has never tasted fish and chips or sung a nursery rhyme; it works from patterns.
This distinction matters because fluent language can sound authoritative even when the underlying answer is wrong. AI can “hallucinate”—producing plausible but invented information. Genealogists must therefore treat it as an assistant, not as evidence.
The art of asking a useful question
A vague question such as “My grandfather was John Smith—what can you tell me about him?” gives an AI system almost nothing reliable to work with. A stronger request includes known facts: an approximate birth date and place, relatives, occupations, addresses and the exact research problem. Steve likened a good prompt to a recipe. The better the ingredients and instructions, the more useful the result.
He recommended telling the system what role to adopt—perhaps “an experienced genealogist”—then stating the goal, supplying relevant context, specifying the desired output and adding safeguards such as: do not guess; distinguish facts from suggestions; identify uncertainty; and tell me what needs checking. If the first response is too technical or disappointing, the user can ask for simpler language, a different format or advice on improving the prompt.
A new GeneAIlogy area for members
The meeting introduced the Society’s new GeneAIlogy section in the members’ area. It includes beginner-friendly guidance, principles covering accuracy and disclosure, advice on prompt writing and a growing library of reusable “AI recipes”. Members can choose a task, enter details into a form and generate a structured prompt to copy into their preferred AI service.
Examples already available or demonstrated included help with a brick wall, timelines, biographies and analysis of a death certificate. The recipe approach is intended to make good prompts reusable, consistent and accessible to people who have never used AI before. Steve invited members to suggest further prompts so that the library can develop around practical family-history needs.
Reading a death certificate
The first substantial demonstration used an uploaded death certificate. AI extracted details relating to Lou Calvert, including her sex, place of death at Strangeways in Manchester, family and address information, occupation as a weaver, and a cause of death described as a fractured dislocation of the neck caused by judicial hanging. It then suggested possible next steps, including the 1921 Census, inquest and coroner’s material, prison records and criminal records.
The demonstration showed both the power and the weakness of the technology. Much of the certificate was read accurately, but one address or number was doubtful and required human inspection. Steve emphasised that handwriting remains difficult and that every transcription must be compared with the original. AI was most useful not because it magically found the ancestor, but because it drew out clues and proposed a research route.
Headstones, handwriting and imperfect images
A difficult headstone image provided a sterner test. Members could make out fragments of the inscription themselves, but the live AI transcription was slow and hampered by connection interruptions. The episode became a useful lesson in its own right: artificial intelligence is not infallible, poor images remain poor evidence, and a live demonstration does not always behave as expected.
Members discussed their own experience of automated transcription. One had found Ancestry’s transcription assistance useful for wills, although it could seriously misread an occasional word. Another noted that checking a handful of doubtful words was still much quicker than producing the entire transcript manually. Sixteenth-century handwriting was recognised as a particularly difficult area, with Transkribus mentioned as a specialist service, though experience of it was mixed.
Turning a PDF into a structured spreadsheet
One of the evening’s most convincing demonstrations involved a typescript of settlement records supplied by Stan Merridew. Steve asked AI to list the individuals by surname, forename and date and to create an Excel workbook. The resulting file contained 28 named individuals drawn from 17 source entries, along with a separate sheet for unnamed wives, children and family groups. It retained source wording for audit purposes, added filters and noted inferred surnames and reading certainty.
This suggested immediate possibilities for Society projects. Printed parish-register transcripts, settlement material and other PDFs might be converted into spreadsheets or database-ready tables, provided that the desired columns are specified and every result is checked. A particularly useful instruction is to forbid guessing and require uncertain words to be placed in a separate review column.
Photographs as collections of clues
Members considered whether AI could decide if two photographs taken many years apart showed the same person. Steve advised against relying on facial comparison alone. Clothing, estimated date, studio setting, buildings, landscape, objects and other contextual details may all contribute pieces to the identification puzzle. AI may not name the person, but it can help identify a likely period or draw attention to overlooked features.
Sheila Harris described using Google’s AI mode to investigate a business once run by her great-grandfather in Newcastle upon Tyne. It returned links to websites and photographs she had not previously found. This was a good example of AI widening the search rather than supplying a final conclusion.
From old church image to artwork
A lighter demonstration used an image of Langcliffe Church from a 1914 book identified in the transcript as North Craven Churches. A reusable prompt asked for a crisp black-ink topographical drawing with cross-hatching and a gentle watercolour wash. The source photograph, taken from a printed page, showed a moiré pattern of coloured bands. After a lengthy wait, the AI produced an attractive interpretation that members greeted warmly, while the inherited banding remained visible.
The example showed how conversational refinement works: a user can ask for more watercolour, a different artistic treatment or an improved prompt, and continue the exchange until the result is suitable. It also raised questions about copyright and the right to upload or transform source images.
Writing, organising and Society administration
Steve described using AI to draft website guides, translate material connected with Czech research, prepare an article on Poor Law records, develop website code and create meeting reports from Zoom transcripts. A consistent instruction can turn a transcript into a detailed website report within minutes, after which someone who attended the meeting checks and edits it.
Further ideas included proofreading the Society journal, extracting names from articles for a searchable database, converting printed text into structured data, and using Codex locally to rename and sort files into folders. These are repetitive, time-consuming jobs for which AI can be especially effective, although local file access and automated changes should be used carefully.
Members’ experiences and discussion
The discussion repeatedly returned to research strategy. Helen Reeves had found AI frustrating when asked to locate an ancestor directly: it could lead the user down a long path without producing the hoped-for “silver bullet”. She found it far more effective for suggesting search strategies and explaining social history, including the possible nonconformist origin of a name.
Lynda Balmforth shared a striking example concerning James Webster, a cabinetmaker whose death and burial had proved elusive. A long AI-generated analysis suggested searching the whole workhouse, infirmary or hospital in the 1841 Census, trying the forename without the surname and carefully examining institutional images. It did not solve the mystery, but it offered lines of enquiry she had not considered. Her description neatly captured the evening’s theme: AI opened her eyes to possibilities that had previously seemed like “a cloud in the distance”.
Questions also covered free and paid access. Steve stressed that free versions of ChatGPT, Gemini and similar services are sufficient for experimentation, and members should try several to see which suits them. A subscription may be worthwhile for intensive use or advanced features, but it is not necessary for getting started.
Privacy, copyright and verification
The most important cautions were practical. Members should not enter passwords, bank details or sensitive personal information. Photographs and records should only be uploaded after considering privacy, consent, copyright and the terms of the website from which an image was obtained. The group discussed the uncertainty surrounding the reuse of census images downloaded from commercial services; this should be checked against the relevant provider’s terms rather than decided by AI itself.
Above all, every factual answer must be verified against original records and reliable sources. AI can transcribe, summarise, compare, suggest and organise, but it cannot replace a citation or the genealogist’s judgement.
A successful first step
Despite slow image generation and the occasional interruption, members responded enthusiastically to the informal format and supported the idea of a future follow-up session. The evening succeeded because it showed AI realistically: impressive in some tasks, fallible in others, and most useful when directed by someone who understands the evidence.
Steve’s concluding advice was both practical and memorable: start with a free tool, have a conversation, ask targeted questions, and do not be frightened to correct the system or try another one. “If you can talk to people, you can use AI.” For family historians, the opportunity is not to hand research over to a machine, but to gain a patient assistant capable of revealing clues, organising information and suggesting the next sensible question.
Memorable Quotations
- “It doesn’t think in the way that we do.” — Steve Miller
- “It’s chat. It’s the art of conversation.” — Steve Miller
- “If you get all the right ingredients, then you’ll get something good out of it at the end.” — Steve Miller
- “Think of it more as an assistant for you.” — Steve Miller
- “It won’t perform miracles, but it’ll certainly help you.” — Steve Miller
- “If you can talk to people, you can use AI.” — Steve Miller
- “It opened my eyes to something that was like a cloud in the distance.” — Lynda Balmforth