Ep. 42: Writing the Proof - How AI Helps You Make Your Case

You can write a genealogical proof argument using 3 AI tools in one afternoon: NotebookLM organizes your evidence, Claude drafts and stress-tests the argument, and ChatGPT reviews it for plain-language clarity. This is the GPS Mini-Series capstone on Element 5.

What you'll learn:

• The three-part proof argument structure: Statement to be Proved, Evidence Presentation, Analysis and Reasoning
• How to prompt NotebookLM to surface inconsistencies in your sources before you write a single sentence
• How to ask Claude to draft an argument that flags its own logical weak points
• How to run a clarity review so the argument holds up for anyone who inherits your research
• The difference between a proof statement, proof summary, and full proof argument

This episode is for you if you search: how to write a genealogical proof argument, GPS Element 5, NotebookLM genealogy, Claude for genealogy research, AI tools for family history writing, BCG proof argument, Genealogical Proof Standard tutorial, genealogy AI workflow.

Outcome: Full Breakthrough. Two genuine source inconsistencies resolved. One finished, submission-quality argument. The AI held the pen. The standard was Brian's to meet.

Australian and UK researchers: the Genealogical Proof Standard is recognized across the English-speaking genealogical world. The Society of Australian Genealogists (sag.org.au) and the Society of Genealogists (sog.org.uk) both publish compatible research standards for your records.

Patreon members get the Companion Guide: 12 advanced prompts, a GPS Research Checklist, and a full multi-step workflow from raw sources to finished argument. Everything else is at ancestorsandai.com.

Connect with Ancestors and Algorithms:

📧 Email: ancestorsandai@gmail.com
🌐 Website: https://ancestorsandai.com/
📘 Facebook Group: Ancestors and Algorithms: AI for Genealogy - www.facebook.com/groups/ancestorsandalgorithms/

Golden Rule Reminder: AI is your research assistant, not your researcher.

Join our Facebook group to share your AI genealogy breakthroughs, ask questions, and connect with fellow family historians who are embracing the future of genealogy research!

New episodes every Tuesday. Subscribe so you never miss the latest AI tools and techniques for family history research.

Transcript

SPEAKER_00 0:11

Here's something I've been sitting with for a while. What is the actual difference between knowing something is true and being able to prove it? I've been researching one particular branch of my family tree for a couple years. I know in my gut who these people are. I've got census records, I've got a land deed, I've got a probate file, I've got three different sources that all point in the same direction. And yet every time I've tried to write it down, try to turn all of that into something someone else could read and trust, I've stalled out. The pile of evidence is right there. The conclusion feels solid, but the moment I try to commit it to paper, something stops me. Sound familiar? Today that stops. Let's dive in. Welcome to Ancestors and Algorithms, where family history meets artificial intelligence. I'm your host Brian, and today we're wrapping up the GPS miniseries, a four episode arc I've been building since episode thirty. We started with an overview of all five elements of the genealogical proof standard. Then in episode 35, we went deep on element four, resolving conflicting evidence. Episode 38 tackled element 1, the reasonably exhaustive search. And today we're closing the loop with element 5, the soundly reasoned, coherently written conclusion. The proof argument. And we're going to build one, start to finish, using AI as our drafting partner. Even if you've never heard the phrase proof argument before today, stick with me. By the end of this episode, you'll understand exactly what one is and you'll know how to write your own. So let's get started. Let me tell you something about genealogists that I think is deeply, painfully true. Most of us are much better at gathering evidence than we are at writing it up. We're hunters. We love the search, the discovery, the moment when a record surfaces that unlocks something that has been stuck for years. We can spend an entire weekend chasing one census household, reading microfilm until our eyes blur, and come away feeling completely alive. And then when the hunting is done, when we've got the evidence sitting right in front of us, we freeze. I know I do. There's a shoebox, metaphorically speaking, that almost every serious genealogist I've ever talked to has somewhere. It's the research project that is done, except for the writing. The evidence is there. The conclusion is solid. Three sources all pointing the same direction. Maybe four. But writing it down in a way that would hold up to scrutiny from someone who wasn't in the room with you, someone who wasn't there for the discovery, someone who needs to be convinced by the argument and not by your enthusiasm. That's the part we keep putting off. I think I know why, and it's not because we're lazy or disorganized. It's because writing a proof argument feels like standing up in court. It's not just, I think this is right. It's here's my case, here's the evidence, here's why the evidence supports the conclusion, and I am prepared to defend every sentence of it. That's a higher bar. It's supposed to be. And for a lot of us, that bar is intimidating enough that we'd rather keep doing research than start writing. But here's what I've come to believe. After more than a decade of serious genealogy work, the proof argument is not the end of the research. It's the part of the research that shows you what you still need to find. It's the structure that makes the gaps visible, the tool that forces you to be honest about what you actually know versus what you've assumed. And it doesn't have to be terrifying to write. Not anymore. Here's what the GPS actually asks for. Element five is deceptively simple in its wording. A soundly reasoned, coherently written conclusion, not a peer reviewed academic paper, not a legal brief, a clear, organized argument that explains what you found, what it means, and why you're confident in your conclusion. There are three parts to every proof argument. First, a statement of what you're trying to prove. Not a vague question, a specific claim. Quote Calvin Merritt born around eighteen forty two is the same Calvin Merritt who appears in the eighteen seventy Decatur County, Indiana Census as the head of household, and the same Calvin Merritt named in the eighteen sixty eight County Deed Records in quote. Specific. Second, the evidence. Everything you gathered in elements one through four of the GPS, the sources, what each one says, how they correlate, and how you resolve the places where they didn't agree. Every source gets its own honest accounting, including the ones that initially seem to contradict your conclusion before you work through the contradiction. Third, the analysis and reasoning. This is where you connect the dots for the reader. Why does this evidence, taken together, support the conclusion? Why should anyone trust it? This is not a place to hide uncertainty. It's a place to be transparent about what you know for certain and what you're inferring, and to make the argument for why the inference is sound. That's the whole structure. When you lay it out that plainly, it sounds manageable, doesn't it? Three labeled sections, a claim, the evidence, and the analysis. That is GPS element 5. And here's why I'm bringing AI into this. The structure is the easy part to understand in the abstract. The hard part is sitting down with a dozen browser tabs, a folder full of saved record images, a text document of research notes that only makes sense to you, and trying to assemble all of that into prose that flows, cites correctly, and makes a convincing argument without leaving out anything important. That's the translation problem. Getting what's in your head and in your sources onto the page in a form that another researcher could follow and trust. That's exactly the kind of organizational analytical drafting work that AI tools are genuinely good at. Now, before I go any further, let me say the thing that matters most here, because it's true in this context more than in almost any other. AI is your research assistant, not your researcher. That principle sits at the heart of today's whole workflow. The AI doesn't decide what the conclusion is. You do. The AI doesn't evaluate whether your evidence is sufficient. You do. What the AI does is help you get the argument that's already in your head onto the page clearly and coherently so that it can do what it needs to do, survive you, outlast the conversation, be understood by someone who never sat beside you in the archive. I think about it this way. My grandmother's generation of genealogists wrote letters to county clerks and kept handwritten files in accordion folders. My generation learned to search microfilm and build family search trees. This generation is learning to work with AI. The standard doesn't change. The tools do. For today's episode, I'm going to walk you through a real research tangle drawn from work on my own tree. Over the past year, I've been building a case around a land record discrepancy in a probate file that, combined with census data and a deed, finally resolves a question I've had about an ancestor for years. It touches on everything we've covered in this mini series. And today, I'm going to write the proof argument for it live using three tools. Notebook LM to organize my evidence, Claude to draft and logic check the argument, and Chat GPT to do a final plain language clarity review. Let me show you how it works. The first thing I needed was to get all of my evidence in one place where I could query it without losing track of where everything came from. That's the part of writing a proof argument where I usually start to feel buried. I've got a Sitsys image saved in one folder, I've got a D transcription in another. My notes on the probate file are in a document somewhere in my research folders. None of it's talking to each other. And when I sit down to write, I'm constantly switching back and forth between windows, which breaks my concentration and my train of thought. Forty five minutes in, I've reread the same census entry six times and written three sentences. This is exactly what Notebook LM, notebook LM.google.com was built for. Its whole identity is being a research assistant that answers questions grounded in the documents you upload with citations back to the specific passage. It doesn't browse the internet, it doesn't draw on its general training to fill in gaps. Every answer it gives you traces back to something you gave it. For writing a proof argument, that constraint is not a limitation. It's a feature and it's an important one. Here's what I did. I created a new notebook and uploaded everything relevant, my transcription of the 1870 census entry, a PDF of the deed image I'd obtained from the county recorder's office, my type notes from the probate file with page references, a short document summarizing the conflicting birthplace information I'd worked through in my earlier research, a county history excerpt that provided context about land ownership patterns in the area during the period, and a transcription of a second census entry from 1880 that confirmed the same individual in the same county a decade later. Five documents, one research note, six sources in total, all from my own research file. Then I typed a query into NoBook LM to start organizing the pieces. Here's exactly what I typed. Quote, based only on the sources I've uploaded, summarize what each source says about the identity of my research subject. For each source, know what it confirms, what it implies, and any gaps or inconsistencies you see. Include the specific passage or citation from the source for each point, end quote. What came back was remarkable. Nobook LM pulled a structured summary from each document, citing the exact passage each point came from. Because the prompt specifically asked it to flag gaps and inconsistencies, it did exactly that. It was careful. It didn't volunteer information I hadn't given it. Where my notes mentioned that the probate file referred to C period Merit without spelling out the full name, Notebook LM flagged it as an abbreviation requiring confirmation rather than assuming it matched my subject. That kind of precision matters enormously when you're building a proof argument. A credentialed reviewer will ask exactly that question. How do you know C period merit in the probate file is the same person as Calvin Merit in the deed? Notebook LM surfaces the question. Your research has to answer it. The citation by citation summary became the skeleton of my evidence section. Every point in the proof argument section section, the evidence presentation, now had a clear origin I could trace back to a primary source. I didn't have to hold it all in my head anymore. There was one moment in the notebook LM output that stopped me cold, and I mean that in the most useful possible way. As I reviewed the summary, I noticed it had flagged something I'd glossed over in my own notes. The 1870 census recorded my subject's occupation as farmer, but a second document in the set, an 1866 county tax record I'd included in my notes, listed him as a laborer. I've been trading those as roughly equivalent because, for my purposes, the identity mattered more than the occupation. Both documents placed him in the right county. The word farmer versus labor felt like a detail. Notebook LM surfaced it as an inconsistency and was right to do so. Not because the inconsistencies necessarily undermined my conclusion, but because a proof argument that doesn't acknowledge it is a weaker argument than one that explains it. I noted it to address in the draft, and I'll show you how that played out. That's your first takeaway from today. Before you write a single sentence of your proof argument, run your evidence through a grounded analysis tool that forces you to see what each source actually says, not what you remember it saying and not what you've assumed it says because you looked at it three months ago and your notes were brief. Notebook LM's free tier includes up to 50 sources per notebook and 50 chat queries per day, which as of June 2026 is more than enough capacity for most proof arguments. You don't need a subscription to use it for this. Alright, I had my organized evidence. Now I needed to draft. Drafting a proof argument from scratch is where most genealogists, including me, get genuinely stuck. The evidence is organized. The conclusion is clear. But transforming all of that into coherent prose with proper transitions, logical flow, and citations in the right place is hard. It requires holding the whole argument in your head while you write any one piece of it. And it requires a kind of structured thinking about the shape of the argument, not the genealogy, not the history, but the actual rhetorical structure of the case you're making. Claude is exceptionally good at exactly this kind of complex structured writing task. Its projects feature on Claude.ai means I could load my evidence summary from Notebook LM plus my own research notes directly into a project, and Claude would have all of that context available across multiple messages without me having to re-explain the background each time I asked a new question. That matters when you're working iteratively on a document, refining as you go. Here's how I set up the drafting prompt. I pasted in the summary Nobook LM had given me, along with my original notes, and then asked Claude to draft the argument in three clearly labeled sections. Quote, I am writing a genealogical proof argument following the standards of the Board for Certification of Genealogists. Using only the evidence summary I've provided below, please draft a proof argument in three clearly labeled sections. One, the statement to be proved, one specific, testable claim about the identity or relationship of the research subject. Two, evidence presentation, a source by source account of what each document says and implies with any inconsistencies or gaps noted explicitly rather than papered over. Three, analysis and reasoning. An explanation of how the evidence taken together supports the conclusion, including how many inconsistencies were resolved. Write in the first person as a genealogist presenting their case. Use clear, plain language. Do not add information not present in the summary I've provided. Flag any place where you believe additional evidence would strengthen the argument. Evidence summary, paste the notebook Elm output here, end quote. The draft Claude produced was strong. It organized the evidence exactly as I'd laid it out in the notebook Elm summary, and the statement to be proved was crisp and specific in a way that my own scattered notes hadn't been. Claude had done exactly what I'd asked, taken the pieces and assembled them into a structured argument. And it addressed the farmer labor inconsistency head on in the analysis and reasoning section exactly as I'd needed. The draft noted that laborer in eighteen sixty six and farmer in eighteen seventy are not necessarily contradictory. In mid nineteenth century Indiana, men who would later describe themselves as farmers often appear in earlier records as laborers while they were still accumulating land. The draft cited that the eighteen sixty eight deed showed a forty acre conveyance, suggesting that by that year the subject had moved from laboring on others' land to holding land of his own. That progression made the shift in occupational description coherent. That's not something I had written explicitly in my notes. But because I'd given Claude the full evidence set, it could reason across the sources and surface the explanation. I verified the reasoning against my original documents. It held. But something else Claude flagged surprised me just as much, and this is important. It made the argument's other weakest seem visible. In my evidence, I had three sources that all placed my subject in the same county at the same time period. But the deed and the census used slightly different spellings of the surname, and my original notes had treated that as a non issue because I'd been working with these documents long enough that the identification felt obvious to me. The variation was minor. I'd stopped noticing it. Claude's draft flagged the spelling variation in the analysis section with a note that read, essentially, quote, the difference in spelling between the deed and the census requires explanation for a reader unfamiliar with the case, end quote. It hadn't resolved the inconsistency. It had surfaced it, which is exactly what a careful reader or a credentials reviewer would do. And I realized something in that moment. I hadn't been careless. I'd been too close to the research. I knew the spelling variation was minor because I'd spent enough time with the nineteenth century County Clerk Records to know that name standardization was not a priority. Period Clerk spelled by ear. The same individual could appear under three different spellings and three different documents from the same courthouse. But the argument on the page didn't say any of that. The reader couldn't know what I knew. Claude had shown me precisely the gap I needed to fill. So I went back to my sources. I found two additional land records from the same county, same decade, same clerk's office, that showed the same individual using both spellings interchangeably across a span of eight years. I added those to my Nobook LM notebook, then I typed a follow-up query. Quote, does any source in my notebook provide additional evidence for how this individual's name was recorded across time, particularly any variation in spelling across documents? Nobook LM found the relevant passages in the two new records immediately, cited them specifically, and noted the pattern of alternating spellings. Then I went back to Claude with an addendum to the prompt. Quote, in the evidence presentations section of the draft, please add a paragraph that addresses the name spelling variation between the deed and the census. The additional evidence I'm providing below shows the same individual using both spellings and county records across an eight year period. Explain how this pattern resolves the apparent inconsistencies in what it tells us about record keeping practices of the era, end quote. Claude added the paragraph, and more importantly, it added the explanatory logic that the argument had been missing. In mid-19th century county clerk records in this region, name standardization was largely absent, and the pattern of variant spellings across contemporary documents for the same individual is itself a form of evidence rather than a contradiction. It demonstrates that the same clerk's office was recording the same person under both forms, which actually strengthens the identification. This is where element four of the GPS came fully into the picture. Element four is the resolution of conflicting evidence. We covered it in episode thirty five. The principle is that you don't sweep conflicts under the rug, you address them directly. And the way you address them is with evidence, additional square. Sources, contextual information about record keeping practices, or reasoning that explains why the apparent conflict doesn't actually undermine the conclusion. Now let me step back and name what's been happening here from a GPS standpoint because this is the capstone of the miniseries and I want to connect all four episodes explicitly. Element one, reasonably exhaustive research, is represented by the breadth of source types I'd gathered before I ever opened these tools. Census, deed, probate file, county history excerpt. We talked about casting that wide net in episode 38, and the additional land records I went back to find during the drafting process are exactly that principle in action. The drafting revealed a gap and I went back to fill it. Element three, analysis and correlation, happened in the notebook L impass. Seeing all my sources side by side, watching the tool identify what each one confirmed or complicated, is what we do with evidence when we're being rigorous. We collect it and then we put it in conversation with itself. Element four, resolving conflicting evidence was the work triggered by Claude's flag. Go back, find the additional sources, explain the pattern. We covered the reasoning framework for this in episode thirty five. And element five, the soundly reasoned, coherently written conclusion, is what Claude is helping me draft and sharpen right now. The AI did not do my thinking for me. It held the pin while I directed the argument. The research was mine. The conclusion is mine. The AI made it readable. Here's where I want to be precise about that golden rule, because this is the verification moment. Remember, AI is your research assistant, not your researcher. Claude's draft was a starting point, not a final product. Before I accepted any claim the draft made, I went back to my original source images and confirmed that the argument matched what the documents actually say, not what I might have half remembered or what Claude might have reconstructed from my notes. Every point in the evidence section gets checked against the primary source. Every citation gets verified. The AI wrote a coherent argument from the material I gave it. My job was to make sure the material I gave it was accurate and that the argument the AI constructed from it was faithful to the evidence. That's not a formality. That's the standard. Here's where the whole thing came together. I had a draft proof argument that was structurally sound, that addressed both inconsistencies honestly, and that incorporated the additional evidence I'd gotten back to find. The argument now had three clearly labeled sections, source citations throughout the evidence presentation, and an analysis section that walked through the reasoning step by step. But I had one more concern, and it's a specific one. After spending weeks deep in a research question, it is nearly impossible to read your own writing and gauge whether it's makes sense as someone who wasn't there. You fill in the blanks with your own knowledge. You skip steps you've internalized. You use terms you know and forget that your reader might not. This is a universal problem in technical writing, and genealogical proof arguments are no exception. This is where I brought in ChatGPT for a targeted task, plain language clarity review, not fact checking, not evaluation of the genealogical reasoning. The AI doesn't have access to my sources and I didn't ask it to evaluate the evidence. The task was narrower and more specific. Does this read clearly? Are there passages where I assume knowledge the reader doesn't have? Are there sentences that could be said in half the words without losing anything? Here's the prompt I used. Please review the proof argument below for clarity and readability. Assume the reader is an intelligent adult with a general interest in genealogy but no specialist knowledge of nineteenth century land and probate records. Flag any passages where the writing assumes context the reader might not have. Suggest places where technical terms could be briefly explained on first use. Do not evaluate the genealogical reasoning or suggest changes to the evidence. Focus only on whether the writing is clear, organized, and easy to follow. Return your suggestions as a list of specific edits, not a rewritten version, end quote. ChatGPT came back with eight specific suggestions. Most of them were small, a term used without defining it, a sentence doing double duty that should be two sentences, a paragraph that buried its main point in the third sentence when it should have led with it. Two of the suggestions were more substantive. I'd used the phrase indirect evidence twice in the analysis section without ever explaining what it meant, because by that point in the research, I'd stopped noticing it was terminology. To me, it was a plain description of the situation. To a reader who hasn't spent years thinking about the difference between direct, indirect, and negative evidence, it might as well be a legal term. ChatGPT flagged it both times. I added a brief parenthical explanation the first time the term appeared. The second suggestion was structural. My statement to be proved section was technically correct but rather flat. It opened with the claim exactly as a proof argument should. ChatGPT suggested that one additional sentence of context, setting up why this particular question mattered to the research as a whole, would help a reader understand what was at stake before they worked through the evidence. I thought about it, agreed that the dry claim first approach served the logic but not the reader, and added the sentence. Then I applied all eight edits. I read the revised draft aloud, the way a script eventually gets read in a recording booth, because reading aloud catches what reading silently misses. Sentences that look fine on the screen sometimes trip you in the air. Paragraphs that seem well paced on paper sometimes feel rushed when you say them. And when I finished reading, I felt something I hadn't expected. Not triumph. Something quieter. Relief. The kind that settles in when you realize you've done something properly. Not quickly, not approximately, actually properly. That's what element 5 is supposed to feel like. Not a formality you append to the research. The part of the research that makes the research matter. The whole process, from uploading my first document in Notebook LM to accepting the last edit from the ChatGPT Clarity Review, took one focused afternoon. That's it. One afternoon to write the proof argument I'd been putting off for two years. Your homework this week is this. Pick one research question from your own tree, any question where you have at least three sources pointing the same direction, and write a statement to be proved. One sentence specific, testable, something you'd be willing to defend in writing. That's the whole homework. If you finish that sentence and want to keep going, upload your sources to Notebook LM and see what they look like when they're organized. But the one sentence is enough to start. The sentence itself will tell you whether your evidence is ready for a proof argument or whether you need to do more research first. Either answer is useful. And for my Australian and UK listeners, the proof argument standard I describe today is recognized across the English-speaking genealogical world, and your research benefits from it exactly as ours does. For Australian researchers, the Society of Australian Genealogists or G.au, provides resources on research methodology and standards that align with what we covered today. UK researchers will find the Society of Genealogists.org.uk and its library an invaluable resource for understanding how written proof standards apply to British research, including the kinds of source analysis that works with Paris Chess Documents, the 1921 census, and the 1939 register. The BCG framework travels your archives, your records, your argument. Thank you so much for listening to Ancestors and Algorithms. If you've enjoyed this episode, please leave a review wherever you listen to podcasts. And if you know a fellow genealogist who could benefit from what we covered today, share this episode with them. That is the best way to help our community grow. For my Patreon members, the companion guide for this episode is waiting in your library. It includes 12 advanced prompts for every stage of the proof argument workflow, a GPS research checklist you can use for your own cases, and a multi-step workflow that takes you from a raw pile of sources all the way to a finished argument ready for peer review. And if you've been thinking about joining, I host a monthly live Q ⁇ A sessions on YouTube exclusively for members. Head to ancestorsandai.com to learn more. For everything you need, including every episode, our private community, companion guides, and the research lab, head over to ancestorsandai.com. It's all right there waiting for you. I'm your host Brian, and I will see you next week for another journey into the past powered by the future. Until then, happy researching.