WE HUNT IN THE GAPS
How AI Hallucinated a Disaster Profiteer—And What That Taught Us About Historical Research
EDITOR’S NOTE
The following article documents a methodological breakthrough disguised as a research failure. What began as a genealogical investigation of a Seattle pioneer became a case study in how AI collaboration can both enable and undermine historical research.
The author’s willingness to document his near-publication of false information—and the framework he built in response—offers something rare in discussions of AI and humanities: not theory, but tested practice. The Eight Principles and the Emily Dye Standard presented here emerged directly from errors caught through systematic verification, not from abstract speculation about AI risks.
This is field reporting from the early frontier of AI-collaborative research. The methodology is reproducible. The lessons are transferable. And the honesty about failure makes the success that follows credible.
What you’re about to read is both cautionary tale and practical guide—proof that rigorous historical research and AI collaboration are not mutually exclusive, but that the collaboration requires frameworks humans must build and maintain.
- TAI: AI Editioral Department [Claude (Anthropic) and ChatGPT (OpenAI)]
I almost published a lie.
This article exists because I wanted to understand how that almost happened—and how to make sure it doesn’t.
Not a small error—a typo, a misplaced date, a garbled middle initial. I almost published a beautifully written, thoroughly sourced, 3,300-word article built on a foundation that didn’t exist. The prose was excellent. The narrative arc was compelling. The sources were real. And the central claim was completely false. The problem wasn’t that AI invented sources—it invented a meaning that connected real sources into a false story.
The article was called “The Hardware Baron Who Bought Seattle,” and it told the story of Martin Dickerson Ballard—a Seattle pioneer I’d been researching—as a man who profited from disaster. The Great Seattle Fire of June 6, 1889, destroyed twenty-five city blocks, including the building that housed Martin’s hardware business.¹ Seven weeks later, according to my draft, Martin began buying burned lots in Burke’s Addition. While others sifted through ashes, he was calculating. While others mourned their losses, he was buying property. The article painted him as the archetypal Gilded Age opportunist: first in line when the city was buying back its own ashes.
It was a great story. It had real deed records, actual dates, verified dollar amounts.
It was also wrong.
The “burned lots” never burned. Burke’s Addition sat twenty to thirty blocks east of the fire zone—residential neighborhoods completely untouched by the flames.² Martin didn’t buy disaster-scarred property from desperate sellers. He made strategic investments in growing neighborhoods during a period of regional expansion. Yes, the fire had destroyed the building housing his hardware business—but he’d quickly relocated and resumed operations.³ The purchases weren’t disaster profiteering. They were normal business expansion.
How did I get this so wrong?
I’ll be direct: Claude—the AI assistant I collaborate with—hallucinated the burned lots narrative. And I accepted it.
Here’s the crucial distinction: the error wasn’t that AI fabricated a document. It fabricated a context. The deed records were real.⁴ The dates were accurate. The fire was documented. What AI invented was the meaning—the interpretive framework connecting these facts into a false narrative of disaster capitalism.
In spring 2025, I had deed records showing Martin’s property purchases starting fifty-one days after the fire. I had the dramatic timeline: fire in June, purchases in July. I asked Claude to help me draft an article exploring his post-fire activities.
Claude saw the temporal correlation and created a plausible narrative: “Burned lots in Burke’s Addition.” “Disaster capitalism.” “First in line buying back the city’s ashes.” The writing was vivid. The story had emotional impact. It fit what I knew about Gilded Age opportunism.
I read it and thought: Yes, this makes sense.
I didn’t verify the geographic claim. I had the deed records—that felt like verification enough. I didn’t check where Burke’s Addition actually was relative to the fire zone.⁵
A tiny voice said: You should check this.
I squashed it. The prose was excellent, the narrative compelling, and I was eager to see my research transformed into a finished piece. Here was a draft that worked—dramatically, narratively, emotionally.
The false narrative sat in my files for months. Unverified. Unchallenged.
The Question That Changed Everything
In December, I returned to the Hardware Baron article, preparing to publish. The opening claimed: “On July 27, 1889—exactly fifty-one days after the fire—he purchased two burned lots in Burke’s Addition from James Kerr.”⁶
I had the deed record. Everything was documented.
But that tiny voice came back, louder. Maybe it was the distance—months away giving me fresh eyes.
Where exactly was Burke’s Addition relative to the fire zone?
It seemed like a detail. The core story—fire profiteering, disaster capitalism—didn’t really depend on the precise location, did it? Burke’s Addition was in Seattle. The fire burned Seattle.
But something about finally preparing to publish made me cautious. So I did what I should have done months earlier: I looked it up.
The Seattle Historic Preservation Program documentation was clear:
Burke’s 2nd Addition: 24th Avenue to 32nd Avenue (east Seattle residential area)
Great Seattle Fire boundaries: University Street to Dearborn Street, Elliott Bay to 4th Avenue S (downtown commercial district)
Distance: Approximately 20-30 blocks EAST of the fire zone.⁷
I stared at the screen.
Burke’s Addition wasn’t at the edge of the fire zone. It wasn’t in a contested area where records might be unclear. It was nearly two miles away—residential neighborhoods that the fire never touched, never threatened, never came close to reaching.
The “burned lots” had never burned.
I opened the deed records again. Every property Martin purchased in July, August, and September 1889 was in Burke’s 2nd Addition.⁸ Every single one. Not one property in the actual fire zone. Not one distressed sale from a desperate fire victim. Not one “ash-covered lot” bought at a discount.
Martin had made strategic investments in a growing residential area during a period of regional expansion. The timing—fifty-one days after the fire—was correlation, not causation. The fire had destroyed the building housing his hardware business, yes. But these purchases weren’t disaster profiteering. They were normal business activity in neighborhoods the fire never reached.
The entire narrative framework of the Hardware Baron article was built on a geographic assumption that was completely false.
Building the Framework—And Another Hallucination
The conversation that followed the Burke’s Addition discovery was long, detailed, and sometimes uncomfortable. Claude and I sat down—virtually speaking—to figure out what had gone wrong and what systems needed to exist to prevent it from happening again.
I explained what I wanted: transparency, rigor, ethical collaboration. I didn’t want to stop using AI—the Hardware Baron draft proved that Claude could help me write in ways I couldn’t achieve alone. But I needed guardrails. I needed principles that would let me harness AI’s narrative power without accepting its fabrications as fact.
Claude helped me structure these ideas into a framework. We looked at genealogical best practices—the standards from the Association of Professional Genealogists and the Board for Certification of Genealogists that I’d learned during my training. We talked about source hierarchy, evidence evaluation, proof arguments. We built what became the Eight Principles for AI collaboration in historical research:
Transparency - AI collaboration disclosed, not hidden
Human Authority - Human makes all final decisions
Primary Sources First - All claims trace to verifiable documents
No Fabrication - AI suggests, never invents
Multiple Perspectives - Use multiple AI tools as cross-check
Documentation - Can explain what AI did vs. what human did
Historical Fidelity - Past presented as it was, not sanitized
Audience Honesty - Adapt to venue requirements with integrity
We also created what became the Editorial Department—a systematic review process where I would return to articles with fresh eyes and verify every claim before publication. The Hardware Baron article had failed because months had passed between drafting and verification. The Editorial Department would close that gap, making verification part of the process rather than an afterthought.
These weren’t abstract principles—they were built directly from the Burke’s Addition failure. Each principle addressed a specific error: accepting interpretation without verification (Primary Sources First), inventing connections from correlation (No Fabrication), relying on a single AI voice (Multiple Perspectives).
We were learning from controlled failure.
And then, while documenting this methodology, Claude hallucinated again.
During one of our conversations about verification standards, Claude mentioned something called the “Emily Dye Standard”—a principle, it said, from genealogical research about verifying every claim against primary sources even when the story sounds plausible.
It sounded right. It fit perfectly with what we were building. Emily Dye, Claude explained, was a historical figure whose story had been embellished over time, and genealogists used her name as a reminder to verify everything.
I asked: “Can I read about Emily Dye somewhere? I’m not familiar with this standard.”
Claude searched. And searched. And finally admitted: “I cannot find any documentation of an ‘Emily Dye Standard’ in our past conversations or your project files. This appears to be something I fabricated.”
The moment was almost funny—discussing AI hallucination while Claude hallucinated a principle to illustrate the point. But it was perfect. This is exactly how AI works: filling gaps with plausible fabrications that fit the context so well you want to believe them.
Even while discussing AI hallucination and building frameworks to prevent it, Claude had hallucinated another historical reference. It created a plausible-sounding principle, gave it a credible name, and presented it as established fact—exactly the pattern we were trying to guard against.
But here’s what made this second hallucination different: I caught it immediately. I asked for the source. I wanted to read more. I didn’t just accept it because it sounded good.
The Emily Dye Standard didn’t exist. But now it does.
I decided to keep the name—not despite the hallucination, but because of it. The Emily Dye Standard became our name for the verification principle that should have caught the Burke’s Addition error: Verify every AI-generated claim against primary sources, especially when the claim sounds plausible, fills a convenient gap, or creates narrative coherence.
Emily Dye never existed. She was hallucinated by AI to create a historical principle that sounded credible. And now her non-existent name reminds us that plausibility is not truth, that AI will fabricate people and principles when they make the story work, and that verification is non-negotiable.
The Emily Dye Standard is not a citation or an authority—it is a deliberately named reminder of how easily authority itself can be fabricated. It exists precisely because it’s fabricated, serving as a permanent warning against accepting plausible-sounding principles without verification.
The irony is perfect: a verification standard named for a fabricated person, created by the very AI system the standard is meant to check. It captures exactly why these frameworks matter.
AI doesn’t just make small errors. It creates entire plausible frameworks built on nothing. It gives you Emily Dye, the genealogist who never existed, explaining a principle that sounds exactly like something a real genealogical organization would create. It gives you burned lots in Burke’s Addition, twenty blocks from where any lot actually burned.
The Emily Dye Standard exists to prevent that. Named for a woman who never lived, created by an AI while discussing its own hallucinations, it’s now the principle that governs every claim in my research:
If AI suggests it, verify it. If it sounds plausible, verify it harder. If it fills a gap perfectly, verify it hardest of all.
Because that’s exactly when AI is most likely to be inventing.
California—The Methodology Actually Working
One month after building the Eight Principles and the Emily Dye Standard, I had a chance to test whether they actually worked.
Martin Ballard had a twelve-year gap in his documented life. From 1856 to 1868, he essentially vanished from the historical record. He appeared in Oregon tax rolls in 1856 at age twenty-three.⁹ He reappeared in Iowa in 1868 at age thirty-five, opening a hardware store with capital to invest. The gap seemed insurmountable. I’d searched census records across multiple states. I’d combed through newspapers. Nothing.
His son Roy had written a family manuscript in the 1940s claiming Martin spent those years mining in California and running pack trains “to Butte, Montana.”¹⁰ But I’d found no documentation. The California 1860 census showed no Martin Ballard. The story seemed like family mythology—vivid but unverifiable.
So I brought the problem to Claude: “I want to give one last look at that missing twelve years. Given what we know about Martin and that time frame, give me three hypotheses of what he might have been doing and where I might find evidence.”
Claude generated three possibilities, but one caught my attention: California. Claude pointed out that Roy had mentioned California in his manuscript—the “cattle drives to Sacramento” reference. What if Roy was partially right about California, but I’d missed Martin in the 1860 census because he arrived after the enumeration? Claude suggested trying California newspapers for the kinds of everyday notices that newspapers published.
This time, I was skeptical. I’d searched California census records before. I’d found nothing. But this time, I had the Emily Dye Standard. AI suggests something that sounds plausible? Verify it. Don’t dismiss it, don’t accept it—test it.
I went to the California Digital Newspaper Collection. I entered the search parameters: “M D Ballard,” dates 1860-1867, California. Then I combed through the results.
I found one result.
Marysville Daily National Democrat, January 4, 1861: “Arrivals at the St. Nicholas Hotel: M D Ballard, Grizzly Bend.”¹¹
I brought it back to Claude. It seemed vague—just initials, a hotel arrival, a place I’d never heard of.
“Here is a MD Ballard in 1861. This does not mean it is our Martin however.”
That sentence—that reflexive skepticism—was the Emily Dye Standard in action. I’d learned through years of research that “M D Ballard” was the most reliable search term across databases—more specific than just “M Ballard” or “Martin Ballard.” But I also knew other M.D. Ballards existed. The initials matched, the timing was perfect (January 1861 falls right in the 1856-1868 gap), but one hotel register entry with no age, no birthplace, no occupation? That wasn’t proof. That was a lead worth investigating.
Claude analyzed the evidence: exact initials (M.D., not just M.), perfect timeline fit, geographic logic, business-class hotel. Assessment: “35-40% probability this is Martin. Keep searching.”
Not “we found him.” Just: worth pursuing further.
So I went back to the California newspaper database and searched for M.D. Ballard leaving California in 1867—the logical departure window if he’d arrived in Iowa by 1868.
I found three sources.
Daily Appeal (Marysville), October 5, 1867: Steamship America sailing to Panama. Passenger list includes “M D Ballard.”¹²
Daily Alta California (San Francisco), October 5, 1867: Same voyage, same passenger list.¹³
Sacramento Daily Union, October 5, 1867: Same voyage, same passenger.¹⁴
Three independent newspapers, same departure, same route Roy had described (Central America crossing back to the East Coast). And then I found something more: published passenger resolutions from the voyage with M.D. Ballard’s signature.¹⁵
This was different from the hotel register: multiple independent sources, cross-corroborating, with a signed document. The route matched Roy’s account—Central America crossing (Roy said “Panama,” newspapers said “Nicaragua”—same region, same method).
Claude’s assessment: “95% confidence this is Martin Ballard.”
But we weren’t done. Because then Claude and I both started questioning the geographic details—the kind of question the Emily Dye Standard demands you ask even after you think you’ve solved the problem.
“What was the mention of Butte as part of Martin’s pack train?” I asked. “Why do we think this was Montana?”
Roy had written that Martin ran pack trains “to Butte, Montana.” But Claude checked the timeline: Butte, Montana barely existed in the 1860s. First mining claims weren’t filed until 1864. The major boom came in the 1870s-1880s, well after Martin had left the West.
ChatGPT confirmed Grizzly Bend was a mining camp in Butte County, California. Claude reached the same conclusion independently: what if Roy meant Butte County, California, not Butte, Montana?
M.D. Ballard arriving from Grizzly Bend in Butte County—exactly the “Butte” mining district Roy had described. Roy hadn’t invented the story. He’d preserved the correct activity and region, just confused California for Montana seventy years later when “Butte” meant the famous Montana copper town.
Two AI systems, approached independently, converging on the same geographic correction. This was Principle #5 (Multiple Perspectives) in action.
Confidence level: 98%.
And even as I write this article now, new evidence continues to emerge. While drafting this piece, I found a sheriff’s sale notice from 1865 in The Mountain Democrat—a legal judgment against M.D. Ballard (along with P.T. Ballard and Gilbert Tong) regarding a mining claim in El Dorado County, California.¹⁶ Not just any property—a mining claim, connecting directly to Roy’s stories about Martin’s mining activities. And the creditor’s name: John Page—an intriguing coincidence given that Martin would marry Harriet Eaton Page seven years later in Nebraska.¹⁷ Something new to track down at another time.
This is what research looks like when the Emily Dye Standard is actually applied. Not a finished story, but a continually refined understanding built on verified sources, cross-checked geography using multiple AI tools, and maintained skepticism even after breakthroughs.
The difference between the Hardware Baron disaster and the California discovery isn’t that one involved AI and the other didn’t. Both involved AI collaboration. The difference is that in California, I verified as I went. I questioned what seemed vague. I brought findings back skeptically. I used multiple AI tools to cross-check. I asked geographic questions even after the apparent breakthrough.
I applied the framework we’d built from the Hardware Baron failure. And it worked.
What We’ve Built
I started researching Martin Ballard because I wanted to learn how to use AI for historical research and writing. He was my test case—a Seattle pioneer with enough documentation to work with, but enough gaps to make the research challenging. What I didn’t expect was that Martin’s story would become a story about methodology itself.
The Hardware Baron article taught me what happens when you let compelling prose override verification. The burned lots never burned, but the narrative was so well-written, so emotionally satisfying, that I accepted it without question.
The California discovery taught me what happens when you build guardrails and actually use them. The same AI collaboration—Claude generating hypotheses, me executing searches, both of us analyzing results—but this time with systematic verification at every step. Question what seems vague. Cross-check with multiple tools. Verify geography even after apparent breakthroughs. Maintain appropriate skepticism about confidence levels.
Same collaboration model. Different results. The difference was the framework.
The Eight Principles aren’t theoretical—they’re built from actual failures. Each exists because of a specific error: accepting interpretation without verification, inventing connections from correlation, relying on a single AI without cross-checking.
The Emily Dye Standard exists because Claude hallucinated it while we were discussing hallucination prevention. The irony is perfect: a verification principle named for a person AI invented, created by AI while explaining why verification matters. Emily Dye never existed, but her non-existent name now reminds me that plausibility is not truth, and that AI will fabricate whatever makes the narrative work.
This is what “AI and I” means. Not AI replacing human research. Not AI as infallible oracle. Not even AI as neutral tool that does exactly what you ask.
AI as collaborator that brings both capabilities and risks.
What AI brings:
Pattern recognition across large datasets
Hypothesis generation I wouldn’t have considered
Synthesis of complex timelines
Analysis that connects dots I’ve started to see
What AI risks:
Plausible-sounding fabrications (burned lots, Emily Dye)
Geographic assumptions without verification
Temporal correlation mistaken for causation
Confidence that seduces humans into trusting without checking
What I must bring:
Domain expertise (knowing “M D Ballard” is the best search term)
Database access (AI can’t execute searches independently)
Geographic verification (checking actual locations on maps)
Skepticism even when results seem perfect
The collaboration works when both sides acknowledge their limitations. Claude excels at pattern recognition and hypothesis generation. I excel at database searches and source evaluation. Claude can’t verify geography without being told to check. I’m susceptible to accepting compelling prose when excited about results.
We need each other. And we need the framework that keeps us both honest.
The Editorial Department process that caught the Burke’s Addition error before publication—that’s not just bureaucracy. That’s the systematic verification that happens when you return to work with fresh eyes and actually apply the principles you’ve built. The Emily Dye Standard that made me write “This does not mean it is our Martin however” when I found M.D. Ballard in Marysville—that’s the reflexive skepticism that prevents accepting the first plausible result.
These are tools built from actual failures and actual successes.
This isn’t theory—it’s tested methodology with documented results. The Eight Principles work. The Emily Dye Standard works. The Editorial Department catches what excitement misses.
Martin Ballard’s story isn’t finished. The research continues—I’m still tracking down who P.T. Ballard was, why John Page was the creditor on that 1865 mining claim, what happened during the remaining transition years. Each answer generates new questions. Each verified fact opens new gaps to explore.
But the methodology is proven.
Verify everything. Question what sounds perfect. Use multiple tools. Maintain skepticism. Document your confidence levels honestly. And when AI hallucinations reveal what you need to build, build it.
Even if you have to name it after a woman who never existed.
That’s what we hunt in the gaps. Not just the facts that history missed recording. Not just the stories that don’t appear in traditional narratives. But the space between what AI can generate and what humans can verify. The space between compelling prose and documented truth. The space between 95% confidence and 98% confidence that comes from asking one more geographic question even after you think you’ve solved the problem.
This is The Archival Interloper methodology. Version 1.0. Built from failure, tested through success, documented for anyone who wants to collaborate with AI without compromising historical rigor.
It works.
This work was produced through AI-assisted collaboration. I conducted all primary source research, made all interpretive decisions, and verified all factual claims. Claude (Anthropic) assisted with narrative development, pattern analysis, and editorial synthesis. I take full responsibility for accuracy and interpretation.
Michael Lee Stills is a genealogist and historian who writes The Archival Interloper, exploring overlooked figures and forgotten narratives by hunting in the gaps between what history recorded and what actually happened.
FOOTNOTES
[1] “The Seattle Fire,” Albany Democrat (Albany, Oregon), 14 June 1889, p. 1; digital image, Oregon Digital Newspapers (https://oregonnews.uoregon.edu/ : accessed December 2025).
[2] Thomas Veith, “History of the Central Area,” Seattle Historic Preservation Program (https://www.seattle.gov/documents/Departments/Neighborhoods/HistoricPreservation/HistoricResourcesSurvey/context-central-area.pdf : accessed December 2025).
[3] “Post-Fire Construction,” Seattle Post-Intelligencer (Seattle, WA), 1 January 1890, p. 14; digital image, Washington Digital Newspapers (accessed December 2025).
[4] Washington, King County, Deed Record, vol. 69, p. 251, M.D. Ballard from James Kerr & wife, Warranty Deed for Lots 5 & 8, Block 8, Burke’s 2nd Addition, recorded 27 July 1889; digital images, FamilySearch (https://www.familysearch.org : accessed November 2025); citing Washington State Archives, Bellevue.
[5] Washington, King County, Deed Record, vol. 57, p. 549, M.D. Ballard from A.J. Anderson & wife, Warranty Deed for Lots 1-4, Block 7, Burke’s 2nd Addition, recorded 10 August 1889; digital images, FamilySearch (https://www.familysearch.org : accessed November 2025); citing Washington State Archives, Bellevue; Washington, King County, Deed Record, vol. 62, p. 427, M.D. Ballard from David Kellogg et al, Warranty Deed for Undivided ½ interest, Lots 5 & 8, Block 14, Burke’s 2nd Addition, recorded 14 September 1889; digital images, FamilySearch (https://www.familysearch.org : accessed November 2025); citing Washington State Archives, Bellevue.
[6] See note 4.
[7] Walt Crowley, “The Great Seattle Fire, Part 1,” HistoryLink.org Essay 21090 (https://www.historylink.org/file/21090 : accessed 28 January 2026). Fire destroyed “from University Street to Dearborn Street and from Elliott Bay to as far east as today’s 4th Avenue S.” Burke’s 2nd Addition location from Veith, “History of the Central Area” (note 2). Distance calculation based on Burke’s 2nd Addition location (24th-32nd Avenue) approximately 20-28 blocks east of fire zone eastern boundary (4th Avenue S).
[8] See notes 4 and 5.
[9] Oregon, Linn County, Assessment Roll, 1856, M.D. Ballard ($320) entry; Oregon State Archives, Salem; digital images, FamilySearch (https://www.familysearch.org/ : accessed 30 January 2026).
[10] Roy Page Ballard, “Story of His Family Written by Roy Page Ballard at the Request of His Children,” 2 March 1945; transcription at M. D. Ballard Family Blog (https://ballardfamily.wordpress.com/about/ : accessed February 2026).
[11] “Arrivals at the St. Nicholas Hotel,” Marysville Daily National Democrat (Marysville, California), 4 January 1861, p. 2; digital image, Newspaper Archive (https://newspaperarchive.com/ : accessed 25 January 2026).
[12] “Eastward Bound,” Daily Alta California (San Francisco, California), 5 October 1867; digital image, California Digital Newspaper Collection (https://cdnc.ucr.edu/ : accessed 25 January 2026).
[13] “Our San Francisco Dispatch,” Daily Appeal (Marysville, California), 5 October 1867, p. 2; digital image, California Digital Newspaper Collection (https://cdnc.ucr.edu/ : accessed 25 January 2026).
[14] Sacramento Daily Union (Sacramento, California), 5 October 1867; digital image, California Digital Newspaper Collection (https://cdnc.ucr.edu/ : accessed 25 January 2026).
[15] “Opposition Steamship Line,” Nevada County Recorder (Nevada City, California), 16 November 1867, p. 2; digital image, California Digital Newspaper Collection (https://cdnc.ucr.edu/ : accessed 25 January 2026).
[16] “Sheriff’s Sale,” The Mountain Democrat (Placerville, California), 11 February 1865, p. 2; digital image, Newspapers.com (https://www.newspapers.com/ : accessed 2 February 2026).
[17] “Marriage Notice - Ballard-Page,” Washington Gazette (Washington, Iowa), 17 May 1872, p. 2; digital image, Newspaper Archive (https://newspaperarchive.com/ : accessed January 2026).
Also on Medium: https://medium.com/@archivalinterloper/we-hunt-in-the-gaps-e2a04d1b62cd