How I Create

🧠

How I Create. I research and write Possibility Pages with Claude, an AI model, as a research and drafting partner. I choose the questions, sources, and arguments; Claude helps me find and read sources, check facts, and draft. I verify quotes against the original documents and recordings, and every claim is cited so you can check it too. This is an emerging method, and I'm documenting the process as a researcher concerned with the ethical use of AI in scholarship. The views and any mistakes are mine.

Who does what

I choose the questions. Every Possibility Page starts with something I can't stop orbiting: a vote, a photograph, a line from a podcast, a name in a county record. The question decides everything after it. I choose the sources and the argument, and I decide where I land.

Claude, an AI model made by Anthropic, is my research and drafting partner. It helps me find sources, read long documents, and pull out what bears on the question. It drafts paragraphs, especially the sourced, reporting-heavy ones. It checks dates, spellings, and citations, and it builds the reference list. When I ask, it also argues with me about whether a claim holds.

What I don't hand off: the questions, the stance, and the "I." When a sentence speaks for me (what I think, what I remember, where I'm still unsure), I write it, or I rewrite it until it's mine.

Drafting is the use most academics are wary of. In a 2026 Chronicle survey of 460 academics, only 11 percent of those open to AI use found drafting acceptable, compared with 70 percent for finding previous writing on a topic (Vance, 2026). I’d rather tell you than leave you guessing.

The tools

This isn't a free-tier experiment. I've invested hundreds of hours and my own money into paid professional AI models, because the method takes sustained practice to do well. No company pays me to use or recommend them.

The shoulders I stand on

I exiled myself from the academy, not from the scholarship. I brought my research library with me, along with more than a decade of teaching, research, and invited presentations and publications in higher education. I trained as a researcher at the methodological edge: participatory action research in my master's program, arts-based research in my Ph.D.

Every book on my shelves and every article I've read, I read the long way. That slow work is creative friction meets sweat equity meets neuroplasticity, and I wouldn't trade it. It is also what lets me judge what an AI hands me back. You can't check a reading you never learned to do yourself.

I'm a bridge here, too. I learned scholarship as large language models were just arriving; I had Grammarly in my Ph.D. program, not Claude Opus. We don't ask mathematicians to work without calculators, and I don't think we should give up too early on the ethical use of AI in scholarship, including in the humanities, and not just for processing large quantitative datasets. Unlike a calculator, though, a language model can be confidently wrong, which is why the checking below isn't optional.

Citation is how I show whose shoulders I'm standing on. Even that phrase has a lineage: Newton borrowed it, and it goes back to Bernard of Chartres in the twelfth century. Some of those shoulders were never asked, though. Scholarship has long stood on stolen ground: ideas taken without credit, land taken from the people who lived on it, and the wisdom that held the two together, knowledge of a place that grew from living in it. These models carry the pattern forward. They were trained on writing whose authors didn't consent, some of it pirated. A model can't tell you whose work it learned from. I can tell you whose work I used, so every claim here points back to a person. That isn't a formality; it's what rigor means.

Claude also helps me look around corners and find voices I would have otherwise missed. Machine processing meets human curiosity, and that leaves more of my own bandwidth for the part that matters: the dialogue between sources gathered from many timelines, and a pause long enough to make sense of it. It's like a fantasy dinner party where the guests have time-traveled in to meet one another. The machine has blind spots of its own, so I still check who's missing from the guest list.

How a page gets made

The question comes first. The writing decides what gets read, not the other way around. Each section names the question it must answer, and only the sources that answer it get a close reading. That keeps the research from swallowing the writing.
Reading. Claude reads what it can reach and labels how it knows what it knows: whether it read the text itself or only saw a title or a summary. A summary is never passed off as a reading. When a source is paywalled or blocked, I get the document and read it myself, or we leave it out. We don't route around a block.
Drafting. Drafts carry visible tags on anything not yet checked: verify, you to supply, decision open. A gap stays a gap until someone closes it. An ugly draft with visible gaps beats a clean draft with invisible guesses.
Checking, then voice. Nothing publishes with an open tag. Then I go back through the draft for the lines that are mine to say.

How I check

Quotes against the originals. I check quotations against the recording, transcript, or document they came from, with timestamps. For Part III, I checked every quote from Pamela Colloff's podcast interviews against the transcripts. Four had drifted: a host's framing credited to his guest, a remark about one case stretched into a general claim, a detail overstated, and "one turning point" turned into "the first." Each was fixed before publication. That is why I do the checking.
A confabulation audit before publishing. Language models can confabulate: produce a quote, a date, a title, or a citation that sounds right and isn't. So before a page goes out, every quotation, name, number, date, and reference is traced back to a source someone actually read. Anything that can't be traced gets cut or is marked as uncertain in the text. The audit catches the AI's errors too. While Part III was being drafted, our working notes "corrected" Pamela Colloff's book title to To Catch a Devil. The publisher's listing says Catch the Devil, and that's the title that ran.
Every claim cited. Each page ends with a full reference list in APA style, with links in the web version, so you can follow any claim back to its source.
Independent sources. Three articles that repeat the same original source count as one source, not three. Corroboration has to come from somewhere else.

Images

Images in the body of a post are not AI-generated. They are photographs, documents, and public-domain scans, such as newspaper pages from the Library of Congress, and their sources are credited.

Site photos and branding are made with Claude, in Canva, or both. Some feature images start from the archive: the one for Part III began as three 1936 newspaper pages from the Library of Congress, which Claude cropped and toned with code. The headlines are real; only the paper's texture was added.

Why I disclose

In October 2026, an AI detector flagged part of a draft as likely machine-written. I could have reworded until the score dropped. Instead, I'm telling you how I work. A detector measures how a sentence sounds. It can't tell you whether a quote is accurate, whether a source exists, or whether anyone stands behind the argument. Those are the things I can show you, so those are the things I show you.

I’m in the minority here, again: 65 percent of academics in a 2026 Chronicle survey had never used AI to prepare writing for publication (Vance, 2026).

A byline has never meant that no one else helped. It means someone answers for the work. Here, that's me.

Still unsettled

I call this method emerging because it is. Questions I'm still working through:

The environmental cost of the computing behind it.
Whose writing trained these models, and on what terms.
What the method does to my own reading, attention and patience.
Where a disclosure ends and an endorsement begins.
Who can afford tools like these, and what that means for whose research gets done.

The environmental cost isn’t abstract here. Anthropic, the company behind Claude, has signed a $19 billion, 20-year lease on a data center campus in Hancock County, just east of Owensboro, and the same developer plans a campus near Ashland that could draw up to two gigawatts (Hansen, 2026).

Corrections

If you find an error, please tell me through the corrections form. You can submit it anonymously; leave your name or email only if you'd like a reply or credit. Corrections are made on the record and noted at the end of the post with the date, never silently patched.

The short version