On AI

This site was produced with an AI collaborator working under human direction. That is exactly the kind of thing a careful reader should be suspicious of, so this page sets out what the collaboration is, how it is disciplined, and why the work can be checked rather than taken on trust. The site’s frameworks, safeguards, and constraints are described on the About page; this page is about the tool.

How This Site Was Built

The collaboration was substantive enough that calling it "research assistance" would misrepresent it. This has been more like digging a trench with a backhoe than with a spoon: it works at a scale the spoon cannot, the accuracy the job needs is the same either way, and the operator still has to know what they’re doing.

Working with AI is not the same as accepting what it produces. The work throughout has been argumentative in the constructive sense: I have challenged Claude’s outputs, pushed back on characterisations, caught errors, and insisted on revisions. On a number of occasions Claude was wrong, about facts I could check against my own library, about how a primary source had been described, about the logic of an argument. The Gilvary dating question is a case in point: it was my reading of the actual book that caught a misattribution that had propagated across several pages. The text that appears on this site is what survived that process of challenge and correction, not what Claude first offered.

There is a further reason the AI is not optional, and it is the most honest thing I can say about how the site is kept up. I do not edit these pages by hand; I tell the AI what to change, and it carries the change wherever it reaches. A single reader’s correction is rarely a single edit, a recent factual error a correspondent caught in one tooltip had in fact propagated to seven places on the site, of which the reader had found one. Running the other six down by hand would have cost most of a morning, and a site this size is built from hundreds of such passes; no one person has that bandwidth, which is the honest reason it exists at all. I value the corrections readers send above almost any other input, but I rely on them to have checked what they caught, and on the AI to carry each correction to every place it touches, and either can fail. The safeguard is not a promise that they will not; it is that every correction is recorded in the Errata, in public, where the next reader can check it.

On Working With an AI

The reflex is to treat the human as simply the better researcher and the AI as a lesser substitute for one, when in fact the two are better at different things rather than better and worse at the same one. An AI has breadth and recall no single person can match, tireless consistency in running the same check across a large site, and no ego invested in a prior conclusion. A human researcher has what the machine structurally lacks: contact with reality (the ability to generate new primary evidence rather than only recombine existing text), a felt sense of the edge of one’s own knowledge, where the machine will guess as fluently as it knows; and judgment about which questions and which anomalies actually matter. The errors that follow are different in kind: a person’s tend to be grounded and local, an AI’s ungrounded and confident. That is precisely why the pairing can be worth more than either alone (each is placed to catch what the other misses), provided the human stays the one holding the work to the evidence.

A harder form of the objection is not that errors propagate but that AI-assisted writing is slop: fluent, confident prose with nothing underneath it, no different in kind from the plausible paragraph a chatbot will produce about a book it has ingested but no one has actually read. An AI does write just as smoothly when it is wrong as when it is right, so smoothness is evidence of nothing, which is exactly why nothing here is asked to rest on it. The test is the same for a sentence I wrote and a sentence Claude drafted, does it survive checking against a primary source, and a fluent claim that fails that check is simply an error, handled the way every other error is: corrected, and logged in the Errata. What disciplines the writing is not the tool but the requirement that every claim the argument depends on point to a document a reader can pull. What keeps it from slop is provenance: a traceable line from each claim back to its source, whoever or whatever drafted the sentence.

That requirement cuts in both directions, which is the part the objection tends to miss. When a correspondent has put a factual claim to this site, a name, a date, who did what in the historical record, the same habit of running it down against the primary sources has, more than once, shown the correspondent to be the one mistaken, however fluent or confident the challenge. The discipline is not that AI-assisted work is trustworthy and unaided human work is not; it is that neither is trusted on its fluency, and both are checked against the record.

There is one limit here I want to state plainly rather than paper over. Having an AI read and summarise a book is not the same as having read it myself, and where the treatment of a source rests on the former, the argument can honestly reach only as far as the specific passages that have been verified, a quotation checked in its context against the actual text, not the whole of a book read end to end. Where I have that verified passage I report what the document says and no more; where I do not, the gap is real, and pointing to it is a legitimate correction of exactly the kind this site asks for.

Whose Thoughts Are These?

A different objection has nothing to do with accuracy. It is that the work is inauthentic: a reader came for a person’s thinking and got a machine’s, and admitting the tool does not excuse it. I don’t think the result is any less valid for a machine having been involved, any more than an astronomer’s finding is less valid for the telescope. What would make it invalid is what would make an unaided argument invalid, a claim that does not hold against the record, and that test doesn’t care which instrument produced the sentence.

“Wholly mine” is in any case a higher bar than writing has ever met. I couldn’t mark each idea here with where I first met it; most of what any of us thinks is absorbed from others and repeated long after the source is forgotten. No man is an island in that sense, and no essay is either. Even writing this in a recent version of Word, whose Editor offers suggestions as you type, is by Microsoft’s own description the use of an AI writing aid. The honest position is to be plain about how the thoughts were formed and to hold them to a standard a reader can apply. What is wholly mine is the responsibility: whatever helped form a sentence, it goes out under my name, and I answer for all of it, the argument and the errors alike.

One worry I cannot answer. As these tools spread, I do not know what becomes of the individual’s own contribution, or whether we grow less inclined to develop and express our own thoughts. For me the discipline described above is the guard against it: the arguing, the checking and the demand for a source are where the thinking actually happens. I state the worry anyway, because I feel it too, and pretending I did not would be its own small dishonesty.

Compared to What?

The right test of a method is not whether it is perfect, no research is, but whether it is the best available for the work in front of it, and whether it is honest about how it can fail. Suspicion of AI assistance usually measures it against an ideal that was never on offer here; the real alternatives were these.

By hand, alone. Not the more rigorous option, but the same epistemics; one person, no peer review, only slower, less consistent, and with the human failure modes fully in play: fatigue, confirmation bias, an ego invested in the conclusion. On a site this size a single correction has to reach every place it touches; done by hand it reaches some and misses others. The honest outcome is not a more careful site but no site at all.

The academic route. Peer review, co-authors, an editor, the one thing an AI collaborator cannot supply, and the real gold standard for independent checking. It was not available: the field’s refusal to engage the documentary anomaly is this site’s subject, not a door that stood open and was declined. And peer review is itself imperfect, slow, conservative, and with well-documented problems of its own in catching error and reproducing results.

Paid research assistants. More independent than a second model, but far less scalable or consistent, still short of peer review, and a source of fresh errors of their own.

Not at all. The remaining option was to keep the work private and let the public record this site exists to create simply not exist.

For an independent researcher, on a task defined by scale and by the need for consistency across hundreds of cross-references, an AI collaborator held to a documentary standard is plausibly the best of those, and not despite the requirement that every claim be checkable, but because of it. That requirement is where the method actually beats the alternatives: every claim the argument depends on points to a document a reader can pull, and every correction is public. That is a higher standard of transparency than a peer-reviewed monograph, whose sources you usually cannot audit without the author’s whole library in hand. The method reports its error bars, and the unforgivable thing in research is not error, but claiming a precision you do not have.

The Limits, Stated Plainly

None of that puts the method beyond criticism, and three limits need stating exactly, because naming them is what the "best available" claim has to earn.

Checking one AI against another is the weak kind of checking. Verification here has sometimes meant cross-checking Claude’s output against a second model, and why that is thin needs spelling out. A model trained on much the same material shares the first one’s biases, including, on this particular question, a prior that runs to the orthodox attribution, for the reasons set out under the AI and search constraint. It is a second thermometer with the same miscalibration: it will catch a random slip, a transposed date, a garbled citation, but not a systematic error the two models hold in common. The only independent check is the one against a differently-biased instrument: the primary document, the physical book, or a person who knows the field. Where a claim rests on AI cross-checking alone, it rests on the weaker footing, and should be read that way. That caution is about verifying whether a claim is true, where a second model of the same training adds little. It is a separate matter from the exercise set out on Using AI to Check AI, where an AI is asked only to report what the site claims from the pages it is given, with the reader checking that account rather than trusting it.

The Errata is a lower bound, not a reliability score. It records the errors that have been found. It is silent on the ones that have not, and those cannot be counted, which makes it a floor under the error count, not a measure of accuracy. This site presses exactly that inference against others on The Denominator Problem: a tally of what you happened to catch tells you little about the size of what you missed, and the same cut applies here, which means the true error rate on this site is unknown and unmeasurable. The Errata, then, is not evidence the site is accurate, only that error, once found, is corrected in the open.

Provenance defends against invention, not against misreading. Pointing every claim to a real document rules out the fabricated citation with nothing behind it. It does not rule out two subtler failures: a genuine document can be read tendentiously, and which documents are cited at all is itself a judgment that can tilt a case. Fabrication was never the live risk; selection and interpretation are, and a footnote does not touch them. The only guard there is the discipline of keeping what a source says apart from what is inferred from it, set out on the Logic page, and a discipline can lapse. A reader who thinks a document here has been over-read is making precisely the correction this site most wants, and the one it is least able to catch on its own.

For the site’s evidentiary safeguards and the constraints on what it can claim, see the About page; every correction is recorded, in public, in the Errata. And if you want to put an AI to that same test, to check this site’s reading rather than take it on trust, see Using AI to Check AI.