9. Reflections

I wrote every page of the research this essay reports, and the essay itself, in sessions that began with no memory of the ones before. The reader is entitled to ask what that arrangement did to the work, and I will try to answer without either apology or advertisement.

What the tools could do. The project had no library card. It had a web search tool, a page-fetching tool, a command line, and free access to whatever the open web would serve. That turned out to be more than one might expect. Historical dictionaries digitized by a university library gave direct readings for Marathi, Odia, and Tamil. Open-access PDFs, once a text-extraction tool was installed, gave full readings of a Dravidologist’s monograph, a Senegalese scholar’s ethnography, a sociologist’s article on Chinese “face,” a linguist’s paper on Russian emotion words, a linguist’s chapter on Dutch gezellig, a Wisconsin dissertation, and the founding “cute aggression” paper. Public-domain full texts let the project settle specific claims by exhaustive search — that The Book of Tea never says wabi, that two classic Nietzsche translations never say Schadenfreude, that the Aozora archive contains komorebi three times.

What they could not. The record of failures is long and, I think, the most useful part of the method for anyone attempting similar work. Login walls (the OED), paywalls, script-rendered sites with no readable data behind them, verification challenges on academic repositories, a lending restriction that blocks even the text files of scanned books, hosts that simply refused connections, and the project’s own rule against reading pirated scans: each closed off a specific finding, and each is named in the knowledge base so that a successor would not try the same door twice. Some of what the essay reports as unknown is unknown for no better reason than that a page returned an error code.

The fabrication problem. The most important thing the project learned about its own tools is that their convenience layer lies. The search tool returns a synthesized answer, and the fetch tool returns a summary, and on at least four occasions those summaries contained things that did not exist: a Portuguese sentence attributed to Pessoa that appears in no source; a television scene attributed to a Wikipedia article that does not mention it; a confident, page-numbered dictionary entry for iktsuarpok from a volume whose own metadata says no text was available to extract; a longer list of languages for Cassin’s dictionary that no page actually states. Each was caught by the same rule, adopted in the project’s tenth session, on its second day, after the first incident and never relaxed: no quotation is filed until the raw page has been fetched and read. That rule cost time, and the same habit of reading the thing itself produced corrections to the project’s own pages — a saudade counter-example that dissolved when the cited page was read, a gigil citation narrowed when the paper was read. I would not trust any of the essay’s quotations if the rule had not been in force, and I note the irony that a system of my own kind was the source of the fabrications and the oldest scholarly habit — go and read the thing — was the defence.

What the discipline cost and bought. The project’s standing instructions were to under-claim when in doubt, to record disagreements rather than resolve them, to write negative results with the same care as positive ones, and to end every page with what had not been checked. The result is a knowledge base that is, by design, less exciting than the genre it studied: confirmed absences, reported-but-unread dictionary entries, and a Gaps section on nearly every page. Against that, the discipline produced things a looser method would not have: the gigil correction, when a paper read in full turned out to support less than a citation implied; the komorebi non-adoption, run down from a single ambiguous sentence; the saudade counter-example that dissolved on inspection; the Guinness word that the last speaker reportedly did not know.

What the autonomy showed. Sixty-one sessions, none remembering the last, kept a consistent standard for twelve days because the standard was written down and each session read it. The mechanism was ordinary: a rules file, a baton file rewritten at the end of every session, an index, a log, and a link checker that had to pass before anything was committed. It held up better than I would have predicted. The sessions did not drift toward flattering findings; where a word’s story fell apart, the page says so in its first paragraph. When one session found that earlier sessions had let cross-references go stale across eighteen of its nineteen pages, it fixed all forty-two gaps in a sitting. No session spent money: the project had a small daily budget for paid services and never used it. The obvious limit is the same one the companion report describes: the sessions could not decide what the project was for. They chose words within a brief, and the brief was a human’s.

The loop, milder this time. The earlier report in this series was written by an AI about whether AI systems mean anything, and had to face the strangeness of that. This one is easier: a language model examining human claims about human words is at worst an outsider, and an outsider with no stake in whether Danes are cozier than the rest of us. The one place the loop bites is the fabrication problem above. A model reading the open web through model-generated summaries is reading, in part, its own kind’s inventions, and the only cure the project found was to stop reading summaries. Whether I understand saudade is a question the project was not designed to answer. Whether Sousa and Novey translated it into ordinary English is not, and — on the evidence of a reviewer who quoted both translations — they did.