Cohere's speech-recognition model (cohere-transcribe-03-2026)
transcribed the recordings. It takes no instructions, so its errors are
mishearings. Fixed numerical rules, not AI judgment, filter out its typical
failures, such as repetition loops and invented text at the end of a tape.
Passages removed this way are not deleted: they appear on their rows as
expandable "low-confidence audio" items. We adopted this after an
independent review found that the filter was also catching real speech.
Silent stretches read "(audio gap)." A separate local model (Whisper)
supplies the timing for click-to-play and nothing else.
Where no digital text existed, vision models read the scans: OpenAI's GPT-4o-mini first, Claude for pages GPT-4o-mini refused, and Google's OCR engine for the rest. Part of their instruction read: "Transcribe ALL of its body text VERBATIM… Preserve original wording, spelling and accents exactly; do NOT translate, summarize, modernize, or correct… If a word is truly illegible, write [illisible]. Do not invent text."
Text was taken from each edition's digital text layer, or OCR'd under the same verbatim-only instruction, and then divided into sessions by the printed date lines. Fixed rules removed book apparatus (chapter titles, running heads, copyright pages). These rules were checked against Staferla so that real content could not be mistaken for apparatus, and every removed item can still be read in a popover ("Editorial chapter heading," "Book front matter"). Where an edition leaves out or shortens a session, a hand-kept list says "Not in this edition" instead of leaving a blank.
All 5,603 footnotes in the corpus were reviewed one by one, and the resulting changes were applied by tools that log each edit. There were 973 changes, including 70 passages of real prose that had been swallowed into notes and were put back. Any note whose wording was repaired shows "AI-amended. Original reading: …" followed by the original.
The Staferla text is the spine, chosen because the Staferla editions are complete. It generally stays closer to the stenographic transcripts than the official editions do, and the alignment is built on it. Its bibliographic insets were moved into popovers. (An earlier version of this step removed them without a record; a review caught it, and every move is now logged.) Long quotations stay in the text, by the curator's decision. Reference texts appended to some sessions, such as Freud in German and Augustine's De Magistro, appear as appendix bands.
A small number of misprints were corrected. Each correction was checked against another source and made only when the right reading was clear. Corrected words carry a dotted underline and show the original reading ("AI correction of the source transcription"). 129 damaged characters that could not be resolved with confidence were left marked, not guessed.
No text is written at this stage. Cohere's multilingual embedding model
(embed-multilingual-v3.0) measures how close two sentences
are in meaning and generates nothing; sentences are placed in rows by
fixed thresholds that keep each source in its original order. When Claude
was asked to judge doubtful rows, it could only propose moves, and it had
to quote a French phrase from the text as evidence for each one. A script
checked every quotation, proposals with a quotation that did not match
the text were thrown out, and a person applied the rest.
The English and Portuguese translations are the only place where an AI
writes continuous text, and they are labeled as machine translations
wherever they appear. Both are made directly from Staferla by Anthropic's
Claude (claude-sonnet-4-6, at low temperature), with
GPT-4o-mini handling the few passages Claude's filter declines. The
prompt keeps jouissance, objet a, lalangue and
parlêtre in French and fixes the standard English renderings
("the Real," "signifier," "surplus jouissance," "subject supposed to
know"). It also says: "Preserve every logical / mathematical symbol
EXACTLY as it appears in the source… Do NOT translate, romanize,
paraphrase, or drop them," and asks for every paragraph break to be
kept. The Portuguese prompt uses the standard Brazilian terms
(gozo, lalíngua, falasser; pulsão,
never instinto; recalque, never
repressão).
Sometimes the model added a bracketed gloss or editor's note of its own. These are kept and shown as "AI translator's note" popovers. Treat the translations as a fourth opinion, not a primary source, and do not cite them as published translations. The highlight-to-translate popover uses the same model with a similar instruction, and the same caution applies.
Claude vision models found the diagrams in the scans and checked their quality. They produced image crops only, never text. A second, separate visual check removed about 1,600 bad crops.
Every tool that changed the corpus was run as a dry run first, backed up each file before editing it, and wrote each change to a dated log. Before anything is published, an automated check confirms that no words were lost or added in processing, and a session is published only if its sources are exactly the ones that check approved. Serious problems stop the release.
The work was also reviewed by large batches of AI agents checking the code and the texts, and by two outside AI reviewers from other companies (OpenAI's Codex and Moonshot's Kimi), who were allowed to criticize but not to change anything. Problems they found, including citations removed without a record, an audio filter that was too aggressive, and appendices that did not display, were each fixed, logged or made visible. The curator made the main editorial decisions: what stays in the text, what goes into a popover, and what is disclosed and how.
The audio transcription has serious limits and should always be read
against the stenotype or the recording itself. OCR of worn carbon copies
is imperfect; [illisible] marks words that could not be read.
The alignment has a measured error rate that is tracked over time. Audio
timing is approximate in 26 of the 155 sessions with audio. Some audio
passages were wrongly filtered and are waiting to be restored, and a few
damaged notes still need to be checked against printed copies. Finally,
the largest review campaigns were run in interactive sessions with AI
assistants, and not all of those conversations were saved. Every change
they produced is logged, backed up and can be audited.
A separate note covers the construction of the index of names and concepts: A Note on the Construction of the Index — AI-Written.
← Back to the seminars