Digital Humanities Lab Accused of Over-Indexing 'General Vibe' of the Carolingian Renaissance

The machine delivered its verdict on the Carolingian Renaissance at 4:12 on a Thursday: 0.81, on a scale the Orris Institute's Textual Dynamics Lab calls the Carolingian Vibe Index, with an attached confidence note reading "strong sense of renewal." The number was printed, stapled, and circulated to three departments before anyone asked what the model had read.

I asked. I pulled the training manifest, which is stapled to nothing and sits behind a request form, and the mechanism has a name that does not appear anywhere in the lab's press materials. Of the roughly 412,000 documents in the corpus, about 397,000 were published after 1860 and concern the Carolingian renaissance. None were produced during it. The model was fed a century and a half of admiring scholarship, the adjectives of Giesebrecht and his inheritors, the adjective-heavy forewords of museum catalogues, the word "flourishing" appearing in a form the ninth century never used, and it returned, with 0.81 confidence, the average opinion of historians who think Charlemagne's court was a nice time.

"It has quantified the Victorian hangover," said Professor Emmanuel Crace, a philologist who has spent eleven years on a single capitulary and can name the manuscript that a mid-career scholar once misread in his presence. "At least a stemma admits it is guessing. This thing guesses at volume and calls it a finding."

The lab's reply was careful and, in its way, honest. "We never claimed to read the manuscripts," said Dr. Anneke Vos, who built the index with two graduate students and a cluster funded by the Lazlo Foundation, which has not commented. "We claimed to measure the shape of the scholarship, and the shape is real. Scholars have described this period as a renewal for four hundred consecutive years. That pattern is data."

It is data about something. The test the lab ran afterward showed it plainly. Set against a well-known modern guidebook's foreword, the model scored 0.94. Set against a facsimile page of the Poetae Latini Aevi Carolini, it scored 0.46 and helpfully noted that the passage "lacks renewal markers." A ninth-century poem was judged insufficiently ninth-century by a machine that had never been shown a ninth-century poem as anything other than the thing being admired.

Chandra's Bench Notes, as always, keeps the failed part: the defect is not the model, it is the label set. Someone, somewhere in the pipeline, mapped "vibe" onto sentences written by modern people praising the past, then treated the praise as a property of the past. Subtract the post-1860 material and the index drops to 0.44 — a score the press office has not reissued.

Vos's team intends to relabel and rerun. Rerunning costs about nine days of cluster time, which the grant does not have twice. Meanwhile the index has been cited twice, once by a podcast and once in a conference abstract with a colon-heavy title.

A philosopher of history at the coffee machine offered the shortest version I have heard: "The machine didn't find the vibe of the ninth century. It found the vibe of everyone who has written about the ninth century, which is us, at conferences, agreeing."

The lab's next target is the Investiture Controversy. Surviving primary material for the training window: three letters and a disputed chronicle. Everything else will be, once more, the historians.