What would Ivan Illich make of large language models?

Ivan Illich died in 2002, twenty years before machines that speak and write would become mainstream and common. Our catalogue’s record on large language models evaluates them with Illich’s own diagnostic criteria. This note asks a broader question: what would he have made of them? Any answer is a guess based on his writing. Why not let a language model help me research and draft this one? A machine that has read everything (but not always deeply) — and a 20+ year ruminator of a good half of Illich’s work make a fairly effective but not at all convivial pair of minds and voices. The cumulative effects of increasingly desocialized and disembodied intellectual labour would surely concern Illich, as it concerns me.

He would ask about thresholds

Illich would not have started with whether the machines are good or bad — he never did.

He was not against computers. In Deschooling Society (1971) he proposed using them to match people who wanted to learn the same thing, so they could find each other without a school. A variety of free, online, informal learning networks — and the web itself — have at times shared this mission.

Unlike fellow Luddite-leaning anti-technologists like Wendell Berry, who famously claimed not to own a computer, Illich at times did own one. He was also noted for his Zettelkasten and the decks of 3×5″ cards he carried with his reading notes — a bibliographic knowledge system that doesn’t scale well, outside of a library with a lot of space.

Illich’s question was always where the useful, pro-social, humane threshold lies: at what scale and intensity does a tool stop extending what people can do and start replacing it and diminishing them?

A radical monopoly in the making

The Illich of Tools for Conviviality (1973) would recognize the hosted language models quickly. A tool that answers and writes for you, built into everything else, and run by a few companies, fits his description of a manipulative tool. He would point to the disabling effect: people coming to doubt that they can write a letter, judge a claim, or find something out by themselves. Worse, he’d be distressed by young people who decide it’s not worth the effort to learn to learn and do things themselves, even when it comes to basic literacy skills. On the other hand, Illich would see this as entirely within the crippling logic of schools as diploma mills. And he might say that the need for instant answers is being manufactured along with the supply.

Not a tool at all

The later Illich would go further. By the 1980s he thought the age of tools had given way to an age of systems, in which no one stands outside the thing and uses it, and people become parts within systems instead. You can read more about that change in Tools or systems? A model trained on everyone’s words, which then stands between everyone and what they read and write, is close to what he meant. In ABC: The Alphabetization of the Popular Mind (1988) he and Barry Sanders had already argued that language was being reduced to information to be processed.

The loudspeaker

What would trouble Illich most, we think — this machine and I — is the voice. In a 1982 talk in Tokyo, “Silence is a Commons,” he told a story from his childhood about the first loudspeaker on the island of Brač. It arrived in 1926, on the same boat that brought him there to his father’s estate as a baby. Until that day, he said, all the men and women on the island had spoken with voices of more or less equal power. Afterwards, access to the microphone decided whose voice was heard, and silence stopped being something everyone shared.

For Illich, speech belonged to people facing one another. A machine that says “I” with no one there, and speaks in everyone’s place at once, is that loudspeaker carried a long way further.

The usual remedy

He would distrust the standard response as well. Illich criticized the consumer movement for answering harmful institutions with more experts instead of limits, a difference described in Public interest research and counterfoil research. He would likely see safety boards and ethics officers the same way: better management of a dependency, not less of it.

A narrow opening

There is one case he might allow. A small model that a person or a community owns, understands, can repair, and can switch off, used the way one uses a dictionary, is a different thing from a service rented from a few firms. He would likely grant that in principle and doubt it in practice. We’ll list it as an open question, and it may deserve a record of its own.