Past the second watershed. Of the six balances: 2 lost, 3 tipping, 1 not assessed. Position at least 88 of 100. The balances not assessed can only move it further.
- Biological degradation Tipping Does this tool damage the living world, or the body, faster than either can recover? Tipping: the tool has begun to cause this problem, without going past the limit.
- Radical monopoly Tipping Can people still meet the need another way, without this tool? Tipping: the tool has begun to cause this problem, without going past the limit.
- Overprogramming Tipping Does using this tool require training or certification that people cannot give each other? Tipping: the tool has begun to cause this problem, without going past the limit.
- Polarization Lost Does this tool move power or advantage to a few, at others’ cost? Lost: the tool is clearly past the limit.
- Obsolescence Lost Does this tool force people to replace what still works? Lost: the tool is clearly past the limit.
- Frustration Not assessed Does this tool give back less than it takes, counting everything it takes? Not assessed: nobody has looked at this yet, so it is not counted as holding.
What it does
Produces text in response to text, as a service run by a company: answers, summaries, translations, drafts, and computer code. It works by predicting likely next words from patterns found in a very large body of writing.
Where it stops
It doesn’t, by design. A model answers any question in fluent prose whether or not it is right, and its makers aim it at every kind of work. What it says, what it keeps of what you ask, and whether it is there tomorrow are decided by the company that trains and runs it.
How easily you can stop using it
Easy for one person today: you can still read, write, and look things up yourself. Harder each year, as the models are built into search, office software, and schoolwork. A hosted model can be changed or withdrawn at any time.
What is scored
- The unit: A large language model offered as a service by the company that trained it, reached through that company’s website, apps, or other software. Left out: models a person downloads and runs themselves (see Local language models).
- The job it is tested against: Producing fluent, relevant text on request. It does that reliably. It is not tested here on whether the text is true, although answers, summaries, and translations are only useful if they are. If truth is the job, this record belongs below the first watershed; see the open questions.
- Where and when: The large general-purpose services of 2023 to 2026. The energy figure is the International Energy Agency’s projection; the rest is argument from how the services are offered, not measurement.
A disclosure first: a large language model drafted this record, at the editor’s direction, and the editor revised it. Read the scoring with that in mind.
A large language model is software that produces text in response to text. It is built by having a computer find statistical patterns in a very large body of writing, and it answers by predicting, a piece at a time, what words are likely to come next. Ivan Illich did not live to see one. But he and Barry Sanders described the kind of mind they belong to, and the catalogue’s record on Literacy is the best place to start.
A machine made of literacy
These models could not exist in a culture without writing. They are made from the stored text of literate societies: books, articles, letters, forum posts, and computer code, collected by the billions of pages. Walter Ong wrote that writing lets words sit still so they can be compared, listed, and analyzed. A language model is what happens when a machine does that to nearly everything written down.
Ong also described a secondary orality: the telephone, radio, and television, which sound like speech but depend on writing to exist. A conversation with a language model goes a step further, and this reading is ours, not Ong’s. It has the form of talk between two people, but one side of it is produced entirely from text.
Plato’s complaint, again
In the Phaedrus, Socrates objects that writing gives the appearance of wisdom without the substance, and that a written text cannot answer a question: ask it something and it goes on saying the same thing.1 A language model looks like the answer to the second objection. It replies to anything. But it sharpens the first. Its replies are fluent whether or not they are true, and there is no one behind them who knows the subject or can be held to what was said. Emily Bender and her co-authors named this in 2021, when they described such a system as a “stochastic parrot.”
Language as information
The last part of ABC: The Alphabetization of the Popular Mind (1988) argues that the literate mind is being replaced by one modelled on the computer, in which language is treated as information to be processed and exchanged. Illich made a related argument in a 1987 talk, “Computer Literacy and the Cybernetic Dream.” He and Sanders were writing about how people had begun to think and speak. A language model is the same idea built as a machine: it handles words as quantities, with no speaker and no occasion.
Illich’s idea of lay literacy applies as well. He argued that writing reshaped people who never learned to read, because they came to live among documents. In the same way, people who never use a language model will read letters, news, school reports, and official decisions that one has written.
The opposite of a book
The record for The book says that a book says the same thing to every reader, cannot watch who reads it, and cannot be withdrawn once it is in your hands. A hosted language model reverses each of these. It says something different every time. It can keep a record of what you ask. And its maker can change it or switch it off. A model that runs on your own computer is closer to a book in the last two respects, and has its own record: Local language models.
The pattern of literacy, repeated
Literacy became a requirement, then a credential, and was taught in the ruler’s language. The same steps are visible here. “AI literacy” is already offered as something everyone must acquire, and schools are asked to teach it. The models work best in the languages with the most text online, English above all, which adds to the pressure on smaller languages. And the writing they were built from was taken from millions of people who were not asked. For First Nations in Canada that is a question of sovereignty over a people’s own data as well as of authorship; see Local language models for the OCAP principles and for models that communities hold themselves.
The six balances
Scored as set out in How we evaluate tools. The first question is whether the tool reliably does its job. It reliably produces fluent, relevant text. It does not reliably produce true text. We have scored it as doing its job, and listed that decision as an open question.
- Biological degradation: 1. Training and running these models takes large data centres. The International Energy Agency expects data centres’ electricity use to more than double between 2024 and 2030, to about 945 terawatt-hours, with AI the largest cause.
- Radical monopoly: 1. The models are being built into search, office software, phones, and schoolwork, so that using one is becoming the default way to find something out or write something down. People can still do both without them.
- Overprogramming: 1. Writing is one of the ways people work out what they think. A tool that writes for them can take the place of that work. The evidence on what this does to learning is new and contested.
- Polarization: 2. Building a leading model takes computing power and money that only a handful of companies have. They built their models from the writing of millions of people, without asking, and they set the terms of use.
- Obsolescence: 2. Hosted models are replaced and retired on their makers’ schedules, often within a year or two. People who build their work on one must rebuild it when the model changes.
- Frustration: Not assessed. The return is real for some tasks. The cost is that every answer has to be checked by someone who could have found it without the tool.
Where the threshold falls
Past the second watershed, at 88. Two balances score 2, and any balance at 2 puts a tool past the limit: 70, plus 2.5 for each of the seven points in its total. One balance is not assessed, so this is the lowest the position can be. That is the same position as Compulsory schooling, and close to The algorithmic social feed.
The score describes the tool as it is mostly offered now: hosted by a few companies, general in purpose, and added to other tools whether or not their users asked. It is a first reading of a tool that is changing quickly, and it should be scored again. For a guess at what Illich himself would have made of these machines, see the research note What would Ivan Illich make of large language models?
Open questions
- Does a language model reliably do its job? If the job is true statements, it may belong below the first watershed. If the job is fluent text, it is past the second.
- A model that runs on a person’s own computer now has a separate record, Local language models. Check the two against each other as both change.
- What does writing with a language model do to a person’s own writing and reading? Find the studies, and say what they do and do not show.
- Check the description of ABC’s last part and of “Computer Literacy and the Cybernetic Dream” against the printed texts, and add page references.
- Add figures for water use, and for the energy cost of a single answer, with sources.
- How does this record relate to Literacy? It could be read as literacy’s latest stage, or as its replacement.
- Plato, Phaedrus, trans. Alexander Nehamas and Paul Woodruff (Indianapolis: Hackett, 1995), 274c–275e. ↩︎
Sources
- Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?.” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. Association for Computing Machinery, 2021. https://doi.org/10.1145/3442188.3445922.
- Illich, Ivan, and Barry Sanders. ABC: The Alphabetization of the Popular Mind. San Francisco: North Point Press, 1988.
- Illich, Ivan. In the Mirror of the Past: Lectures and Addresses, 1978–1990. New York: Marion Boyars, 1992.
- International Energy Agency. Energy and AI. Paris: International Energy Agency, 2025. https://www.iea.org/reports/energy-and-ai.
- Ong, Walter J. Orality and Literacy: The Technologizing of the Word. London: Methuen, 1982.
- Plato. Phaedrus. Translated by Alexander Nehamas and Paul Woodruff. Indianapolis: Hackett, 1995.