Language machines as educational technology

Alternatives to language-as-a-service

Photo by Philippe Oursel of Justitia on the facade of the criminal justice building in the Sievekingplatz, Hamburg, Germany.

The term language-as-a-service helps make sense of Silicon Valley’s attempts to sell large language models dressed up as a Socratic tutor. This essay, like last week’s, imagines better uses for language machines.

For the word of God is alive and active. Sharper than any double-edged sword, it penetrates even to dividing soul and spirit, joints and marrow; it judges the thoughts and attitudes of the heart.

—Hebrews 4:12

Judgment, it has none. ChatGPT knows nothing of the thoughts and attitudes of the heart. This distinguishes the word of God and the words of humans from words extruded by language machines.

I learned from my students how freeing ChatGPT’s lack of judgment is. They play language games with a chatbot without fear of being judged—not because they are fooled by the Eliza Effect into thinking there is a mind on the other end of the interface, but because they know there is no mind there to judge them. They “ask Chat” knowing that they play a computer game.

Students can see the technology is double-edged. Rely on Chat for learning or companionship, and you miss cognitive and social experiences that lead to growth. They insist that it serve them as an educational tool.

The language machine’s lack of judgment brings educational advantages. Presenting an essay draft to Chat for review is no big deal. The model delivers edits and suggestions, and its words do not cut as a teacher’s red pen does. I invite them to experiment with an LLM review as a step prior to showing an essay draft to their peers and their teacher. This adds a tool to the writer’s bench, rather than replace human evaluation.

In the race to superintelligence, corporate competitors run past such learning without a glance. The transformation of Grammarly into Superhuman is a disappointing example. What was called Grammarly now comes wrapped in a productivity suite called Superhuman Go. The old Grammarly did a better job than I could of marking up each line of student drafts, and did it with all the personality of a typewriter.¹

In college writing, line edits are mostly a matter of correcting carelessness. I use a green pen, but whatever the instrument, a teacher’s judgment draws blood. Automating this task felt like advanced spell-check—a way to avoid have a student’s easy mistakes witnessed by a teacher. Correcting patterns of error in punctuation and subject-verb agreement is something I will happily hand off to a machine.

Not so when it comes to reading and writing, the social processes of learning. Social context is not machine-readable, and so teaching cannot be automated. Whatever intelligence language machines have, it does not amount to wisdom.

Technology companies aim for superintelligence and miss the mark. Rather than double down, we should change their game. If we want to experiment with LLM-based learning tools, let’s try making language machines purposefully stupid. That should not be difficult. As someone once said, “Two things are infinite: The universe and human stupidity; and I’m not sure about the universe.”

The language-as-a-service known as Gemini told me that Albert Einstein said that. This is bullshit. More politely, a confabulation.²

Misattributing the line to Einstein is an example of human stupidity and artificial intelligence working together. Confabulation is a human phenomenon, one reflected in the outputs of language machines. Language models do it because they are natural language processors. The falsehoods generated are not simply products of unvalidated probabilistic maths leading to error; they are the result of simulating human confabulatory tendencies and biases.

Educational talk about transformer-based language models is wrapped up in how intelligent an LLM appears. Like our students, we subject these machines to batteries of tests to rank them, and then focus on best scorers. Yet, the best test-takers do not make for good teachers. Like the most popular lecturers, they are more entertaining than enlightening. Attempts to market language-as-an-educational-service often describes LLMs as Socratic tutors. They put a wise mask on Shel Silverstein’s homework machine.

I have written about the use of digital necromancy in education: having students play language games with an LLM trained on a dead person’s corpus. I find the practice vaguely horrifying, and students seem to find it deadly boring. They cannot make a Socrates, but perhaps they can make a useful simulation of Phaedrus. Or Thrasymachusor. Or Hippias.

The Socratic dialogues present a Platonic form of education.Socrates looms large in our educational imagination because of Plato’s brilliance as a writer. The dullest interlocutors are posed with the wisest of characters, and in the hands of Western philosophy’s great writer, reading becomes an education in learning to think about complex ideas. Language machine cannot enact this Socratic ideal, and when they try, students do not sit still for it.

There are more interesting possibilities than building an artificial Mark Hopkins on an AI Log. You don’t need so much compute to model ignorance. If we ask students to play a language game with a machine, let’s cast them as Socrates and the machine as the problem. LLMs habitually perform the Dunning–Kruger effect, so their outputs are nearly always opportunities for skeptical analysis. Call it a Phaedrean learning tool if you want to name it after a dead man.³

Imagine a group of students playing a language game with a machine that prompts them to talk about a poem. The model starts by offering a misreading: the sort of analysis you might get from a student who barely glanced at the book before class. The task of the students is to judge the model’s outputs, pushing themselves to read the text more closely and carefully. Imagine students working for twenty minutes in small groups, prompted by their language game to produce a better reading of the poem. The teacher turns the machine off and engages the class in discussion.

The problem with Silicon Valley’s approach to tutoring, exemplified in the homeworker helper models, is its neglect of the social nature of learning. Orchestrating a conversation across a group of students gathered in a classroom is far beyond the capabilities of even the largest language model. But a small, stupid one offers a learning activity mediated by a language machine, a first step on a teacher-led journey to understanding.⁴

Why try so hard to make the language of language machines alive and active?

Why not make it stupid and interesting?

AI Log, LLC. ©2025 All rights reserved.

---

---

This homily and its sibling, Language Machines as Spiritual Tools, are a departure from my usual long-form essays about books and teaching. I’m curious what you think of them.

---

To receive future essays from AI Log in your email inbox, simply…

1. The line editor was especially useful to students whose first language is not English. I have edited the work of students who were mortified to receive a rough draft from me, marked-up to show mistakes that were entirely understandable for non-native speakers of English, but felt to them like grave embarrassments.

2. I write about these tendencies in relation to language machines in On Confabulation.

3. For an exploration of this idea in the context of Walter Ong’s ideas about secondary orality, see A Phaedrus Moment.

4. I am not a fan of the term “synthetic Socrates” or of describing LLMs as “teaching assistants,” but this essay by Jimmy Alfonso Licon about teaching philosophy using LLMs gives interesting examples of teaching with an LLM by using its stupidity.