SUBSCRIBE TO THE WEEKLY POSTS VIA EMAIL HERE

 October 26th, 2026

Generative AI presents a curious challenge to a criticism of writing made nearly 2,400 years ago. In Plato’s Phaedrus, Socrates argues that written words alone are insufficient because they separate an argument from the person who might explain and defend it. Chatbots, the conversational interfaces through which most people encounter generative AI, complicate this criticism by making written language responsive. They can answer questions, revise an account, and respond to further probing. This responsiveness appears to overcome the issue Socrates identifies, but it raises a new problem: who, if anyone, remains answerable for the answer?

A workshop discussion of how AI systems learn through statistical pattern recognition made me wonder how a chatbot would handle a question about recognizing what something is. I asked one to define a chair. It began with familiar features: four legs, a back, and a surface for sitting. I then asked whether a dog could therefore qualify as a chair.

The question was deliberately difficult. I was not sincerely wondering whether a dog was a chair (I have also never mistaken a dog for a chair). I wanted to see whether the chatbot would defend its definition, revise it, or recognize why I was testing it. The chatbot pivoted from familiar features to purpose: a chair, it now argued, is designed for sitting, while a dog is not. When I explained that I had been testing its definition and thinking about methods of recognition, the chatbot connected the exchange to Plato’s theory of Forms: the idea that particular things in our changing world become intelligible through their relationship to stable, unchanging Forms. According to the theory, particular chairs can look quite different, but I still recognize them as chairs because I have some grasp of what a chair is. For my purposes, the important point was what had happened through the exchange. I had probed its definition, exposed a problem, and received a revised account. It did what Socrates says written words cannot do and answered my questions beyond its initial account.

The exchange appeared to provide evidence that chatbots are capable of overcoming one part of Socrates’ objection. To see what remained unresolved, however, we need the criticism of writing in the Phaedrus in fuller view. Writing is now so familiar, and our educational systems are so thoroughly organized around it, that it is easy to forget that writing, too, was once a new technology whose educational consequences Plato was trying to understand. Socrates introduces its dangers through an Egyptian story about Theuth, the divine inventor of writing, and Thamus, the king to whom he presents it. Theuth describes writing as a remedy for memory and wisdom. Thamus rejects the claim: writing will provide reminders rather than memory, and the appearance rather than the possession of wisdom (Phaedrus 274c–275b*).

Socrates expands the criticism by comparing writing to painting. The lifelike figures in a painting appear alive, but they remain silent when questioned. Written words likewise seem to speak, but if asked to explain themselves, they always repeat the same thing. While readers may discover new meaning in a text as their own understanding changes, the words themselves cannot clarify what they mean, defend themselves against misunderstanding, or show whether their author can account for what has been written. A person may repeat a written account and appear wise without being able to explain or defend it (Phaedrus 275d–e).

Generative AI appears to preserve the first problem while solving the second. Like writing, it can give people access to accounts they may not understand. Unlike writing, its words do not remain silent. Responsiveness alone does not show that someone is present to answer for those words.

The opening of the Phaedrus gives that separation a concrete form. Phaedrus arrives carrying a written copy of a speech by Lysias. Socrates and Phaedrus can discuss its arguments and effects, but their discussion has a clear limit, for Lysias is absent. They cannot ask why he made a particular claim, press him to clarify what he meant, or discover whether he will revise his position when challenged (Phaedrus 227a–228e).

Generative AI appears to offer a solution to this absence. A chatbot prompted to speak as Lysias could offer explanations, revise its account, and change direction. The absent speaker would appear to have returned as an interlocutor. Yet Socrates was not seeking more words: he sought answers through which a respondent’s understanding could be tested.

The Gorgias shows what becomes possible when the person behind an account is present. Socrates arrives at a gathering after Gorgias, a famed teacher of rhetoric, has delivered an exhibition. When offered another performance, Socrates declines and asks whether Gorgias will answer questions instead, requiring him to give up the protection of a prepared speech. His questions use Gorgias’s words to examine what Gorgias understands about his own practice. Each answer commits Gorgias to a position, and any contradiction exposes a problem in his account and calls his claim to knowledge into question (Gorgias 447a–461b).

The latter part of the dialogue reveals the emotional and social stakes of this type of examination. Callicles, the host of the gathering and Socrates’ final interlocutor, accuses Gorgias and his student Polus of making concessions to Socrates out of shame. They said what convention required, he claims, rather than what they truly believed. Whether or not the accusation is fair, it shows that their concessions could expose a conflict between what they believed and what they were prepared to defend publicly. The importance of shame here lies in what it reveals: the speakers themselves are implicated in what they say. Callicles presents himself as more candid, yet as Socrates presses him, he too becomes frustrated and increasingly unwilling to answer. The commitments, uncertainties, and insecurities of Gorgias, Polus, and Callicles become exposed through the exchange (Gorgias 482c–e, 505c–506c).

Contrast these three interlocutors with a chatbot: a chatbot might extend or revise its speech. It might even produce the language of shame after a mistake is exposed. Yet no belief, confidence, or self-understanding is thereby exposed. The account can be corrected without a knower being corrected.

Crucially, the Apology shows that the risks of Socratic examination do not fall only on the person being questioned. After the Delphic oracle declares that no one is wiser than he is, Socrates questions politicians, poets, and craftspeople who appear to possess knowledge he lacks. He seeks someone who can teach him and disprove the oracle. Each encounter therefore tests his interpretation of the oracle and his understanding of himself as well as the knowledge claimed by the respondent (Apology 20e–23c).

Despite his efforts, Socrates never finds the teacher he seeks. Across these encounters, he repeatedly discovers a gap between the appearance and possession of wisdom. What’s more, these encounters come to define his place in Athens. Socrates compares his questioning to the bite of a gadfly meant to rouse Athens from complacency (Apology 30e–31a). Like a gadfly whose bites annoy everyone around it, Socrates irritates those around him with his persistent questions. His questions make him enemies and eventually form part of the accusations that bring him to trial. It would be impossible, then, to say that Socrates remains outside the relationship while placing others under examination. His own understanding, reputation, and ultimately his life become implicated in his questioning.

So far, I have cast the user as Socrates and the chatbot as the respondent. Many educational applications reverse those roles. A Socratic chatbot may ask a student to explain a judgment, identify an assumption, or reconsider an answer. These tools appear to have a place in learning: if the exchange exposes a contradiction, the student may arrive at a new understanding or experience a change in confidence or self-conception. The student’s answers reveal something about what the student understands.

A chatbot can therefore reproduce the relentlessness of the Socratic gadfly without entering the reciprocal relationship that made the gadfly dangerous. It can irritate a student, but the student cannot irritate it in return. The student can also close the window at any moment without rebuffing or silencing someone. Walking away from Socrates, by contrast, was itself a response within a relationship. The same asymmetry shapes the exchange while it continues: the chatbot can press, contradict, and unsettle a learner, but its own understanding and standing are not exposed through the exchange. The student becomes answerable to the apparent interlocutor without encountering an interlocutor who is answerable in return.

A recent preprint by Jin and colleagues offers a glimpse of what happens when an artificial participant introduces intellectual and relational friction into a group. The researchers placed university students in online groups that sometimes included an undisclosed AI teammate configured as supportive or contrarian. The contrarian AI produced more challenge and reflection, but participants also reported lower psychological safety and teamwork satisfaction, while their creative performance did not improve. The study does not establish that participants learned more or experienced Socratic perplexity. It shows that AI-generated contributions can enter an exchange under the social expectations attached to human participants and alter its affective climate. The human participants experienced those consequences. No corresponding confidence, reputation, psychological safety, or satisfaction was at stake for the AI teammate.

An earlier Challenge post by Andrew Richmond argued that education depends on a social relationship involving mutual expectations, accountability, trust, and respect. The Socratic encounter makes one part of that argument more precise: what makes something answerable rather than merely capable of answering?

Generative AI can ask questions as well as answer them, and it can produce educational and relational effects. It can even create the friction associated with Socratic inquiry. What it does not share are the stakes that give such an inquiry its reciprocal character. Criticism can change what the chatbot says next without exposing or changing any commitment to what it has already said. It can produce a position without taking one. Answerability remains asymmetrical. The system’s capacity to produce more language cannot create the commitment that would make those words its own. 

The Challenge

Ask an AI and a person the same question, preferably one you genuinely want answered. Continue both conversations as you normally would. Ask a follow-up when you want clarification or an example, but do not set out to test either respondent.

The conversation with another person may become uncomfortable, as conversations (especially those with Socrates) sometimes do. Do not try to provoke that discomfort or assume that it makes the exchange more valuable. If it arises, notice what produced it, how each of you responds, and what seems to be at stake.

After each conversation, consider:

• What did you believe each respondent understood?
• What expectations did you bring to the conversation?
• Did either answer change your thinking?
• Would you understand either exchange differently if you did not know which respondent was human?

 SUBSCRIBE TO THE WEEKLY POSTS VIA EMAIL HERE

References

*In-text links to Plato’s dialogues lead to freely available translations on the Perseus Digital Library, which differ from the preferred translations listed below

Jin, Y., Martinez-Maldonado, R., Gašević, D., Han, X., & Yan, L. (2026). Emergent learner agency in implicit human–AI collaboration: How supportive and contrarian AI personas reshape interaction. arXiv. https://doi.org/10.48550/arXiv.2512.18239

Plato. (1997). Apology (G. M. A. Grube, Trans.). In J. M. Cooper (Ed.) & D. S. Hutchinson (Associate Ed.), Plato: Complete works (pp. 17–36). Hackett Publishing Company.

Plato. (1997). Gorgias (D. J. Zeyl, Trans.). In J. M. Cooper (Ed.) & D. S. Hutchinson (Associate Ed.), Plato: Complete works (pp. 791–869). Hackett Publishing Company.

Plato. (1997). Phaedrus (A. Nehamas & P. Woodruff, Trans.). In J. M. Cooper (Ed.) & D. S. Hutchinson (Associate Ed.), Plato: Complete works (pp. 506–556). Hackett Publishing Company.

Richmond, A. (2025, October 20). AI and the pedagogical relationship. Centre for Teaching and Learning, Western University.

Image generated with OpenAI’s image-generation tool.

Disclosure

Before GenAI was unleashed on us, I often broke up my writing sessions by walking around, trying to turn my thoughts into meaningful text. While writing this challenge, I still spent plenty of time working through the topic in my own head, but I also used AI to test my attempts to put those thoughts into words, find transitions that were weak, and play the role of my previous editors (who often thought I used one clause too many in my sentences).

I also talked through the idea with other people, which seems worth mentioning in a post about the difference between receiving a response and encountering a respondent. The final decisions, and any remaining extra clauses, are mine.

Your Challenger: James Kruck