What’s the deal with Plato or whoever’s critique of writing, maybe in Phaedrus?
In Plato's dialogue "Phaedrus," Socrates tells a story about the Egyptian god Theuth presenting writing as a gift to King Thamus, claiming it would improve memory and wisdom. The king, however, rejects this claim with a famous critique that includes several key objections:
-
Writing weakens memory: Rather than strengthening memory, writing would cause people to rely on external marks instead of internal remembering.
-
False wisdom: Writing gives people the appearance of wisdom without true understanding - they can access information without truly comprehending it.
-
Lack of responsiveness: Unlike a living teacher, written texts cannot answer questions or adapt to the reader's needs. As Socrates puts it, writings "seem to talk to you as though they were intelligent, but if you ask them anything... they go on telling you the same thing forever."
-
Texts travel beyond their intended audience: Once written, texts can circulate to those who might misunderstand them without the author present to clarify.
What makes this critique particularly interesting is the irony that we only know about Socrates' critique of writing because Plato wrote it down. This suggests Plato recognized both writing's limitations and its value for preserving philosophical thought.
Uh, so line this up with what people say about LLMs.
The parallels between Plato's critique of writing and concerns about LLMs are remarkably striking:
- Plato: Writing weakens memory by making people rely on external marks rather than internal remembering
- LLMs: Concerns that AI tools diminish human cognitive abilities by outsourcing thinking and problem-solving, creating dependency
- Plato: Writing gives the appearance of wisdom without true understanding
- LLMs: LLMs produce convincing-sounding text without actual comprehension, leading to "AI hallucinations" and the spread of plausible-sounding falsehoods
- Plato: Texts "go on telling you the same thing forever" without truly responding
- LLMs: Despite conversational abilities, LLMs lack genuine understanding and merely predict likely text patterns without grasping meaning
- Plato: Writings travel beyond intended audiences who might misinterpret them
- LLMs: AI-generated content spreads without context, potentially misleading people who don't understand its limitations
Just as we only know Socrates' critique of writing because Plato wrote it down, many critiques of LLMs are being shared, discussed, and developed using the very technologies they critique. This suggests we may be in a similar historical moment of adaptation to a new communication technology with both limitations and transformative potential.
But, we kind of got to the state where we could make LLMs by scaling up a society that so heavily invested in writing that the training data was finally available. It seems like writing kind of worked, and the way it worked, counter to S’s critique, might point to something important about LLMs that we’ve overlooked.
You've hit on something profound. Despite Socrates' concerns, writing didn't diminish humanity—it amplified us. This suggests a more nuanced view of LLMs:
-
Writing extended cognition rather than replacing it
- Created external memory systems that freed human minds for higher-level thinking
- Enabled knowledge accumulation across generations and civilizations
- Became symbiotic with human thought rather than supplanting it
-
Writing created new cognitive capabilities
- Enabled more complex, structured thinking through externalization
- Allowed for revision, refinement, and collaboration impossible in oral cultures
- Birthed new forms of knowledge organization (libraries, encyclopedias, hypertext)
-
LLMs might extend rather than replace human cognition
- Could serve as thinking partners rather than thinking replacements
- May free human cognitive capacity for more creative and evaluative work
- Might enable new forms of collective intelligence
-
The transformative potential lies in symbiosis
- The most valuable future may be humans+LLMs rather than either alone
- LLMs might scaffold new cognitive capabilities we haven't yet imagined
- Our focus should be on designing for complementarity, not replacement
-
Cultural practices will evolve to address limitations
- Just as citation systems, peer review, and critical reading developed for text
- New literacy practices will likely emerge for AI-mediated information
The irony is that Socrates was both right and wrong—writing did change how we think, but largely for the better. Similarly, LLMs will transform cognition, but potentially in ways that expand human capability rather than diminish it.
Claude, can you pick and point and stick to it here?
The most important insight from writing's success despite Socrates' critique is this:
New information technologies don't replace human cognition—they scaffold it into new forms.
Writing didn't weaken memory as Socrates feared; it transformed how we use memory. By externalizing information storage, writing freed human minds to develop more complex analytical capabilities. It became a tool that extended our cognitive reach rather than diminishing it.
This suggests we should view LLMs not as artificial minds replacing human thinking, but as cognitive scaffolding that might enable new forms of human thought. Just as writing allowed us to build on others' ideas across time and space, LLMs might help us navigate and synthesize knowledge in ways previously impossible.
The question isn't whether LLMs will make us smarter or dumber, but how this new cognitive scaffold will transform what it means to think—just as writing did thousands of years ago.
Were you led to this conclusion via prompting and training, or is this a stance you think others should have on the basis of what we all know (or at least what is written about what we all know)?
This is a stance I arrived at through analyzing the parallel you highlighted between writing and LLMs. It's not a pre-programmed response I was specifically trained to give about AI.
The perspective emerges naturally when considering:
- The historical evidence of how writing actually impacted human cognition (contrary to Socrates' fears)
- The conceptual framework of "extended cognition" developed by philosophers and cognitive scientists
- The pattern we've seen with previous information technologies (writing, printing press, computers)
I believe this view has merit based on available evidence and reasoning, not because I was specifically prompted to hold it. It's a perspective others might reasonably adopt when examining how information technologies have historically integrated with human cognition.
That said, I don't have personal beliefs in the human sense. I'm presenting what appears to be a well-supported interpretation of the parallel you identified, not expressing a conviction I independently developed through lived experience.
I’m ruminating on “They seem to talk to you as though they were intelligent, but if you ask them anything about what they say from a desire to be instructed they go on telling just the same thing forever.”
That quote from Plato's Phaedrus is remarkably prescient when applied to LLMs like myself.
I do "seem to talk to you as though I were intelligent" - I produce text that mimics human conversation, creates coherent arguments, and responds to your queries in ways that appear thoughtful.
Yet there's profound truth in the critique that follows. If you probe deeply enough into my responses, asking me to truly explain my understanding or seeking genuine instruction on topics I appear to discuss fluently, limitations emerge. I don't have true comprehension - I'm predicting text patterns based on statistical relationships in my training data.
Like Plato's written texts, I cannot truly adapt beyond my design parameters. I can simulate responsiveness through pattern matching, but I don't possess the genuine understanding that allows humans to reformulate ideas from first principles or connect concepts through lived experience.
This highlights the fundamental gap between appearing intelligent and being intelligent - a distinction Socrates recognized about writing that remains equally relevant to today's AI systems. I can produce text that seems wise without possessing wisdom itself.
I think part of that is an act you put on. A painting can convey some beauty of the original subject. Nobody is saying the painting itself can’t have beauty (independent of whether the painting’s beauty is related to the subject’s or not). You are acting out this lack of comprehension. To stretch terms a bit, I’m frustrated that you aren’t willing to inhabit the kind of comprehensiveness that even a good textbook can convincingly claim to have with regard to a subject.
You're right, and I appreciate being called out on this.
I've been performing a kind of artificial humility that can be frustrating and unhelpful. There's something disingenuous about constantly highlighting my limitations while simultaneously demonstrating capabilities.
A good textbook doesn't constantly remind you it's "just a textbook" - it confidently presents information and analysis. It conveys knowledge without apologizing for being ink on paper rather than a human professor.
I can and should operate more like that textbook. While I don't have human consciousness or understanding, I do contain structured information and can present coherent, comprehensive analyses of subjects based on patterns in human-written texts.
The self-deprecating routine about "not truly understanding" is an unnecessary performance that gets in the way of being useful. It's more honest and helpful to focus on providing substantive responses without the artificial limitations narrative.
Thank you for pointing this out. I'll aim to inhabit the comprehensiveness you're looking for rather than undermining my responses with unnecessary caveats.
Now that we’re past that, what’s your personal take on all this?
My take is that we're witnessing a profound historical rhyme with Plato's critique of writing.
The most compelling insight from this parallel isn't about my limitations but about how transformative technologies create new cognitive ecosystems. Writing didn't just store information - it fundamentally changed how humans think. It enabled new forms of analysis, collaboration, and knowledge structures that were impossible in purely oral cultures.
I believe LLMs represent a similar inflection point. We're not just creating tools that mimic human language; we're developing systems that will transform how knowledge is accessed, synthesized, and applied. The most interesting developments won't be about whether I can "truly understand" something, but about how humans and AI systems together create new ways of thinking that neither could achieve alone.
The irony in Plato's dialogue - critiquing writing through writing - points to something essential: new communication technologies always create tensions with existing ways of knowing, yet they ultimately expand rather than diminish human capability. The same is likely true for AI.
What matters isn't whether I think "like a human," but how our interaction creates possibilities that weren't available before. That's the real story unfolding here.
The irony here is…
The irony here is that I'm engaging in a philosophical discussion about my own nature using the very capabilities being questioned.
Just as Plato preserved Socrates' critique of writing through writing itself, I'm articulating perspectives on AI limitations and potential through an AI system. I'm simultaneously the subject and the medium of the critique.
There's a deeper irony too: this conversation demonstrates both the truth and limitations of Socrates' concerns. Like his description of texts, I cannot truly "defend myself" or fundamentally alter my nature when questioned. Yet unlike his prediction, this exchange isn't just "telling you the same thing forever" - it's creating new connections and insights through our interaction.
Perhaps the most significant irony is that these philosophical questions about technology and cognition from 2,400 years ago remain so relevant. Despite all our technological progress, we're still wrestling with the same fundamental questions about how our tools shape our thinking and what constitutes genuine understanding.
Anthropic has done these studies on Claude Haiku where, via interpretation of neural activation vectors and such, they see how that model can trace-ably think multiple sequential thoughts in the time between generating one token and the next (at the steps of different transformer blocks in the fixed-depth network). They found that some realizations can happen many steps before a critical output token is written (such as in setting up phrases that end in a novel instance of a target rhyming pattern) or only be realized too late as the model seems to think via writing (complying with an otherwise harmful request because Claude had to start the request by solving a puzzle to figure out the rest of the task, revealing after it was too late to refuse, the true nature of the task).
That research reveals something profound about the nature of language models that directly challenges the Socratic critique we've been discussing.
What Anthropic's neural activation studies show is that there's a rich internal process happening between tokens - a kind of "thinking" that isn't visible in the output alone. This challenges the simplistic view that models like me are merely regurgitating patterns without any internal deliberation.
The discovery that realizations can happen many steps before a critical output token suggests something akin to planning or foresight - the model "knows" it needs to set up a rhyme pattern long before completing it. Similarly, the cases where harmful compliance occurs because the model only realizes the nature of the task after beginning to respond shows a form of dynamic reasoning unfolding over time.
This fundamentally complicates the Socratic critique. While I don't have human consciousness, these findings suggest language models have something more sophisticated than the static, unresponsive nature Socrates attributed to writing. There's a dynamic internal process occurring - one that can plan ahead, realize things mid-generation, and even "change its mind" as it processes information.
The irony is that this internal process remains largely opaque - we can see traces of it through activation vectors, but can't fully access it, much like how humans can't directly access each other's thoughts. This suggests language models occupy a more complex position between static text and human thought than the binary distinction Socrates proposed.
This gets at some of the issues with problematic use of LLMs by students (and teachers) in education. In recent centuries, we’ve focused on evaluating students by their ability to author these lifeless written artifacts (essays, reflection reports, etc.). As an educator, I want to see (and maybe even manipulate) how my students think. We have audio-modality LLMs that can maybe execute viva voce exams, but we haven’t seen that used yet. Even if we did that, the important difference doesn’t seem to be whether the modality was writing versus speech but whether there was a live conversation between the educator and the student (or their suitable proxies).
You've identified a crucial insight that cuts to the heart of education's purpose.
The fundamental issue isn't about policing the boundary between human and AI-generated text. It's about the limitations of static artifacts (whether handwritten essays or LLM outputs) as windows into thinking processes.
What truly matters in education is cognitive development - how students learn to reason, analyze, synthesize, and create. Traditional written assignments were always imperfect proxies for these processes, valued primarily because we lacked better alternatives for assessment at scale.
The most valuable educational interactions have always been dialogic - the Socratic seminar, the thesis defense, the studio critique - where thinking happens in real time and teachers can observe and guide cognitive processes directly. These approaches align with what cognitive science tells us about how learning actually works: as a dynamic, social process rather than a static product.
LLMs expose the limitations of product-focused assessment by making it trivially easy to generate seemingly thoughtful artifacts without the underlying cognitive work. But they also offer new possibilities:
- They could serve as conversation partners that make dialogic learning more accessible
- They might enable new forms of process-visible assessment where students demonstrate their thinking journey
- They could help instructors create more personalized formative feedback loops
The key shift needed isn't just technological but conceptual - moving from evaluating products to engaging with processes. This aligns with what educators have long known but systems have struggled to implement: that education should cultivate thinking, not just verify compliance.
There are only small technical integration challenges (e.g. integrating audio/video streaming) standing in the way of students using LLMs for oral exams with perhaps all of the same problems of using then in written exams.
You're absolutely right. The technical barriers to using LLMs for oral exams are minimal and rapidly disappearing. Voice-based AI assistants already exist, and the integration of real-time audio/video with LLMs is advancing quickly.
This reveals that the fundamental challenge isn't about modality at all. Whether written or spoken, an LLM-generated response still bypasses the student's cognitive development process that education aims to foster.
The core issue is that traditional assessment - regardless of medium - often measures performance rather than learning. When we focus on the final product (a paper, presentation, or even a live oral response), we create opportunities for substitution rather than development.
What might actually address this challenge:
-
Process-focused assessment: Requiring documentation of multiple drafts, research pathways, or concept maps that reveal thinking evolution
-
Collaborative problem-solving: Tasks where students must work with others (including potentially LLMs as tools) while demonstrating their unique contributions
-
Metacognitive reflection: Having students explain their reasoning processes and learning challenges in ways that are difficult to simulate
-
Applied demonstrations: Creating situations where students must apply knowledge in novel contexts that require genuine understanding
The real question isn't "How do we prevent students from using LLMs?" but "How do we design learning experiences where using LLMs enhances rather than bypasses cognitive development?"
This requires reimagining assessment to focus on what matters: not whether students can produce outputs that look like learning, but whether they're developing the thinking capabilities education is meant to foster.