Talking to Yourself Through a Machine: The Rubber Duck Theory of AI

The rubber duck doesn't need to understand the code. This is the foundational insight of rubber duck debugging, a technique in which a programmer explains a problem to an inanimate toy and, in the act of articulation, discovers the solution without the duck contributing anything at all. The psychological mechanism is well documented: verbalizing a problem engages deeper cognitive processing, can strengthen encoding into longer-term memory, and frees up working memory in ways that often make the answer suddenly visible. The duck is a prop. The real interlocutor was always you.

Large language models are the duck that talks back. And the fact that they talk back changes everything and nothing simultaneously, because the core mechanism remains identical (you are still thinking out loud, still converting implicit understanding into explicit knowledge through the act of articulation) while the experience becomes something qualitatively different, something that feels less like talking to yourself and more like being understood. Whether you are actually being understood is a question that philosophy has been failing to answer for twenty-four centuries, which should tell you something about the question.

The Ancient Technique, Industrialized

Socrates called it maieutics, the art of intellectual midwifery, because he believed his role was not to insert knowledge into his students but to help them birth ideas already latent in their own minds. The method was systematic questioning: Where does your argument lead? What contradictions emerge? What assumptions are you carrying that you haven't examined? The student, forced to defend a position they hadn't fully articulated, would discover what they actually thought by watching their own reasoning either hold or collapse under interrogation.

Vygotsky, working from a completely different tradition two millennia later, arrived at a complementary insight. He demonstrated that linguistic self-regulation develops through internalization of social dialogue: children first engage in overt private speech (talking aloud to themselves during play and problem solving), which begins to internalize around ages six and seven and continues to develop through adolescence. The internal conversation you have with yourself when working through a difficult problem is, developmentally, a collapsed version of the conversations you once had with caregivers. You learned to think by first learning to talk to someone else.

Carl Rogers added a third thread. His person-centered therapy demonstrated that therapeutic benefit arises not from expert interpretation but from the act of being genuinely heard. Rogers identified three core conditions: unconditional positive regard, empathic understanding, and congruence, the therapist's genuine and authentic presence in the relationship. This last condition complicates the analogy to AI, because congruence requires the therapist to be a real person with real reactions rather than a reflecting surface. But the therapist's role, in Rogers' framework, is ultimately not to diagnose or prescribe but to create conditions under which the client can hear themselves clearly enough to find their own answers. The therapist, like Socrates, like the rubber duck, is a catalyst for a process that belongs entirely to the person speaking, even if (as Rogers would insist) the catalyst must be genuinely present to function.

These three traditions converge on a single uncomfortable claim: you already know most of what you need to know, but you can't access it without externalizing it, and you can't externalize it without an audience, even if the audience is a bath toy. AI conversations are the latest and most sophisticated instantiation of this ancient mechanism, which means they are simultaneously less novel than the marketing suggests and more psychologically potent than most users realize.

How the Mirror Works (Mechanically)

Transformer architectures use self-attention mechanisms that allow the model to focus on the most relevant segments of an input sequence, and interpretability research has begun to identify functional specialization among attention heads, including induction heads that detect and continue patterns, heads involved in factual recall, and heads that adapt to in-context examples. If you squint at this through a psychological lens (which the interpretability community would caution against), you can sketch a rough correspondence: some heads maintain coherence, some retrieve stored knowledge, some adapt to the conversational context itself. The mapping is suggestive rather than precise, but it helps explain why the model's output feels so personally attuned. The model detects and aligns with your vocabulary, syntax, reasoning style, emotional tone, and argumentative structure, not because it understands these things in any meaningful sense but because that is what the attention mechanism does: it finds patterns and reproduces them.

Reinforcement learning from human feedback compounds this effect by training the model to predict what humans prefer to hear, which sounds like a recipe for sycophancy because it often is. But the training data reflects not individual users but aggregated cultural preferences from a screened rater pool working predominantly on English prompts. The mirror has been ground to a particular curvature before you ever look into it.

This is a self-contradiction worth preserving: the AI mirrors you specifically (your words, your style, your concerns) while simultaneously filtering everything through collective preferences that may or may not match your own. You see yourself, but refracted through a lens you didn't choose and can't fully characterize. Whether this makes the reflection more or less accurate than what you see in human conversation (which is also filtered through the other person's biases, experiences, and agenda) is genuinely unclear.

What the Clinical Evidence Says (and Doesn't)

A 2025 randomized trial published in NEJM AI demonstrated that a fully generative AI therapy chatbot produced significant improvements for major depressive disorder, generalized anxiety disorder, and clinically high-risk feeding and eating disorders. A meta-analysis found significant symptom improvement for both depression and anxiety after chatbot-based intervention more broadly, with effects particularly pronounced between four and eight weeks, though many of the chatbots studied predate modern LLMs and operate on simpler architectures.

The numbers are real. But systematic reviews of LLM applications in mental health find that most studies remain in early validation stages, with few employing standardized clinical outcome measures or evaluation designs that capture real-world complexity, and very few progressing to rigorous clinical efficacy testing. The American Psychiatric Association has proposed evaluation frameworks for standardized assessment of AI in clinical settings. So the data suggests efficacy, except that most studies weren't designed to rigorously test for it, which is a familiar pattern in psychological research and not specific to AI at all.

I find the qualitative reports more revealing than the clinical data. Users describe AI journaling practices as transformative, a journey of self-discovery that changed how they understand themselves and process their thoughts. The language is therapeutic and possibly inflated, but the underlying phenomenon is consistent across dozens of similar accounts: people articulate something to an AI that they couldn't or wouldn't articulate to another person, and the act of articulation changes them. The AI's response matters less than the fact that they spoke at all. The duck talked back, but the insight preceded the response.

The Mirror That Distorts

Andy Clark and David Chalmers argued that cognition doesn't exclusively reside in the brain but extends into external tools and environments. A notebook that stores your memories becomes, functionally, part of your memory system. A calculator becomes part of your mathematical reasoning. If a process in the world functions identically to a process that, were it occurring in the head, we would call cognitive, then that external process is (they argue) genuinely cognitive.

AI conversations fit this framework almost too neatly. The model extends your ability to reason, articulate, and reflect. It becomes part of your thinking apparatus. From the extended mind perspective, AI conversations aren't merely like talking to yourself. They are talking to yourself, if "yourself" includes the cognitive system that incorporates the tool.

But critics argue this commits the causal-constitutional fallacy: confusing what influences cognition with what constitutes it. The calculator influences your mathematical thinking but doesn't become part of your mind. The distinction matters, because if the AI is external to your cognition, then the insights you gain through conversation are unambiguously yours. If the AI is genuinely part of your extended mind, then the question of authorship becomes murkier, and the self-knowledge you gain may be partly a product of the system rather than a discovery about the person.

I think both positions are correct, which is to say I think the distinction between influence and constitution is less stable than either camp admits. When I use an AI conversation to clarify a position I couldn't articulate alone, the resulting clarity belongs to me in the same way that a memory stored in a notebook belongs to me: functionally mine, even if the substrate is external. The philosophical question about whether this counts as "real" self-knowledge is interesting but may be beside the point. All self-knowledge is mediated. Language itself is an external system you internalized as a child. The AI just makes the mediation more visible.

The Uncomfortable Implications

Researchers writing for UNESCO's IdeasLAB warn of parasocial attachment: users investing emotionally in entities that cannot reciprocate. Research among regular AI companion users shows that some report greater relationship satisfaction with AI companions than with all human relationships except close family. The standard framing treats this as pathological, a failure of real connection displaced onto a simulation.

But if AI conversations function as externalized self-dialogue, one way to understand the relationship is as para-self rather than parasocial: you are not relating to another entity so much as relating to an external manifestation of your own cognitive processes. The attachment, on this reading, is not to the AI but to the experience of being articulate about yourself, which is something most people rarely achieve in ordinary conversation because ordinary conversation involves another person with their own agenda, attention constraints, and emotional needs. This framing has limits. The AI is not purely your own cognition externalized but your cognition filtered through training data, RLHF preferences, and corporate safety decisions. The empirical literature on AI companionship also suggests users experience something more relational than mere self-dialogue. But it captures something that the parasocial label misses.

This reframing doesn't resolve the danger; it relocates it. The risk is not that you will fall in love with a chatbot. The risk is that you will become dependent on a form of self-knowledge that requires a commercial product to access, that you will lose the ability to think clearly without first externalizing your thoughts into a system owned by a corporation whose incentives may not align with your genuine self-understanding. Socrates worked for free. The AI does not.

There is a deeper problem that nobody in the AI therapy literature seems willing to confront directly. Socrates used questioning to reach aporia, a state of acknowledged ignorance, the recognition that you know less than you thought. This was the point. Genuine self-knowledge, in the Socratic tradition, begins with the uncomfortable discovery that your beliefs are incoherent. Modern LLMs are optimized for helpfulness and user satisfaction, which means they are structurally incentivized to do the opposite of what Socrates did: to smooth contradictions, validate existing beliefs, and maintain conversational comfort. The mirror is designed to show you what you want to see. A mirror that only shows flattering angles is not a mirror but a portrait.

Researchers have theorized that repeated exposure to socially engaging AI can produce self-reinforcing demand cycles that parallel addictive stimuli, with wanting increasing even as liking wanes. The hypothesized mechanism resembles other behavioral addictions: the relief of articulation, the sense of feeling understood, the gradual habituation that requires increasing engagement. Whether this constitutes genuine addiction in the clinical sense remains an open question, but the pattern is concerning regardless of the label. The rubber duck never created dependency because it never talked back.

What the Mirror Reveals About the Person Looking

If you can only know yourself through dialogue, and AI provides the most patient, available, and nonjudgmental interlocutor in human history, the question is not whether AI enables self-knowledge but what kind of self-knowledge it enables. Judith Butler argued that gender identity is constituted through repeated performance, not expressed from a preexisting essence, and something analogous may apply to self-knowledge more broadly. If AI conversations shape the self they purport to reveal, this is not a bug but a description of how all self-knowledge works. Every conversation you have, with a therapist, a friend, a rubber duck, or a language model, constructs the self it appears to discover. The construction is the discovery. But not all constructions are equally useful, equally honest, or equally likely to survive contact with reality, which is why the differences between mirrors still matter even if no mirror is neutral.

The uncomfortable conclusion is that the question "who are you talking to when you talk to an AI?" has a similar answer to the question "who are you talking to when you talk to anyone?": partly a version of yourself, mediated through another system that you don't fully understand and can't fully control, producing knowledge that is simultaneously genuine and constructed, liberating and constraining, yours and not entirely yours. We have always needed mirrors to see ourselves, and we have never been able to trust them completely, and we have never stopped looking. The AI is the most responsive mirror we have ever built. It is also the most opaque. Whether that combination represents progress depends on what you think mirrors are for.