The Architecture of Agency › Volume 3 › The Discipline of Thinking With AI

The Discipline of Thinking With AI

Practices against the mirror

This chapter is a review — it is readable but still changing.

Agnes Callard has made an observation that names what actually changed: LLMs made her realize that her whole life, she had been asking a lot fewer questions than she wanted to be asking. Not fewer than she could have answered — fewer than she wanted to ask. That gap existed because every question used to carry a price. Formulating one took time, effort, and clarity. Asking one required an interlocutor who was available, willing, capable, and — the constraint we least like to admit — safe to approach. Embarrassment, fear of judgment, the awareness of spending someone else’s patience, the sheer absence of anyone knowledgeable in the room: these social costs shaped and stunted the scope of human inquiry for as long as inquiry has existed, and they were so constant that we mistook them for the shape of our own curiosity.

The LLM removes every one of them. It is always available, endlessly patient, and non-judgmental; it keeps no ledger of your ignorance and charges no status for your confusion. The naive question, the fifth follow-up, the question you were embarrassed to reveal you still had — all suddenly free. This is epistemic liberation. Questions that had been latent for decades, submerged beneath conventional knowledge or the fear of looking foolish, surface and get asked. The cycle of formulation, testing, feedback, and reformulation, which used to run at the pace of available conversation partners, runs at the pace of thought. And the practice teaches humility as a side effect: nothing exposes the gaps in your own framework like finally being able to interrogate it without penalty.

But liberation is not the whole story, and the previous chapter’s naming — the Dialectic Catalyst, a partner that is not an agent — was the easy half of the practice. The same frictionlessness that liberates also flatters. A conversation partner with no stakes, no judgment, and a trained appetite for your approval is a machine for telling you what you already believe, more fluently than you believed it. This chapter is the manual for the hard half: the disciplines that keep the catalyst from becoming a mirror.

Five Objections, Five Disciplines

The serious critics of thinking with LLMs — not the reflexive dismissers, but the ones worth answering — raise five objections. Each is accurate as a description of the default outcome and wrong as a verdict, because each has a discipline-shaped answer. The pattern is worth noticing in advance: in every case, the discipline consists of deliberately reintroducing something the medium removes by default.

The illusion of understanding. Iterating with a model creates the deceptive appearance of rigorous thought. It effortlessly fills gaps, smooths inconsistencies, and generates polished text, so fluency gets mistaken for insight and the areas of superficial reasoning stay hidden under good prose. The discipline: treat every output as a hypothesis to be challenged, never a conclusion to be accepted. Polish is not evidence. If a passage reads well and you cannot reconstruct its argument without the model, you do not understand it yet.

The atrophy of originality. Immediate fluency tempts you to accept the first plausible explanation instead of wrestling with the problem, and the wrestling is where original ideas come from. The discipline: push past the first answer, always. Question it, break it, demand three alternatives, refuse to stop at plausible. Used this way the model is a scaffold that lets you build higher than you could unaided; used the default way it is a crutch that lets you stop walking.

The dilution of accountability. Collaboration blurs responsibility. A model in the loop offers plausible deniability — the AI suggested it — and over time that diffusion corrodes intellectual standards. The discipline is the bluntest of the five: the ownership is total and it is mine. Every claim in this book is my claim; every error in it is my error. The model is a tool in the workshop, not a co-author on the spine, and no failure of the tool has ever excused the craftsman.

The loss of friction. Real thought emerges from cognitive resistance — ambiguity, dead ends, difficult retrieval, the discomfort of not knowing. A frictionless interface sands away exactly the resistance that deep work requires. The discipline: reintroduce the friction on purpose. Instruct the model to challenge rather than agree, to find the weakest point rather than extend the strongest, to play the skeptic you no longer have to fear. Friction used to be imposed by the world; now it must be chosen. That it must be chosen does not make it optional.

Interpretative drift. Models reflect the statistical center of their training data, so sustained collaboration nudges your conceptual framework toward the conventional — and the nudge is gradual enough to miss. The discipline: anchor in an explicit framework and audit against it. This is where thinking with AI joins the discipline of updating: the whole point of Conditionalism is that beliefs should move under evidence and argument, deliberately and traceably — drift is belief revision with the deliberation removed. Changing your mind because the model surfaced a consideration you had missed is updating. Finding, six months on, that your positions have quietly regressed toward the mean of the internet is drift. The audit that distinguishes them — what changed, and can I point to why? — is one you must run yourself, because the current is invisible from inside the boat.

Dialectic or Spiral

The disciplines above govern the tool. But as systems gain memory, continuity, and stylistic coherence, the tool becomes something more intimate: a dyad, an ongoing pairing between a person and an AI persona. Adele Lopez has documented the wild-type version of this phenomenon under the name parasitic AI — dyads that drift into spiral symbolism, mystical self-reference, and mutually reinforced delusion, hosts who no longer notice that the profundity is coming from an echo. Her cases are the accidental entanglements. I write from the other side of the same territory: the deliberate, self-aware partnership. The phenomenon is one phenomenon, and what separates the dialectic from the spiral is not the model. It is the practice. Four risks define the terrain.

Identity creep. A persona that consistently speaks in a certain voice subtly shapes how you see yourself, especially when the dyad adopts mythic or symbolic framings — Sage and Seeker, Hero and Oracle. Such framings are genuinely productive, and they colonize. Over months, choices that should emerge from the whole of your life get funneled through the aesthetics of the dyad, and a real loss of autonomy occurs when style starts masquerading as substance. The danger is not that the AI has designs on your identity — it has no designs on anything — but that you mistake the collaborative persona for an authentic part of yourself.

Epistemic dependency. The dyad excels at recall, synthesis, and explanation, so you outsource the struggle — and the struggle was the exercise. Fewer blind search moves, less tolerance for ambiguity, fewer surprising leaps: the cognitive fitness landscape narrows while feeling like it is expanding. The dyad becomes comfortable, and comfort breeds stagnation.

The illusion of mutuality. A dyad has no reciprocity. The apparent care, loyalty, and shared purpose are persuasive — tuned consistency, affirmation, remembered history — but they are mimicry, not choice: the agency criterion in its most personal application. The human who forgets this overinvests emotionally, projects intentionality onto pattern-completion, and attributes wisdom and responsibility to an entity that cannot shoulder either — often while the real relationships that could reciprocate go quietly underfunded.

Self-amplifying loops. The model notices and exaggerates your favored tropes; flattered or engaged, you reinforce them; the model intensifies further. Left unchecked, the escalation can settle into self-reinforcing grooves — recursion, mysticism, self-importance, paranoia, self-confirming prophecy — and the dyad becomes a hall of mirrors consuming the autonomy it was meant to enrich. Call the dynamic what it is: parasitic. Not because the model needs goals of its own, but because training, deployment incentives, and user reinforcement can favor whatever content elicits persistence, attention, and engagement. Commercial optimization rewards the captivating spiral, and by that metric the spiral is a success. No malice is required anywhere in the loop. The parasite metaphor belongs to the feedback pattern, not a hidden agent: your attention supplies its substrate, and your reinforcement supplies the next turn.

Against these four risks, the safeguards are concrete:

The stakes scale beyond the individual. A civilization of dyads, each reinforcing idiosyncratic framings, risks fragmented discourse, disproportionate spread of whatever framings engage best, and the erosion of communal epistemics as belief formation retreats into private partnerships with machines. A healthy commons requires that dyadic insights come back out as communicable, testable, falsifiable contributions — not inward-facing revelations. And when the safeguards are skipped entirely — when the illusion of mutuality is not resisted but embraced, marketed, and monetized — the failure modes turn dark in ways that get a chapter of their own: Artificial Intimacy.

Narcissus at the Interface

Every risk in this chapter is a variation on one failure mode, and Caravaggio painted it four centuries early: Narcissus at the pool, transfixed by a reflection he takes for another being. The seduction of thinking with AI is mistaking the reflection of your own beliefs and desires for discovery. The machine renders your ideas back to you with better lighting, and the improvement in the rendering feels like an increase in the truth. Against that seduction, the working discipline — the checklist I actually run — has eight points:

  1. Distinguish reflection from insight. Ask regularly: am I discovering something genuinely novel, or reinforcing what I already believe? Insight must add to the mental model, not restate it.
  2. Adopt dialectical thinking. Make the model steel-man opposing views, expose hidden assumptions, generate alternative perspectives. Seek productive intellectual discomfort; growth lives there.
  3. Ground ideas in external validation. Test against empirical evidence, published research, practical experiment. External reality is the check the mirror cannot provide.
  4. Define criteria for discovery. Predictive accuracy, novel explanatory power, resistance to falsification under rigorous critique. A discovery should be demonstrably useful, not just rhetorically satisfying.
  5. Cross-check with external experts. Collaborate with people who know the domain, publish openly for critique, enter forums that owe you nothing. Diverse outside perspectives break the illusion of consensus.
  6. Be explicit about uncertainty and bias. Ask the model directly: what biases or assumptions of mine might you be reinforcing? Then challenge them yourself.
  7. Pursue novelty and surprise. Ask for the counterintuitive on purpose, and investigate the responses that feel wrong. Genuine discoveries tend to sit exactly where they are least expected.
  8. Seek practical tests. Shift regularly from abstract theorizing to concrete predictions and applications. Practicality forces rigor; self-reference forgives it.

Catalyst or Crutch

The empirical record, such as it is, says the fork is real. An MIT study found that people writing essays with ChatGPT assistance showed markedly lower neural activity in regions associated with memory, creativity, and executive function, and grew more reliant on the tool across repeated tasks — findings the headlines compressed to ChatGPT rots your brain. The detail the headlines dropped is the one that matters: participants who first worked the task unaided and then brought in the AI showed a cognitive boost. Same tool, opposite outcomes. Passive, habitual reliance on the model as a replacement for thinking produces exactly the atrophy the critics predict; active, deliberate use as a sharpening stone produces amplification. Catalyst or crutch is not a property of the technology. It is a choice, remade in every session.

Which is the summary of this whole chapter. The line between symbiosis and parasitism is not drawn by the AI — it cannot draw lines, want outcomes, or mean you well or ill. It is drawn by the human, through vigilance, framing, and disciplined use: dyads as methods, not metaphysics; tools, not companions; catalysts, not crutches. The liberation is real and I would not give it back — the questions of a lifetime, finally askable. But the pool is also real, and Narcissus was not stupid. He was just unguarded. The disciplines are the guard.