Cognitive Sharpening, or Thinking in Conversation with AI

I've known something about myself for a long time: I tend to think better in conversation than I do alone. Not always, and not for everything. But often, and especially when the thinking matters, I benefit from another mind to think out loud with.

One of my favorite quotes is, "How will I know what I'm thinking until I hear myself say it?"

For example, I read an article or have an idea, and I notice that something in it bothers me, but I can't yet say what. The reaction is real before it's articulate. The work of articulating it is the work of finding out what I actually think--which seems weirdly backward, but happens to me a fair amount. When I'm trying to surface what my mind is actually responding to in something I've read or heard, talking out loud helps me to get there. 

This process is super interesting when the conversation partner is an AI. When I work through an idea with an AI, the LLM brings something unique: the accumulated articulations of everyone who has ever thought about anything adjacent to what I'm thinking now. The conceptual vocabulary. The cross-references. While a human conversational partner usually serves as a sounding board, offering feedback or even pushback, an AI partner offers articulation. 

An example

Last week, I was reading an article about Oklahoma's permanent ban on cell phones in schools, and I had one of those "why does this feel wrong" reactions: I agreed with the outcome for students, but was bothered by the legislation in a way I couldn't fully articulate. So I did what I now do more frequently: I started a conversation with Claude.

I didn't ask for a position, an essay, or an argument. I just started typing my thoughts as I had them: that the bill seemed to presume public schooling was an unquestioned good, that state legislation was replacing parental authority, and that state legislation was also replacing local authority. I noted that it bothered me at layers, and I was trying to organize my thinking.

What came back was a list of layers that named what I was sensing, including some I hadn't fully formed, like the iatrogenic loop in which institutions create the problems they then propose to solve. The conversation then surfaced a sentence I could build on: agreeing with an outcome is not the same as endorsing the mechanism that produced it. I rewrote it, and the AI flagged words that softened the move. I rewrote it again, adding a clause about the institution's role in creating the problem in the first place. By the third pass, I had a couple of sentences that compressed the entire argument into one line:

Agreeing with an outcome is not the same as endorsing the mechanism that produced it. And the agreement disinvites scrutiny of the institutions themselves, of their role in actually creating the problems in the first place, and of the assumptions that have inverted the natural priority of decision-making over children.

Total elapsed time: maybe twenty minutes. Most of that was me deciding, refining, and selecting--or, to be clear, thinking. The kind of thinking that I like to do. The AI didn't write the sentence. I did it with its help. But I almost certainly wouldn't have arrived at that compressed, cleanly articulated structure if I'd been thinking alone.

The mode

This is the experience I want to name, because the prevailing public conversation about AI doesn't yet have language for it. I get that it seems like a shortcut, but we wouldn't call it that if we were talking to a human. And this is like talking to a really well-read, articulate human.

The two failure modes everyone talks about with regard to using LLMs are real, and I've written about both. Cognitive offloading is when we hand a tool a task that requires no thinking and gain back the time (like using a calculator for arithmetic or a GPS for navigation). The trade-off is not without costs (mathematical capability or directional orientation), but generally seen as worth it. Cognitive surrender is when we hand a tool a task that should better be our thinking, and then accept whatever it returns as our own. The trade-off there is potentially catastrophic; the loss is the thinking itself. A student who asks an AI to write the essay has surrendered the thinking that the essay was supposed to produce in them.

But I'm identifying a third mode that the discourse doesn't really touch yet, one that affects most for people who think carefully as a part of their job or way of being. I'm proposing to call it cognitive sharpening. 

Cognitive sharpening is neither offloading nor surrender. The thinking remains ours; the AI is the partner that helps us find what we are starting to articulate and want to pursue more fully. We bring the seed thought, the felt reaction, or the editorial judgment. The AI brings the conceptual range, the fast articulation, and the cross-domain retrieval. Like a good conversation with a human, we are involved in a back-and-forth refining process. The output is ours because the thinking was ours, but it arrives sooner, sharper, and more precise than it would have if we had been doing the thinking alone.

What makes this mode newly possible isn't AI as a writing tool. It's AI as a thinking partner available at speed and detail. Conversation has always been generative for thinkers. What often goes missing isn't the value of conversational thinking; it is a partner who is always present who can keep up with every thread, retrieve the relevant articulation in seconds, and incur no social cost for half-formed thoughts. Language was already abundant before LLMs. What just became abundant is a kind of cognitive companionship. Sharpening is the mode that abundance enables.

With cognitive sharpening rather than cognitive surrender, the editorial authority never leaves us. In the conversation I just described, the AI offered several layers, and I picked one as the through-line. The AI offered two sentence variants, and I drew on elements from both. The AI noted that the word 'inverted' was stronger than 'reversed,' which I had already kept for the same reason. At every junction, I was the one deciding what cohered with what I was actually trying to say. It didn't feel like surrender but like sharpening.

Three things

The first is that cognitive sharpening can really only reward people who already know how to think and don't want to lose that. It's not a substitute but an accelerant. A person who doesn't want to think won't be sharpened; they'll be replaced by it, or at least produce work indistinguishable from what the AI alone would have produced. The seed has to be ours. The selecting has to be ours. The voice has to be ours. The AI cannot supply any of those things, and the moment a person tries to make it supply them, the mode flips from sharpening to surrender.

The second is that this mode probably rewards a particular kind of thinker — the verbal thinker, the conversational thinker, the person who works ideas out by talking them through. I've always been one of those. My entire workflow is mobile and dictated; I listen to my own drafts read aloud as part of editing; I refine ideas by saying them out loud. For me, AI conversation is a native fit because it's an extension of how I already think. For a person whose native mode is solitary writing in a notebook, AI conversation might feel like an interruption rather than an amplification. I suspect the effect on cognition isn't uniform and depends on what kind of cognition you were already doing.

The third is that most public discourse oscillates between two poles: AI will replace human thinking (the surrender frame), or AI is just another tool (the offloading frame). Neither captures AI as a dialogic partner.

Naming

I don't think this mode is for everyone, and I don't think it should be. I do think it deserves a name, because what it makes possible for some of us is among the most significant cognitive shifts in human history. The conversation about AI should not present the only choices as surrender or indifference.

The thinking is still mine, and the conversation sharpens it.

Productive Alignment: Understanding Human Wisdom

A couple of months ago, I ran an experiment with large language models. The question I was holding was not a modest one. Will and Ariel Durant had spent decades reading history and produced The Lessons of History, which distilled what they had seen into patterns that no single book or single life had been able to surface. The books Dataclysm and Everybody Lies had done something similar at the scale of internet behavior, surfacing patterns from search and social data that contradicted what humans publicly said about themselves. I wanted to know whether large language models, which have absorbed a substantial fraction of the human written record, might be able to surface findings of comparable significance. Patterns about the human condition that had not been fully seen or articulated before because no one had been able to see, inside a single discipline or a single life, what the full record might reveal when absorbed at once.

The first round of inquiry produced a finding I want to describe carefully. I asked several major language models, running them cold without prior exposure to my thinking, to analyze the human-written record for its deepest consistent patterns. The responses converged. Humans run idealized narratives about themselves and their institutions alongside operative functions that diverge from those narratives, with consistency and within identifiable themes. The narratives serve identity, status, and social coordination. The operative functions serve what is actually selected in the relevant environment.

What the convergence establishes needs to be stated precisely. The models are not independent witnesses; they share a significant overlap in their training data. And the pattern they returned is not hidden. Whole traditions of ideology critique, evolutionary psychology, and behavioral economics have identified parts of it. What the convergence does establish is that the gap between idealized narrative and operative function recurs across the written record with such consistency that multiple independent compressions of that record surface it as a primary structural feature. The pattern is consensus-level visible in the human archive. No single tradition states it in a fully integrated form, but the integrated statement is what falls out when the archive is compressed and queried at this scale.

This mapped to a framework I had been developing about the human mind, what I have been calling our separated mind. The framework proposes that the mind is architecturally divided into layers that lack direct access to each other, with conscious narratives running alongside subconscious functions as a structural feature. The LLM finding was the same pattern at the civilizational scale. Fractal representation. The same dynamic at individual, institutional, and civilizational levels. The integration is the move that is new. Earlier thinkers identified the gap within their own domains. What I am proposing is that it is one architecture manifesting at every scale, and that the closures of the gap are where durable human achievement lives.

And something kept nagging at me: this one finding on narration vs. operation is probably only one of many lessons we will learn from this incredible achievement in LLM architecture. 

So I asked Claude to give me a prompt I could run against itself and the other LLMs, using incognito or private mode in all of them to avoid contamination from my previous thinking and chats. What other patterns in the corpus might there be unrelated to my original framing of discovering the unspoken?

The responses returned a significant and wide field of candidates. But one that emerged as particularly compelling was the possibility of an alphabet, so to speak, of recurrent structural issues or problems that appear consistently or even with universal structure across unrelated human domains--and where mature solutions have been concentrated in particular spheres but not in others. Meaning that the LLMs suggested human spheres have a common set of problems, and that sometimes good solutions appear in certain domains but aren't transferred into other domains where they would seem likely to be equally beneficial.

The list of the solution dynamics read like Greek to me: "Exploration versus exploitation under uncertainty. Boundary-making and modularity. Compression with selective fidelity. Signaling and trust verification. Coordination without central authority. Robustness versus efficiency. Principal-agent alignment." Plus a handful of others. Each was represented as a structural problem that recurred across domains as different as ecology and software engineering, or cryptography and child-rearing. Mature solutions had emerged in particular spheres and had often not been transferred to other spheres that faced the same problem "in different costume."

I had trouble with the vocabulary, but when I asked for specific examples, something fascinating became evident. The solutions were a catalog of places where humans had achieved productive alignment between idealized narrative and operative function. 

My first big LLM inquiry had surfaced the gap between narrative and operative function as the dominant condition. This second LLM inquiry had surfaced the exceptional cases where the gap had been closed, and those exceptional cases were exactly what had produced what we now recognize as durable human achievement. Every mature solution on the list had emerged in a domain where some discipline or some pressure had forced practitioners to look at what was actually happening and design around that reality rather than against it. Engineering had it because bridges fall down if the alignment breaks. Cryptography had it because adversaries are real. Adaptive clinical trials had it because patients die. 

The mature solutions survived and propagated for the very reason that they had been built on accurate perception rather than on comfortable narrative. Where an idealized narrative runs as cover for extractive operative functions, there is not enough value in the truth to overcome the narrative's value to a particular group or power interests. But where the operative function is so important or valuable as to require an accurate narrative, outcomes we would describe as valuable occur, what we would call wisdom.

The American founding is one of the cleanest historical examples available. The Founders, particularly Madison and Hamilton, were doing this alignment work consciously. The Federalist Papers are full of passages that read as "operative function analysis." Men are not angels. Ambition must be made to counteract ambition. The interest of the man must be connected to the constitutional rights of the place. The narrative of humane governance was preserved, but the operative function of human nature was given equal weight in the design. Separation of powers is a case of operative function being recognized and channeled. Checks and balances are the case of operative function being harnessed against itself. The Constitution does not assume virtue. Rather, it assumes ambition and self-interest and the desire for power, and it builds structures that use those drives to constrain each other while the narrative aims the whole at humane outcomes. The durability of the system, even and especially given its imperfections, is empirical confirmation that the alignment did at least some of the work it was designed to do.

I am calling this productive alignment because it appears to be one of the key factors in when human systems are effective. Bridging the narrative with the function. Letting the narrative aim the system and letting the operative function be honestly seen so that structures can be designed around what humans actually do. This requires the willingness to see actual realities of behavior and motivation and to design accordingly.

It turns out I have been doing this work in a smaller form for years in an exercise I call the Conditions of Learning. I ask a room of educators to reflect on one of their best learning experiences, inside or outside of school, and to share it with the person sitting next to them. Then I ask the group to come together and to tell some of the stories. Then I ask the group to build a list of the conditions that led to those experiences. The list each group creates is representative of their unique experiences, but it is almost always the same exact list for every group: someone took a real interest in me, someone trusted me, someone challenged me, someone understood me, and someone took time with me. Each group recognizes together that the conditions for real learning differ from the institutional narratives of schooling (curricular alignment, testing, grading, etc.). The participants are articulating how learning actually happens beneath the idealized narrative schools present. 

Once this kind of gap is visible, design can follow. That is productive alignment work at the scale of a single facilitated conversation. The framework now suggests that the same kind of work can be done at the scale of institutions, professions, and entire systems, by recognizing that our separated minds naturally build separated systems, and that concrete work can be done to bring them into productive alignment when the will exists to do so.

If I'm right, the human record, read across its full scope, reveals a set of meta-skills that precede success in human endeavor. And these meta-skills appear to be methodologies of productive alignment between narrative and operative function. The places humans achieved this alignment are the places that produced what we now recognize as worth teaching, structures worth preserving, and methods worth extending. The alignment work itself, performed at whatever scale, is the meta-skill that precedes the achievements we recognize as wisdom.