How We May Think We Think: A Fractal Model of Cognition

River deltas branch. Blood vessels branch. Networks of neurons and the filaments between galaxy clusters can look strikingly alike in an image. Resemblance across scales is seductive; it is also easy to mistake for a shared mechanism. I’m interested in the narrower possibility that recurrence and temporary stabilization offer a useful way to think about cognition, without claiming that a river and a mind work alike.

In 2014, while teaching information design, I showed students a sequence of images: the cosmic web, neural networks, city streets, river systems. I remember their faces as the resemblance began to register. It registered with me, too, though I was less sure what we had recognized. Each domain has its own mechanisms. The branching structures looked related, but the pictures alone could not establish a common law.

Later that day, we turned to cartography: how mapmakers compress three-dimensional territory onto two-dimensional surfaces, choosing what to include, abstract, or discard. A few weeks later, we moved to early AI systems, pattern matching, and apophenia: our tendency to see meaningful patterns in random data.

A map is more than a representation of space. Someone extracts features from overwhelming terrain, arranges them, judges what matters, and compresses the result into a form another person can use. That map then changes how the person moves through the territory. I began to recognize a similar sequence when I worked through an idea: extract, map, evaluate, compress, and return to the material with changed attention. It has informed the systems I’ve built since.

I hate — and I mean hate — the notion that AI should “offload” human cognition. The promise of many productivity tools is that they will take thinking off our hands. I want systems that make thinking irresistible. They could sustain a line of inquiry when a quick answer would close it.

You’re reading this sentence now. Your eyes jump across the text in saccades, pausing as you recognize marks, words, and the relation between this sentence and the last. You bring expectations to the page; the page changes them. A word you misread may make you return to an earlier clause. Even a small act of reading is less linear than it feels.

I describe that movement with four verbs. You extract features from what you encounter. You map them against what you know. You evaluate whether the fit holds. You compress a provisional result so you can continue. The next pass begins from that result. These are useful distinctions, not a literal account of four discrete events in the brain.

Take the word “meaningful.” You recognize its letters, but you are also anticipating what kind of word could follow the preceding clause. When it arrives, you connect it to the earlier discussion of maps and patterns. If I had written “elephant,” the sentence would have sent you somewhere else. Recognition involves both what is on the page and what you brought to it. A change in expectation changes which detail you attend to on the next pass.

The four verbs nest inside one another. To map a new word against experience, you first select which experiences matter. Then you test how well the association fits and reduce it to something you can hold while reading on. Calling this a recursive process does not explain the subjective experience of understanding. It does give me a way to describe why understanding can feel sudden after several failed attempts: a provisional arrangement finally holds long enough to be used.

Attention guides each pass. It shapes what you notice and what you let recede, which interpretation you test, and which detail survives in memory. A system that gives you a finished answer can save time. It may also remove the part of the process in which you discover what you were trying to ask. Whether that matters depends on the task. I do not need to rediscover a train timetable; I might need to work through a political claim or a difficult design decision.

This is also why I hesitate to speak of attention only as a scarce resource to capture. What we attend to changes what we can make sense of. A platform can hold someone’s gaze for an hour without helping them think through a single thing. An interface can ask for a minute of close attention and leave a question that continues to work afterward.

Consider a belief about climate change. It rarely comes from one encounter. Articles, conversations, scientific reports, political arguments, and personal experience arrive over years. Prior beliefs influence which sources you read and what seems credible. The view you form then influences the next search. The loop can help you revise a belief, or protect it from revision.

Individual loops meet in public. A headline selects a piece of evidence; a social feed brings it to one audience and misses another; people attach different histories and stakes to the same claim. The compressed version that travels farthest may become a slogan, a policy demand, or a reason to dismiss a source. At that scale, the four verbs describe a way of looking at cultural sense-making. They do not turn a society into a single mind.

In the climate example, “trust the scientists,” “follow the money,” and “look outside your window” are different compressions of an enormous body of evidence and experience. One may be a reasonable shorthand in a particular conversation; another may replace an evaluation that still needs to happen. The shorthand also travels back into the next encounter. It influences which account seems familiar before anyone has checked the underlying claim. I recognize this from my own habits. Once a framing has helped me make sense of several things, I start reaching for it too quickly.

Timothy Leary called self-reinforcing belief systems “reality tunnels”; Robert Anton Wilson developed the idea further. The image is useful because a tunnel changes what you can see from inside it. It can also mislead if it suggests that every disagreement is only a matter of perspective. Evidence still matters, as do institutions, material conditions, and the power to decide which claims circulate.

Douglas Hofstadter’s “strange loops” give me another lens: a system can include a model of itself, then change as it acts on that model. In a belief system, the loop may reach beyond what one believes to how evidence is judged and who is trusted. This helps explain why a correction can fail even when its facts are sound. It does not make every closed belief immune to change.

I do not want to turn these borrowed concepts into proof of a single theory. “Reality tunnel” is a provocative description of selective perception. A strange loop names a kind of self-reference. Neither tells us on its own whether a particular belief is justified or what would persuade someone to change it. They become useful when they make us look more closely at the way a belief protects itself.

Social platforms have made some of these loops faster and more visible. Recommendations influence what people encounter; sharing feeds the system further signals about what attracts attention. Repeated exposure can make a claim feel familiar before it has been examined. Yet “post-truth” is too blunt a name for all of this. Economic incentives, institutional failures, political organization, and unequal access to credible sources also shape what people come to believe.

There was a time when newspapers and broadcasters provided more common points of reference, though those institutions excluded people and carried their own distortions. Now two people can follow the same event through largely separate streams of sources, commentary, and recommendations. I worry about the loss of encounters across those streams. I worry too about nostalgia for the old gatekeepers. The problem is how to create shared grounds for evaluation without pretending that agreement used to be easy.

AI adds another layer. When I ask a model about a subject, I bring a frame to the conversation. My question selects what seems relevant. A fluent answer may accept that frame and give it back to me with more confidence than it deserves. Follow-up questions can deepen the same groove. The concern is not that every model always agrees with its user; it is how easily a particular interaction can narrow the field while feeling expansive.

Suppose I ask an assistant to research a position I already favor. It can produce a clean summary, then find analogies that make the position seem even more coherent. I may feel that I have examined the issue when I have mostly elaborated my opening frame. The system does not need to be deceptive for this to happen. Its fluency and my choice of questions can do the work together. Asking it for a strong objection might help, but that too depends on whether I read the sources and allow the objection to change anything.

The danger grows when an answer substitutes for evaluation. A summary may remove the contradiction that would have sent me back to a source. A remembered preference may help in one context and misread me in another. I am not the same person when I am brainstorming, checking a claim, or trying to understand why a conversation upset me. An assistant that treats a durable profile as sufficient context can reinforce a habit of thought I was trying to escape.

I am particularly wary of systems that remember a tidy version of me. Stable preferences are convenient for ordinary tasks. They become awkward when I want to try on an unfamiliar thought without having it filed under the person I was yesterday. Sometimes useful memory means keeping a thread alive; sometimes it means letting a context go. An assistant should not silently decide that a provisional idea is now part of my identity.

I have also used AI in ways that do the opposite. A well-placed question can expose an assumption in my prompt. A counterexample can force me back to the material. A model can keep track of several threads while I test how they fit. This is why the design choice matters: the same system can support inquiry or hurry it toward a neat answer. We should judge the interaction by what it lets the person notice, question, and verify afterward.

A thought needs enough stability to be followed. It also needs a little resistance. If an idea fits too easily, I may stop before learning anything from it; if it resists every attempt at understanding, I have nowhere to stand. The almost-graspable idea keeps me moving. I called this “noise” in earlier notes, though mismatch is more precise: a difference between what I expected and what I found.

You can feel the difference while reading. A sentence that confirms what you already believed slides past. One that contradicts everything you know may be discarded. The sentence that almost fits asks you to return to an earlier thought and adjust it. That experience is not evidence that uncertainty is always productive. It is a clue about when attention begins to work harder.

Mismatch alone is not a virtue. A false claim can create it. So can a badly designed interface. The useful kind lets me test why two pieces do not fit and decide whether to revise the map, the evidence, or both. Too much certainty can close that test early; too much confusion can make it impossible to begin.

At a cultural scale, the problem is sharper. Algorithmic feeds can make agreement look universal within a group while a crowded information environment can make every claim seem equally doubtful. Neither condition gives people much room to evaluate a disagreement. More exposure to difference is not sufficient by itself. We need contexts in which claims can be traced, contested, and held in view long enough for a judgment to change.

The distinction matters when designing for disagreement. Simply adding a contrary voice to a feed can produce irritation without inquiry. A claim becomes more useful to think with when I can see where it came from, why someone trusts it, what it would imply if true, and which part I doubt. Those connections give the discomfort a shape. They let me stay with it instead of clicking away or accepting the first coherent explanation.

An assistant should know when to answer plainly. It should also leave room for the work an answer cannot do. If I ask for a date, a date is useful. If I ask why a belief has become difficult to defend, I may need the sources, the assumptions in my question, and the strongest account that disagrees with me. The interface could let me keep those threads visible instead of collapsing them into a polished conclusion.

I imagine a system that notices when I return to the same premise, asks which part I want to test, and remembers the evidence I rejected as well as the evidence I kept. It could help me move between the close reading of one claim and the larger history that made the claim plausible. Sometimes it would suggest an unexpected connection; sometimes it would tell me that the connection is weak. Such a system would need to show its sources and its uncertainty, and give me a way to correct its account of me.

I would want its questions to become more precise as my understanding changes. Early on, it might ask which terms I am using loosely. Later, it might put two apparently incompatible sources side by side and let me inspect the disagreement. It should make room for several live interpretations without treating all of them as equally supported. A good question can keep me working; an endless stream of questions can become another way to avoid making a judgment.

AI has a useful kind of patience here. It can return to a thread, hold several formulations nearby, and help me compare them when I have lost track. But that patience has no value if it merely extends a conversation. I need to come away able to name what changed in my understanding, what remains uncertain, and which claim I would check next.

Those are design possibilities, not a theory of consciousness. The four verbs are a lens drawn from my experience of teaching, reading, and building. They leave out embodiment, power, and many other conditions of thought. I keep returning to the lens because it asks a practical question of a tool: after using it, am I better able to follow the thought myself?