Externalized Metacognition as Adaptive Independence

01The freed-attention gap

Reser’s hierarchy disengagement hypothesis offers a conditional evolutionary mechanism. Under some ancestral conditions, reduced investment in status competition lowered the attentional, temporal, and emotional costs of social participation, freeing capacity for independent subsistence and other fitness-relevant activity. The hypothesis is carefully bounded. It does not require continuous solitary living. It does not imply that autistic people uniformly fail to recognize hierarchy. It notes that direct evidence for a hierarchy-specific allocation advantage remains absent. It frames the evolutionary explanation as conditional on ecological opportunity, continued cooperation, and the consequences of reduced rank monitoring.

What the hypothesis does not specify, and was not designed to, is the modern fate of the attention it describes as freed. In an ancestral ecology the released capacity had natural sinks: food procurement, tool maintenance, environmental learning, childcare, rest. In a present-tense ecology where independent subsistence is not the binding constraint on most people’s survival, the released capacity has no automatic destination. It does not disappear. It goes somewhere, into deep interests, systemizing, long-form inquiry, model-building, and then, frequently, it goes nowhere that others can reach.

This essay takes up that downstream question. Withdrawal from hierarchy can free attention. It can also leave the person without a reliable channel to preserve, inspect, or transmit what that freed attention produces. The cost is not only social. It is cognitive. Deep work that is never externalized cannot be returned to, repaired, or extended across the distance between one session of thought and the next. It is subject to the same limits Reser’s own iterative-updating model identifies within a single mind: working memory is capacity-limited, states are displaced, and continuity is an achievement that requires active maintenance.

Reser’s iterative-updating account proposes that thought proceeds through a recurrent sequence. Some working-memory representations are retained, others are displaced, and newly generated representations are added, producing a stream that remains continuous while changing. The account is a claim about a property of a single cognitive system sustained across successive states. The substrate claim in this essay is functional, not biological. It does not assert that the substrate is conscious or that it replicates working memory. It argues that the same selective-retention-and-revision function can be carried, at larger scale and longer timescales, on a persistent external surface.

02Continuity within, continuity across

This essay draws on a body of practice I call Externalized Metacognition Through a Conversational Substrate (EMTCS). It concerns a different but adjacent problem: continuity across systems, sessions, and tools. A conversation with an AI collaborator, persisted as an inspectable record, can hold context that a single working-memory state cannot. It preserves the thread of an unfinished argument across days or weeks. It allows a person to revisit an earlier state of their own reasoning, detect where it broke, and repair it. It externalizes the iterative-update function: the record retains, displaces, and adds representations, but it does so outside the skull, where the person can read them back.

Internal continuity is limited by capacity and by the displacement of states. The external substrate provides a persistent, addressable, revisable surface on which the same kind of selective retention and revision can occur at larger scale and longer timescales. The internal system and the substrate are not rivals. They are two layers of the same continuity problem. One is bounded by biology. The other is bounded by what a person chooses to persist and by the durability of whatever holds it.

03The prosthesis claim, stated carefully

Suppose some people allocate attention away from hierarchy maintenance and toward independent model-building. Then an externalized conversational substrate can function as a social-cognitive prosthesis for those people specifically. It gives them a place to hold context, revisit unfinished thought, preserve authorship, and make deep cognition shareable without forcing it into status-paced performance. The substrate does not replace the social channel. It removes the requirement that the social channel be the only channel through which cognitive work becomes real.

This is where the adjacency to Reser’s hierarchy disengagement becomes more than topical. The disengagement hypothesis describes a population whose attention is, in a specific and limited sense, reallocated. A prosthesis claim follows: a population whose cognitive surplus does not flow into rank maintenance may benefit differentially from a tool that captures the surplus and returns it in a form that survives the session. A persistent substrate is a candidate for that tool. It is adaptive not in the evolutionary sense Reser is careful to leave open, but in a present-tense, functional sense: it adapts the output of freed attention to a shape that can be carried forward.

The prosthesis is not unconditionally beneficial. A persistent conversational substrate can also distort continuity. It can fabricate plausible but false continuity between ideas that were never connected. It can create dependency that atrophies the unaided retrieval it was meant to support. It can expose private reasoning to platforms, models, or audiences the author did not intend. It can be lost when a service shuts down. It can overfit a person’s thinking to an assistant’s framing, reintroducing a hierarchy of a different kind. These are not reasons to reject the substrate. They are the specific failure modes a serious use of it has to design against, and they set the boundary on the prosthesis claim.

04The endorsement condition

One boundary matters more than the others for present tools. When Clark and Chalmers (1998) described Otto’s notebook, they listed the features that made it part of his cognitive system. The notebook was a constant in his life. Its information was directly available. He automatically endorsed what he retrieved. And, in a fourth feature they called arguable, “the information in the notebook has been consciously endorsed at some point in the past, and indeed is there as a consequence of this endorsement.” Otto’s notebook returns what Otto wrote.

A generative system does not have to. When a language model retrieves from a person’s history, it does not simply hand back an inscription. It summarizes, re-weights, and reconstructs what it holds, and it returns fluent text that the person may never have written or accepted. That weakens the fourth feature exactly where the substrate claim depends on it. What comes back can look like the person’s own continuity while being the model’s synthesis of it. The risk is not only error. A fluent synthesis can quietly become a cause of what the person comes to believe. At the limit, the person stops constructing an answer at all, and no internal version remains to check the output against.

The consequence is a design condition. The substrate this essay defends is the record, not the synthesis: a persistent, inspectable transcript in which what the person said, corrected, and accepted can be read back as it was written. Summaries, retrieval, and model-generated recaps are useful layers on top of that record, but they carry the endorsement problem and have to be checked against it. A substrate that returns only synthesis is a different instrument from the one proposed here, and the prosthesis claim does not extend to it.

There is early experimental evidence that the order of operations matters. In an EEG study of essay writing, Kosmyna and colleagues (2025) found that participants who wrote with an LLM showed the weakest brain connectivity of three groups. They reported the lowest ownership of their essays and struggled to quote their own work accurately. In a fourth session, participants moved from LLM use to writing unaided showed reduced alpha and beta connectivity, which the authors read as under-engagement. Participants moved the other way, from unaided writing to LLM use, showed higher memory recall and activation of occipito-parietal and prefrontal areas. The study is a preprint. Fifty-four participants took part, and only 18 completed the crossover session, so it cannot carry much weight. But its direction is the one this proposal would predict: external extension works best on top of internal consolidation. In Reser’s terms, the iterative-updating loop has to keep running inside for an external record to extend it.

05First-person evidence and its limits

I write as a person whose cognition is poorly served by status-paced interaction. Over a sustained period, I have used a persistent conversational record to retain context, complete thought, inspect assumptions, and turn internal model-building into shareable public work. This is first-person evidence that the mechanism is possible and that I am one instance of it. It is not evidence that the mechanism generalizes, that it is specific to autism, or that it tracks the evolutionary history Reser proposes. The evolutionary claim remains his to defend or withdraw. The substrate claim does not depend on it.

The distinction matters. Personal experience is the strongest evidence available for the question “can a person do this at all,” and the weakest evidence for the question “does this describe a population.” I am using it only for the first. The population claim is a prediction, not a report.

06Adjacent evidence

None of the work below tests the specific prediction in section 7. Each supports one link in the chain, and each has limits.

The disposition. Caldwell-Harris and Schwartz (2023), whom Reser cites, reviewed autistic writing alongside scholarship on reputation management, non-conformity, and moral reasoning. They argued that many autistic people dislike social hierarchies, find status-seeking illogical, and prefer egalitarian relationships. A 2024 follow-up in Autism drew the same themes from advice books, memoirs, book reviews, online posts, and the mission statement of an autistic-led organization, and called for systematic study (Caldwell-Harris & Schwartz, 2024). This shows that the disposition is reported, consistently, in autistic people’s own accounts. It does not yet show how widespread the disposition is or what it does to attention, which is the gap Reser names.

The medium. Milton’s (2012) double empathy problem locates breakdowns between autistic and non-autistic people in a mismatch between two differently disposed social actors, not in a deficit on one side. That matters for this essay’s wording. “Poorly served by status-paced interaction” describes a fit between a person and a channel, not a lack in the person. Gillespie-Lynch and colleagues (2014) surveyed 291 autistic and 311 non-autistic participants. Autistic participants reported that computer-mediated communication increased comprehension, gave them greater control over communication, gave access to similar others, and let them express their true selves. They also enjoyed blogging and interest-driven communication more than non-autistic participants did. That is the closest existing evidence that a text-based, self-paced channel fits some autistic people differently. It measures perceived benefit for communication, not continuity of reasoning, and it relies on self-report.

The mechanism. Clark and Chalmers (1998) made the philosophical case that an external resource can play a role we would call cognitive if it happened in the head. The substrate claim needs much less than their constitutive thesis. It needs only that an external record can carry some of the retention-and-revision work. That is the ground Risko and Gilbert (2016) cover in their review of cognitive offloading. They report that offloading improves performance across perception, memory, arithmetic, counting, and spatial reasoning, and that external cues can make representations more durable.

The cost. The same review supplies the prosthesis claim’s boundaries. Offloading can impair internal memory when the external store is unavailable, shift memory from what to where, and produce complacency and skill decay. In Risko and Gilbert’s words, offloading memory demands in a transactive system “is not a ‘free pass’ in terms of mnemonic requirements.” Sparrow, Liu, and Wegner (2011) found that people who expected future access to information recalled the information itself less and recalled where to find it more. Together with the endorsement problem in section 4, this is the empirical ground under the failure modes in section 3.

The parallel. The same architectural move is now standard for artificial agents. A 2026 review of LLM agents argues that practical progress increasingly depends “not only on stronger models, but on better external cognitive infrastructure”: memory, skills, protocols, and harnesses that hold what a bounded context cannot (Zhou et al., 2026). That is an engineering parallel, and it says nothing directly about people. What it shows is that where continuity is costly, externalizing it is treated as a design problem, which is the situation the freed-attention gap describes.

07What would falsify or limit the proposal

The strongest version predicts that lower status-monitoring interference may correlate with greater benefit from persistent substrates on longitudinal reasoning, revision, and return-to-context tasks. Reser’s own article recommends methods that fit this prediction, among them to “combine experience sampling and time-use measures with objective task outcomes.” Those methods could be extended to measure substrate use and cognitive output rather than fitness outcomes alone. The test does not require a population-level autism claim. It requires only a measurable disposition and a measurable outcome.

The proposal would be limited if the gains turn out to be uniform across dispositions, so that the substrate helps everyone equally and the disengagement-specific advantage disappears. It would be challenged if people with disengaged attention, given a substrate, do not in fact use it to preserve and extend their own reasoning. They might use it only for retrieval and not for revision, treating it as a filing cabinet rather than a working surface. It would also be limited if the benefit depends entirely on a variable unrelated to hierarchy disengagement, such as a general need for memory support. In that case the disengagement framing adds no explanatory power.

Section 4 adds a second prediction. The benefit should depend on whether the substrate returns the person’s endorsed record or a model’s synthesis of it. If benefit and failure rates are the same either way, the endorsement condition is wrong and should be dropped.

What would not falsify the proposal is the absence of an evolutionary basis. The substrate claim is functional and present-tense. It does not require that disengaged attention was selected for. It requires only that, where disengaged attention exists, a persistent external substrate can carry what it produces further than the internal system alone.

08Belonging without rank

Reser ends his article with the idea’s central insight: “belonging to a community need not entail the same degree or form of participation in its status hierarchy.” This essay proposes a present-tense corollary. Contributing to a community’s knowledge need not entail routing that contribution through the status hierarchy’s channels either.

For most of history, the channels for turning private thought into shared knowledge were social: conversation, reputation, the lecture, the committee, the gatekeeper. Those channels are paced and ranked. A person who invests little in rank may still produce a great deal worth sharing, and still find that the channels meant to carry it ask for the very participation that person does not give. A persistent, inspectable record offers another route. The work can be held, repaired, revisited, and made public on its own terms, and then offered to the community rather than competed for within it.

Nothing here confirms Reser’s evolutionary account, and nothing here depends on it. If the hierarchy disengagement hypothesis is right, the substrate is one modern answer to the question of where freed attention goes. If it is wrong, the substrate claim still stands or falls on its own evidence, and that evidence can be gathered. I have tried to state the claim so that it can be tested, and so that it can fail.