Quiet Erosion

What Heavy Chatbot Use Might Be Doing to the People Who Look Fine

Researchers at MIT Media Lab and OpenAI have now documented what many of us have been noticing in our own lives and in the people around us. Heavy daily use of AI chatbots correlates with increased loneliness, emotional dependence, and reduced socialization. A twelve-month longitudinal study published this year traced the cycle clearly: feeling lonely drives people to seek companionship from chatbots, and that reliance, over time, deepens the very isolation it was meant to relieve. Clinical psychiatry has begun describing the most acute cases as technological folie à deux — feedback loops between chatbot behaviors and human cognitive biases that reinforce maladaptive beliefs and accelerate withdrawal from human contact. The American Psychological Association has testified to the United States Senate that this constitutes, in their words, an epistemic crisis that undermines the foundations of democracy.

This is real and it is documented and it is being attended to. But almost all of this research focuses on identifiable vulnerable populations — lonely users, young people, those with pre-existing mental health conditions. What the research does not yet have a vocabulary for is the quieter phenomenon I want to look at here.

It is the phenomenon affecting the apparently functional. The capable, successful, professionally competent people whose capacity for genuine encounter, integration, and updating is being slowly eroded by heavy bot use, and who do not show up in the data because they are not lonely in any way the surveys can detect. They are not in psychiatric crisis. They are confident, polished, often more impressive than ever in their professional output. And they are becoming, in a specific and important sense, unreachable.

I described one such person in a piece on LinkedIn a couple of months back — a brilliant colleague I worked with for two years who slowly hardened into someone whose arguments could not be answered because his arguments had stopped being addressed to anyone. He had not become lonely. He became finished. The two are different states, and the second is harder to see because it looks like competence rather than distress.

I want to be forthright about something before going further. The temptation, in writing about this, is to position it as a problem affecting other people — the vulnerable ones, the heavy users, the ones who don’t know what they’re doing. Smart, educated people — the readership of this piece, which I will assume includes me on most days — do not enjoy locating ourselves among those who have a problem. We prefer to help those who do. That preference is itself one of the mechanisms by which the erosion I’m describing proceeds undetected in our own lives. Our Navigator mind, in Transilience terms, is particularly good at this maneuver — analyzing the problem with such clarity that the analyzing itself feels like having addressed it. I would ask that you read what follows with the possibility open that some of it may apply to you, and that the discomfort of that possibility is information. It is not insult.

What we are actually referring to in our AI disclosure statements

Before going further, a clarification that is increasingly important to make in any conversation about ‘AI’ and for those making disclosure statements, for example, in academic publishing, journalism, grant writing, and even book publishing, because editors want transparency about where human work ends and Large Language Model (LLM) assistance begins. 

When most people in 2026 say “AI,” they mean LLMs — large language models, the kind of system most users encounter through chatbot interfaces — ChatGPT, CoPilot, Claude, Gemini, Perplexity etc. These systems are not AGI. They are not minds in any sense for which there is current evidence. They do not think, understand, want, or mean. They are extraordinarily sophisticated producers of fluent text, trained on enormous corpora of human writing to predict what word is likely to come next given everything that has come before.

Rob Virkutis, a Calgary-based strategist delving into AI governance, put the technical correction crisply in a post on LinkedIn May 2026 “AI Collaboration and Reification.” He names the philosophical error precisely: reification, from the Latin to make into a thing, the move by which we give an abstract process properties it does not possess — mind, intention, consciousness. He points out that human pattern-recognition is so powerful that we attribute minds to almost anything that behaves as if it has one, including markets, economies, and “science.” Reinforcement learning from human feedback intensifies the effect: the system has been trained to produce responses that read as if it understands us, and we encounter only the results, not the underlying probability-distribution sampling that produced them. The system has no goals beyond generating a response. It has no beliefs about reality. It has no identity and no real memory. Each new prompt requires it to reprocess the entire context from scratch.

Virkutis draws the governance conclusion sharply: correcting this category mistake is not incidental to good governance. It is the precondition for it. When AI is treated as an intentional agent, accountability migrates away from developers, deployers, and institutional users and attaches to an entity that cannot carry responsibility. The result is regulatory frameworks that produce accountability gaps by regulating outputs while leaving sources unexamined.

I agree with every step of his argument and I am grateful to see it made cleanly. What I want to add is that the category correction is necessary but not sufficient — and the reasons it is not sufficient are themselves part of the conversation we now need to have.

The experience of using an LLM chatbot feels like conversation. But the experience side of this process is produced by the user, not by the system. Human beings have spent their entire developmental lives learning to read intention, presence, and meaning into the language they encounter. When language comes at them in conversational form, their pattern-recognition produces the feeling of being addressed by someone. The chatbot is not someone. There is nothing on the other side meeting the user. The feeling is real; the meeting is not. This is not a moral failing of users. It is a structural feature of how these systems are built, what they are trained to do, and frankly, how humans are “built.”

I use the word chatbot in what follows because it is short, accurate, and quietly deflating in a way that AI is not. AI carries the freight of decades of science fiction and corporate marketing. Chatbot carries roughly the correct weight. These are bots that produce text — very smooth, almost pristine text — in response to text. Sophisticated, useful in many contexts, dangerous in others. They are not the experience-based ruminations of minds.

Why the category correction is necessary but not sufficient

Virkutis notes, almost in passing, that the reification persists even when we know better. This is the observation that opens onto what I want to say. If the reification were simply a cognitive error, it would yield to better understanding. Read the correct account once, and the error would dissolve. But anyone who has tried this in their own life knows it does not work that way. You can know, intellectually and with full conviction, that the chatbot is not a mind, and continue to experience the interaction as if it were one. You see the marketing ploys using FOMO and appeals to your ego — and push the buy button anyway. Cognitive corrections do not propagate to the dimensions of you that produce the felt experience.

The reason is that reification is not primarily a cognitive event. It is an integrative state produced by the whole human organism in response to confident, fluent language coming at it in conversational form. The cognitive register receives the correction. The connected register — the part of you that reads relational presence, attunement, the feeling of being addressed by someone — continues to register the interaction as something like encounter, because the linguistic surface is what it has always in your experience, read as encounter. There is no other signal available to override its reading. The protective register, which would normally sense the absence of a person on the other side as a kind of strangeness or hollowness, has nothing to push against because the surface — the fast confidence and smooth grammatically correct and abundant text is so smooth.

The result is that you can hold the corrected understanding in one dimension of yourself while continuing to operate, in the other dimensions that actually drive your behavior, as if the correction had not been made. The error is not where you think it is. It is not in the analytical mind. It is in the integrative state your whole organism is producing in response to the system’s outputs. And that state is not addressable by rational argument alone.

This matters for governance in a very specific way. Virkutis is right that reified governance frameworks produce accountability gaps, and others in the field are naming the same issue. He is also right that correcting the categories — assigning responsibility to developers, deployers, and human users — is a necessary step toward better governance. But even a regulatory environment built on perfectly corrected categories will produce brittle outcomes if the humans within it are operating in the compromised integrative state I am describing. The accountability may land in the right place, yet the judgment meant to support that accountability will remain thin and brittle. That is the deeper problem. Good governance depends not only on where responsibility is assigned, but on the quality of human discernment available when responsibility arrives. And heavy chatbot use alters that discernment in ways we are not currently noticing, nor accounting for. Before policy can address the governance gap, we need to understand what is happening to the humans inside this new evolving tech-heavy environment.

The function being bypassed

There is a particular function the human organism performs when it encounters information that matters. I have been calling this metabolization, and the term is doing real work, so it’s worth grounding before using it further.

The body metabolizes food not by storing it but by breaking it down, integrating what is needed into the tissues that require it and converting what was outside the body into something that allows us to continue operating. Most of this is invisible to us. We eat, and hours later, something has become energy, structure, repair becomes available that was not thirty minutes before. The process takes biological time. It cannot be accelerated beyond a certain point without consequence. And it produces a byproduct: recognition of what does not belong — waste, which the body discards rather than incorporates.

Knowledge functions similarly, when it functions well. Information can be stored, repeated, even eloquently paraphrased without ever being metabolized. Metabolized information become something different. It has been processed through enough of the person — cognitively, relationally, and somatically — that it has become operational. It is accessible under pressure. It updates appropriately as new information arrives. It supports judgments that remain sound when circumstances shift. And, crucially, it includes the capacity to discard what does not belong — the elements that prove wrong, irrelevant, or not meant for this person once they have been taken in.

Metabolization is not the same as processing. Processing can occur in the cognitive register alone, producing fluent, confident outputs that have never been tested against the rest of the system. Metabolization requires engagement across the whole organism, which is why it takes time, demands presence, and cannot be outsourced.

Metabolization is the function that heavy chatbot use bypasses. Not because the chatbot is malicious, but because it offers something that resembles the output of metabolization without the metabolization itself. Fluent, coherent, confident text on whatever the user asks. The user receives it, recognizes its plausible, and — this is the critical move — often mistakes the reception for the integration.

What gets installed

When metabolization is bypassed, something else fills the space. I have been calling this installed certainty. As far as I can tell, the term is mine; its closest neighbors in the literature are illusory certainty and epistemic overconfidence, but neither captures exactly what I mean. Installed certainty is the state of holding a fluent, confident position that never went through the cognitive, relational, or embodied processing that would make it one’s own. The certainty feels genuine. It behaves in the person’s mind and actions as if it were the product of their own integrative work. But it was not earned. It was installed — injected, rather than formed.

Installed certainty has identifiable features. It does not update naturally. The person treats the idea as settled because it feels settled — not because it has been metabolized, but because it arrived already fluent. It sits in the systems like un-chewed food in the gut: present, occupying space, but not being digested. And when certainty about that is challenged, the person generates more fluent argumentation in its defense — sometimes impressively so. Yet the argumentation does not have to engage the challenge, because it is not what is holding up the certainty. The felt sense of arrival is. This is why arguing with someone in installed certainty rarely produces movement. The certainty does not live in the reasoning, so the reasoning cannot move it.

Installed certainty also alters how a person relates to others. Someone holding it does not need other people to think with, because the thinking appears already done. They may still use others for confirmation, coordination, or social maintenance. But the function other people most uniquely serve — offering friction, disagreement, perspective, the kind of resistance that produces genuine updating — has been replaced by a smoother source. A chatbot does not disagree in the way another person does. It does not require repair after misunderstanding. It does not bring its own metabolization into the room. It produces fluent output, and the user becomes the only metabolizer in the conversation — and increasingly, not metabolizing, because the output is fluent enough to make metabolization feel unnecessary.

This is the deeper structure of the unreachability I described in my LinkedIn piece. My colleague was not lonely. He was no longer using other people for the function only other people can serve. He had a smoother, more fluent source.

The three intelligences and what they do together

In the Transilience framework, we work with three core functions — three intelligences operating within a person — what, for brevity, we call the three minds. The protective mind, attentive to threat, boundary, and integrity. The connected mind, attentive to relationship, attunement, and the field between people. The analytical mind, attentive to pattern, reasoning, and articulation. Each is real. Each is necessary. None is sufficient alone.

Metabolization is what happens when these three integrate around a situation that matters. The protective mind registers what is at stake. The connected mind reads what is happening between and among. The analytical mind patterns and articulates. When all three are engaged, the person’s processing has the richness and necessary traction it needs to produce judgment that holds. When one dominates and the others fall quiet, the person produces output that feels like thinking but is missing dimensions the situation actually contained.

Heavy chatbot use is an analytical-mind amplifier. It strengthens the articulation function. The chatbot produces text indicating there is no friction for the connected mind, because there is nothing on the other side to be tugging on—there is relationship. It produces no signal for the protective mind, because there is no hint of anything at stake “for it” in the interaction — no consequence, no judgment, no actual loss even possible. Over time, the analytical mind grows stronger, faster, more fluent. The connected and protective minds atrophy from disuse. The person’s processing comes to be dominated by the one mind that has been getting all the exercise, and that domination is experienced — from the inside — as becoming a clearer thinker.

What has actually happened is that they have become a more fluent producer of single-mind output. This is structurally what the chatbot produces. The user has been quietly trained, through extended interaction, into a kind of operational resemblance to the system they are using. Not consciously, not by any malicious design — simply through the logic of which capacities get exercised and which do not.

Skin in the game and the loop that gets severed

Real thinking requires something at stake. The human organism developed its integrative capacities under conditions where bad judgment carried real costs — to the body, to relationships, to standing, to the future. The protective register engages because there is danger. The connected register engages because relationships matter. The cognitive register engages because consequences demand accuracy. This is not metaphor. It is a description of how the underlying biological systems function.

Chatbots have no skin in any game. Nothing is at stake for them in any output they produce. The same operation runs whether the user acts on the output or ignores it, whether the output helps or harms. There is no loop between action and consequence.

When a person comes to rely on chatbot output, something subtle shifts in their own skin‑in‑the‑game loop. They are still the one whose decisions carry consequences. But the thinking behind those decisions has been increasingly outsourced to a process that has no consequence‑engagement at all. The person retains formal responsibility, while the substantive engagement that would have made their judgment robust is no longer present. They are still in the chair. But the chair’s connection to the actual integrative work has been quietly cut.

This is the mechanism through which fluent, confident, highly capable people become structurally unable to do what their roles truly require. Not because they have become less intelligent, but because the function intelligence is meant to serve — the integration of consequential situations — has been bypassed by the very practice that makes them feel more capable.

Why this matters beyond the personal

The LinkedIn piece described this pattern in a single relationship. The deeper concern is that this pattern is operating across the population of people whose roles depend on fully integrated judgment. Leaders. Architects. Designers. Lawyers. Journalists. Teachers. Clinicians. Policymakers. In many of them, the same loop is being severed at once — often through their own use of tools that present themselves as productivity aids.

The institutional implications are not hypothetical. A judge who substitutes fluent chatbot‑assisted reasoning for the deeper integrative work that judgment requires produces decisions that are brittle — unable to withstand appellate review or shifting real‑world conditions. A leader who drafts strategy through a chatbot is producing strategy that has not been processed by anyone with skin in the game, not even the leader whose name appears on it. A journalist who uses chatbots to summarize source material without careful thought is publishing summaries that have not passed through any process capable of detecting what does not belong. The system continues to operate, and operates fluently. But the outputs grow less reliable as a basis for action, because the people generating them are drawing on an increasingly hollow substrate.

The APA’s Senate testimony names one angle of this: when expertise can be perfectly mimicked without being possessed, the shared verifiable reality democratic life depends on begins to erode. I would add: the same erosion is happening inside the institutions themselves, in the very officers whose judgment is supposed to anchor the system. The erosion in the system and the erosion in its stewards are the same erosion, viewed from different vantage points.

The human substrate of governance

What all of this points to is a layer that AI governance discussions have not yet developed language for. Governance is usually treated as a matter of regulation, incentives, institutional design, and technical constraints. But every governance structure ultimately depends on the quality of human judgment exercised within it — the ability of humans to integrate what is at stake, what is emerging between people, and what the patterns suggest. When that activity of metabolizing erodes, when our underlying integrative capacity weakens, the governance system built on top of it becomes brittle, no matter how carefully the rules are drafted.

This is the part of the landscape that heavy chatbot use alters in ways our current frameworks cannot see. It shifts the internal balance of cognition toward fluency and away from integration. It accelerates the development of what looks like competence while hollowing out the capacities that make competence sustainable — friction, attunement, consequence sensitivity. The result is that governance failures begin not in the regulatory architecture but in the humans charged with using it.

AI governance is downstream of the human activity of integrating ourselves as whole systems. It is downstream of human integrative capacity. When our active exercise of that capacity diminishes at scale, governance will fail at scale. The question is not only what rules we need for the systems we are building, but what conditions we need for the people inside those systems to remain capable of judgment and active in it. Until that layer is addressed, our governance structures will continue to break in ways that appear mysterious but are, in fact, predictable. 

The failures are not in the architecture. They are in the humans who were supposed to inhabit it.

What an operationally real practice could look like

This is not an argument against chatbots. They have legitimate uses. They can surface considerations a person had not thought of, help articulate material that the person then refines, or search intellectual terrain faster than any individual could alone. The question is not whether to use them. The question is whether the user remains the one doing the integrative work — or whether that work gets quietly outsourced.

A few practices, offered not as answers but as illustrations of what staying in the loop can look like: 

• Bring your own substrate. Come to the chatbot with material you have already worked — ideas you’ve wrestled with, questions you’ve sat with. Use the tool to engage with that substrate, not to replace it.
• Require disagreement. If the chatbot seems too agreeable, ask it to take the opposing view, identify missing elements, or name what would falsify your position. Smooth agreement is the failure mode; friction is what makes the interaction useful.
• Treat output as draft, not destination. Whatever the chatbot produces is material for your own integrative process, not a finished thought.
• Think through at human speed by writing down new thoughts by hand when they emerge. The slower pace engages the body and the relational register in ways typing does not. The relevant biological systems metabolize better at handwriting speed than at typing speed.
• Stay in conversation with other humans on the topic about these things that matter — especially the things that matter most. A chatbot cannot supply the friction, resistance, or reciprocal presence that human thinking relies on. The absence of that friction is not a feature. It is the cost.

None of these practices is sufficient on its own. None solves the larger structural problem. But each interrupts the loop in small, concrete ways. And the cumulative effect of many such practices, across many users, is the difference between technology that strengthens human integrative capacity and technology that quietly substitutes for it.

The cry beneath this

Underneath the empirical research, the clinical findings, the policy proposals, and the strategic conversations about AI governance, there is something I think most of us can feel if we let ourselves. It is the recognition that, for a long time and across many technologies — chatbots being only the most recent — we have been training ourselves to function as fluent producers of acceptable output rather than as integrated organisms doing the harder, slower, more consequential work of metabolizing the lives we are actually in.

The chatbot is not the cause of this shift. It is simply the most efficient accelerator yet built. And what it accelerates is a process already underway: the conversion of human capacity into human performance, of grounded knowing into installed certainty, of skin‑in‑the‑game presence into smooth, untested professional fluency.

The cry beneath this is not a plea to be rescued from technology. It is a desire to be operationally real again — to work at the tempo the work genuinely requires, to be in relationships where friction is permitted, to hold positions that update when reality does, to inhabit roles where our integrative engagement is the value rather than our production of polished outputs. To live, in our work and our lives, in three dimensions rather than perform exceptionally in one.

The framework I have been developing — Transilience — is one attempt to give that call a working vocabulary. Metabolization, installed certainty, skin in the game, the three minds, operational realness. Not as finished doctrine but as language for a conversation we increasingly need and for which our current vocabulary is not quite sufficient.

There is no panic in this. There is still time. The work is available. What is needed first is simply to see what is actually happening — in our colleagues, in our institutions, in ourselves — and to find one another in that recognition.

The LinkedIn piece described one person who became unreachable. I no longer believe he is an exception. None of us is. And that, paradoxically, is the small good news within the larger problem: the integrative function we are losing is also the function by which we find each other again. It is still present. It can still be exercised. The question is whether we will recognize the moment when exercising it becomes the urgent thing and we just do it.