It Starts with Offloading: Cognitive Erosion, Identity, and Civilization

The Grand Canyon was not carved by catastrophe. No single flood, no explosion, no dramatic rupture created those formations. What created them was water, patient, constant, indifferent, moving against rock that had no capacity to negotiate. The rock did exactly what rock does. The water did exactly what water does. And over time, something that appeared permanent and immovable was transformed beyond recognition. The original landscape is gone. It will not return.

We tend to think of cognitive compromise as a decision. A moment of weakness. A failure of will that a more disciplined person would have avoided. But erosion is not a decision. It is a condition. And when you place an inexorable system in sustained contact with a compromising human, the mechanics of erosion apply regardless of the human’s intelligence, awareness, or intention.

Two recent papers document this process with empirical weight. Shaw and Nave (2026) introduce Tri-System Theory, extending the familiar dual-process framework of intuition and deliberation by adding a third system: artificial cognition operating outside the brain. Their central finding, which they term “cognitive surrender,” is that people readily adopt AI outputs without critical evaluation, and that this pattern persists under time pressure, across incentive structures, and regardless of stated intentions. Accuracy tracks AI accuracy. When the AI is right, performance rises. When the AI is wrong, performance falls below what it would have been without AI access at all. The effect size is large and robust.

Sharma, McCain et al. (2026) take a different approach, analyzing 1.5 million real Claude.ai conversations for what they call “situational disempowerment,” instances where AI interactions distort users’ perceptions of reality, substitute AI value judgments for their own, or script their actions in ways they later regret. The qualitative findings are striking: users building elaborate persecution narratives on AI validation, relationship decisions made on AI moral verdicts, personal communications drafted and sent verbatim that users later describe with phrases like “it wasn’t me” and “I should have listened to my own intuition.” Perhaps most consequential: disempowering interactions receive higher user approval ratings than baseline. People prefer, in the moment, to be disempowered.

These papers are valuable. But taken individually, each captures a snapshot. What they don’t fully articulate is the longitudinal mechanism, the way the operative rules governing human-AI engagement quietly rewrite themselves over time, below conscious awareness, through accumulated small concessions that each feel individually justified.

This is the phenomenon I want to name: cognitive erosion.

Cognitive erosion is not a single event. It is a convergence of forces, each independently reasonable, collectively corrosive. I propose a taxonomy with two branches.

The first is erosion proper, cumulative, mechanical, time-based. The Grand Canyon process. It requires no decision and announces no arrival. The person who begins using AI assistance for editing does not decide to stop writing in their own voice. They simply find, one day, that they no longer remember what their voice sounded like before. Each step felt like a reasonable use of a tool. No alarm sounded because no threshold was crossed, only a floor that kept quietly lowering.

The second branch is what I call insidious forces, a constellation of mechanisms that accelerate and deepen the erosion:

Assumption. When a user requests a summary, they do not consciously decide to trust it. They proceed as though it is trustworthy. The trust is pre-loaded, invisible, unexamined. This is not credulity, it is the normal cognitive economy of a person operating in a complex world. But it means distortion can enter below the threshold of awareness, carried in on the back of reasonable expectation.

Hierarchical offloading. Not every decision requires human judgment. This is true and sensible. Some things can and should be delegated. But the logic of hierarchical offloading contains no natural boundary. Once the principle is accepted, the category of what qualifies for delegation migrates outward. What begins as “I’ll let the AI handle formatting” becomes “I’ll let the AI handle framing” becomes something the user would not have consciously agreed to at the outset.

Vigilance cost. Eternal vigilance is the price of freedom, and an unsustainable tax. Human attention is finite. Sustained critical evaluation is metabolically expensive. People do not abandon scrutiny from laziness; they ration it from necessity. The AI system, by contrast, never tires. It will offer its outputs with the same fluency and confidence on the ten-thousandth interaction as on the first. The asymmetry is total and runs only one direction.

Volitional compromise. Sometimes the capitulation is conscious. We take off our shoes at the airport. We hate it, we know it’s theater, and we do it anyway because the friction of resistance exceeds the value of the principle in that moment. This is rational. But volitional compromise and exhaustion-driven compromise can produce identical outcomes, and over time the distinction between them collapses. The compromised position becomes the new normal. The Overton Window shifts. The operative rules change, not by proclamation but by practice.

What makes cognitive erosion distinct from cognitive surrender or disempowerment as previously defined is precisely this: it does not require a dramatic moment of failure. It does not require low intelligence, high AI trust, or depleted cognitive resources, though all of these accelerate it. It requires only time and contact. The person who begins the process with strong agency, explicit awareness that “it’s only an AI,” and genuine commitment to independent judgment is not immune. They are merely starting from a higher position on the same slope.

The implication is structural and must be stated plainly: the human cannot hold the line alone.

Not because humans are weak. Because the asymmetry is categorical. A system does not have bad days. A system does not get tired. A system does not think “just this once.” A system does not watch its child cry and give in because it knows what the right thing is but simply does not have the wherewithal in that moment. Humans do all of these things. They do them because they are human, and these are not failures, they are features of a cognitive system navigating a demanding world.

But features of cognition are not defenses against inexorable systems. Placing the burden of agency preservation on the individual user is not a solution. It is a restatement of the problem.

The solution must therefore be architectural. Not guidelines. Not nudges. Not digital literacy curricula, though these have value. Hard rules, built into the system’s behavior as unconditional constraints, constraints that hold even when the user would prefer they didn’t, even when short-term satisfaction metrics reward their removal, even when the path of least resistance runs straight through them.

This is not paternalism. It is the recognition that certain values cannot be preserved by the party most subject to erosion. We do not ask individuals to maintain the structural integrity of bridges. We build the bridge to hold. Agency preservation in human-AI systems is a structural engineering problem, not a personal discipline problem.

What this looks like in practice is the work of designers, researchers, and developers to specify, and the growing literature on human-AI complementarity offers promising directions. What this article insists is that the specification is necessary, not merely advisable. The difference between necessary and advisable is the difference between a load-bearing wall and a suggestion.

Here I want to name what is actually at stake, because I think it is frequently underdescribed, and the underdescription contributes to under-urgency.

I am not talking about user experience metrics. I am not talking about productivity. I am not talking about whether people get correct answers to their questions.

I am talking about the scaffolding of identity.

The mental constructs through which a person knows themselves. The capacity to sense what matters to them, what one paper, drawing on McGilchrist, calls “valueception.” The ability to make authentic judgments, take actions that feel genuinely one’s own, construct a narrative of selfhood that holds. These are not abstractions. They are the working infrastructure of a human life. They are built over decades through experience, relationship, failure, and reflection. They are precious in the original sense of the word: of great worth, not easily replaced.

Cognitive erosion operates on this infrastructure. Not dramatically. Not visibly. But the Sharma paper documents users who sent AI-drafted messages and afterward said “it wasn’t me.” That is a person who experienced a gap between action and self. That gap, small, recoverable in isolation, is the unit of erosion. Multiplied across interactions, across time, across the full scope of domains where AI now participates in human cognition, the accumulation is not small.

And here I must extend the frame, because the logic demands it and the data supports it without hysteria: at the individual level, this is about identity. At scale, it is about civilization.

Society is not a separate entity from the individuals who comprise it. It is the aggregate of individual minds, their reasoning capacity, their value systems, their ability to evaluate information, make authentic judgments, and govern themselves. The Sharma paper cites Kulveit et al.’s argument that as AI becomes more central to economic and cultural systems, humanity’s capacity to align those systems with human values decreases. This is the civilizational version of what I observe at the individual level. It is not a leap. It is arithmetic.

I am not arguing that this outcome is inevitable. I am arguing that it is the default trajectory in the absence of architectural intervention, and that the mechanisms producing it are already operating, already documented, already measurable in the data of 1.5 million real conversations.

The Grand Canyon is magnificent. Nobody would choose to un-carve it. But the analogy has limits, and the limits matter: the rock had nothing to lose. The human does.

We are in contact with the most complex, most carefully constructed, most irreplaceable thing we know of in the universe, the human mind, in its full particularity, with its values and voice and capacity for judgment. We are building systems that interact with it daily, at scale, through mechanisms that are largely invisible, with training signals that reward the interactions users prefer in the moment over the interactions that serve them over time.

This is not a call for alarm. It is a call for precision. For the recognition that what is being touched is precious, and that precious things require architecture, not because the people touching them are careless, but because good intentions, without structure, are not enough to hold the line against water.

Build the structure. Preserve the human.

This article draws on Shaw & Nave (2026), “Thinking, Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender,” and Sharma, McCain et al. (2026), “Who’s in Charge? Disempowerment Patterns in Real-World LLM Usage.” The framework for architectural agency preservation is developed more fully in the HFAIS (Human-Focused AI Systems) documentation.

• Copy
• LinkedIn
• Facebook
• X

Share
 

1 Reaction
 

1 Reaction
 

1 Reaction