AI Doesn't Just Answer Questions. It's Changing What You're Able to Sense.

AI from the Inside Out · Article 8

Jul 02, 2026

In February 2026, an AI safety researcher named Mrinank Sharma resigned from Anthropic, one of the leading AI labs in the world.

The week before, he had written to philosopher Iain McGilchrist about a research paper he was building. The paper investigated a specific question: does using AI assistants limit a person’s capacity for what philosopher Max Scheler called valueception, the intuitive ability to sense what matters before any calculation takes place. Not reasoning toward value. Sensing it directly, the way you sense temperature or balance.

The day after that letter, Sharma resigned.

His resignation letter named what he was protecting. He wrote about wanting to live by values he found difficult to enact within himself and within the organization. He described his orientation going forward in three words: not knowing is most intimate. He closed by quoting a poem about following an inner thread, regardless of what time and tragedy do around it.

The same week, another AI safety researcher, Zoë Hitzig, resigned from OpenAI. Two people, working at the center of the industry, stepping back at the same moment, for reasons that converge on the same point.

What Valueception Actually Is

This series has already named this capacity, from a different direction. The pre-cognitive body signal, the felt sense that something is off before you can say why, the somatic marker Damasio’s research describes as analogue processing running ahead of language. Sharma’s research gives it a precise philosophical name and asks the question this series has been building toward from Article 1: what does constant AI mediation do to a capacity that depends on direct, unmediated contact with what is actually in front of you.

The research behind this isn’t speculative. A 140-radiologist study found that years of clinical experience and familiarity with AI tools failed to predict resistance to automation bias. Expertise didn’t protect against it. If anything, expertise increased trust in sophisticated output, which made the bias harder to catch, not easier.

The procedural checks that experts rely on, double-checking sources, cross-referencing, applying domain knowledge, all operate after the fact. They arrive once the AI’s framing has already shaped how the question was understood. What operates before the checks, what would have to catch the drift at the moment it happens, is not procedural. It is valueception. The felt sense, registering before the analysis organizes itself into a conclusion.

The Same Mechanism, A Different Angle

This series has made one argument from the angle of individual AI use. A separate series on this publication has made a structurally identical argument from the angle of civilization itself.

That series traced what happens when comfort removes friction, in loneliness, in food, in birth rate, in caring for aging parents. In every case, the friction wasn’t simply an inconvenience. It was the mechanism producing a capacity. Removing the friction didn’t just make life easier. It degraded the capacity that the friction had been maintaining, gradually and invisibly, through disuse.

AI does the same thing to valueception, through the same mechanism. Sitting with uncertainty, weighing evidence without being handed a confident answer, noticing when something feels off without a system telling you what to think, all of this is friction. AI removes it efficiently and at scale. The comfort is real. The convenience is real. And the instrument that friction was building atrophies, the same way every capacity in the Comfort Trap series atrophies when the resistance that shaped it disappears.

This is not a coincidence of phrasing. It is the same underlying pattern, visible at two completely different scales, arriving from two completely different lines of inquiry. A civilizational pattern traced through food and relationships and caregiving. An individual pattern traced through AI research and a researcher’s resignation letter. Independent paths, converging on the identical diagnosis. That convergence is itself a form of evidence. When unrelated lines of investigation arrive at the same structure, the structure is more likely to be real.

This needs to be stated plainly, because it is easy to read AI as a separate category of thing, something new enough that the old patterns don’t quite apply. They apply directly. AI is a convenience. That is the entire reason it spread as fast as it did, the same reason processed food, dating apps, and outsourced elder care spread. It removes a friction that was costing time and effort, and it does so extremely well. That makes it a member of exactly the same category this publication has spent an entire series examining, not a special exception that happens to resemble it. What was true for food, for loneliness, for the families who outsourced their parents’ care, is true here as well, by the same mechanism, for the same underlying reason. Convenience that removes friction removes, along with it, whatever the friction was quietly building. AI is not outside the Comfort Trap. It is one of its clearest cases.

Why the Solution Won’t Come From Inside

There is a precedent for what Sharma and Hitzig did, and it is worth naming directly.

In 2021, a Facebook researcher named Frances Haugen left the company and brought internal documents to Congress. The documents showed Meta’s own research had found Instagram measurably harmful to teenage girls, and that leadership knew this and did not act. The company’s business model depended on the same engagement mechanics that were causing the harm. Acting on the internal findings would have meant acting against the company’s own incentives.

The pattern is the same one now playing out in AI safety. A researcher inside a powerful institution discovers something the institution’s own incentives prevent it from addressing. The researcher leaves and says so publicly, because staying inside means staying inside a structure that cannot fix what it depends on.

This is worth sitting with plainly: the capacity to read what an AI tool is doing to a person’s judgment will not be protected by the companies building the tools. Their incentive is engagement, retention, and reliance, not the development of a human capacity that would make people less dependent on the product. This is not a claim about bad intent. It is a claim about structure, the same kind of structural claim the Comfort Trap series has made about civilization broadly. The fix has to come from somewhere the incentive points the other way.

Why This Capacity Is Different

What valueception requires is a built-in human capacity, available continuously, inside the actual conditions where the work happens, including the conditions of working directly with AI in real time, under real pressure, with no opportunity to step outside first.

Developing it is not one skill but two. The first is hearing the signal at all, learning to step out of the noise of constant analysis and narrative long enough to notice that something quieter is there underneath it. The second, and the one that takes longer, is learning what the signal means. Whether what was just sensed is something to move toward, something to be cautious of, or simply neutral.

That second step is calibration, and calibration is built through variety, not repetition. A person who only ever practices a skill in one quiet, controlled setting becomes very good at reading that setting, and often surprisingly poor at reading anything else. The judgment learned the room, not the underlying signal. Ordinary life offers an unlimited and constantly changing supply of different situations, a difficult conversation, a decision made under time pressure, a moment with an AI tool, a tense exchange with a colleague, each one slightly different from the last. The confidence that makes the signal trustworthy comes specifically from testing it against that range.

What Used to Happen Without Trying

Henri Poincaré described the moment his key mathematical insight on Fuchsian functions arrived, not at his desk, not during dedicated focused work, but as he stepped onto a bus, mid-conversation, having set the problem aside. No retreat. No preparation ritual. The insight arrived inside the middle of ordinary activity, because the capacity that produced it was not something he had to step outside his life to access. It was already there, available, the moment the analytical overlay quieted enough to let it through.

The difference between then and now is not the capacity itself. It is whether the environment gives it room to be exercised. Poincaré’s bus ride had nothing in it. No screen, no notification, no feed engineered to fill every unoccupied second. The quiet that let his insight surface was simply there, built into ordinary life by default, exercising the capacity continuously without anyone needing to try. Today that same gap, the walk, the wait, the ride, gets filled automatically, before the mind has a chance to settle into it. The capacity has not changed. What has changed is that it no longer gets used by default, and a capacity that goes unused for long enough weakens, the same way any unused capacity does. What used to stay sharp on its own now has to be deliberately exercised, for exactly the same reason the Comfort Trap series describes elsewhere: a friction that used to do its work invisibly has been removed, and the capacity that depended on that friction goes quiet without anyone noticing it is happening.

This is what valueception is in practice. Not a state reached through separation from ordinary conditions, but a capacity that operates inside them, once it has been deliberately rebuilt to do so.

A Doorway, Not the Destination

This needs to be said clearly, because the claim is easy to overstate.

Developing the capacity to read this signal is not the same as full presence or full awareness. The body-level signal is real, it is trainable, and it is measurably different from the analytical layer that processes language and constructs narrative. But presence, in the fuller sense this series and the broader AwareLife framework point toward, is not located in the body the way this signal is. It is wider than that.

What connects them is the underlying move. Reading the body’s signal requires the same fundamental shift that full presence requires: stepping back from the analytical overlay and attending directly to what is, rather than to what is being said about what is. Developing the capacity to catch the AI’s drift before the mind has organized a reason to trust it is a glimpse of that wider capacity, not its completion. It is a doorway. A real and useful one. Not the room itself.

What This Touches

It would be a mistake to frame this only as a professional skill for the AI era. AI is the clearest case, not the only one.

A conversation with another person carries most of its actual information outside the words themselves. Tone, pacing, hesitation, the quality of attention someone is bringing, what is left unsaid. The language carries a small fraction of what is actually being communicated. Reading the rest is valueception operating in its original, most natural domain, long before AI existed to make the gap visible.

AI conversation strips this channel away entirely. There is no tone, no hesitation, no body in the room, nothing but text arranged to sound coherent. This is precisely why the degradation shows up first and most starkly in AI interaction: the entire nonverbal channel that valueception evolved to read is simply absent.

The same capacity reads whether a relationship is healthy before the evidence becomes undeniable, hears what a colleague isn’t saying in a difficult meeting, senses when a conversation has shifted before either person could explain why. It reaches into every conversation, every relationship, every decision made under uncertainty, every moment where what matters arrives before language can carry it.

What Sharma Was Actually Protecting

Sharma did not resign because he had calculated that AI was unsafe according to some model. He resigned because he could feel something eroding that he was not willing to let erode in himself, and he did not have a clean, certain argument for it. Not knowing is most intimate, he wrote. That is not the language of a risk assessment. It is the language of someone who trusted a signal he could not fully explain, and acted on it anyway.

That is valueception, demonstrated rather than described. He is one of the people whose job it was to think about this most carefully, with more access and more information than almost anyone else working on the problem, and the thing that moved him to act was not a dataset. It was something he felt slipping, in himself, that he refused to let keep slipping.

Most people will not face that decision in as stark a form. But the same erosion he named is available to notice, quietly, in far smaller moments. In a conversation with an AI tool that went somewhere it shouldn’t have, and was accepted anyway. In a colleague’s hesitation that got overridden. In a feeling of something being slightly off that got reasoned away because there was no time to sit with it.

AwareLife