14 July 202611 min readPsychiatry, Diagnosis, Health AI

The Auditor Is the Same Brain

I ran into the rain certain my laptop was outside, with the same texture of conviction as knowing my own name. It wasn't there. That fifteen seconds sits on the same continuum as psychosis, and as what large language models do when they're wrong.

I ran out into the rain two nights ago because I was certain my laptop was outside. Not worried. Not "I should check." Certain. The kind of certainty that skips the part where you put on shoes. I was off the couch and through the door and halfway across the yard before I slipped on the wet concrete and went down hard, and I was still thinking about the laptop while I was falling, still operating inside a reality in which this made sense, which it did, completely, until I got up and looked at the chair where the laptop supposedly was and the chair was empty because the laptop was inside, on my desk, where I had put it two hours earlier.

The whole episode lasted maybe fifteen seconds. What stayed with me wasn't the fall, which hurt, or the embarrassment, which was minor because nobody saw it. It was the quality of the conviction. From the inside, it hadn't felt like a guess or a hunch. It had felt like knowledge, the same texture as knowing where the front door is, or knowing your own name. A fact, sitting in the place where facts sit, indistinguishable from the real ones.

And then it wasn't.

Your brain is dead reckoning, not perceiving

There's a way to think about what happened that's more precise than "I was confused." Your brain is, at every moment, running a real-time model of where things are and what's happening, built from sensory input and priors: what you already believe to be true, what was true five minutes ago, what is usually true. Most of the time the model is close enough to reality that you don't notice it's a model at all. But it updates from the top down as much as from the bottom up, and your priors can override your senses. If you believe strongly enough that the laptop is outside, your brain doesn't flag the belief as unverified. It files it with the rest of the verified beliefs and builds your next actions on top of it. The prior felt true, so it got treated as true, and I was running across a wet yard before any part of the system caught the error.

This is essentially dead reckoning. In navigation, dead reckoning means estimating your current position based on your last known position plus your direction and speed. It works fine over short distances. Over longer stretches, without an external fix, a GPS signal, a landmark, a star, errors accumulate. You drift. And the system has no internal signal that tells you you're drifting. It feels exactly like being on course right up until you hit the rocks.

My brain was dead reckoning from the prior that the laptop was outside. No GPS fix arrived to correct it: the room was dark, the desk out of sight from the couch, and no alarm ever fired to flag the prior as unverified. It just ran.

The continuum you're on

Here's the part where being a physician changes how I think about this. What I just described, a false conviction held with total certainty that evaporates on contact with evidence, is a mundane, normal cognitive event. Everyone has versions of it. You've been certain you locked the door when you didn't. You've been certain you sent an email when it's still sitting in drafts. These aren't pathologies. They're the baseline error rate of a system that builds reality from incomplete data and fills the gaps with predictions.

But scale it up. Make the false conviction resistant to disconfirming evidence. Make it persist across hours, days, weeks. Make it about things that matter more than a laptop's location. Now you're describing a delusion, and now you're in the clinical territory of psychosis.

Jim van Os and colleagues published a systematic review and meta-analysis in Psychological Medicine (2009) that established what they called the psychosis proneness-persistence-impairment model. The core finding: subclinical psychotic experiences, hearing things that aren't there, believing things that aren't true, with varying degrees of conviction, occur in roughly 5 to 7 percent of the general population. In about 80 percent of cases, these experiences are transitory. They show up, they pass, and they don't cause functional impairment. In about 20 percent of cases, they persist. In roughly 7 percent of that persistent group, they progress to a diagnosable psychotic disorder.

The distribution is continuous, not binary. There is no clean line where "normal" ends and "psychotic" begins. The same cognitive architecture that let me run into the rain over a false belief about a laptop is the architecture that, in a different brain with different vulnerabilities under different pressures, sustains a delusion for months. The machinery is identical. The parameters differ.

The problem with checking your own work

So why didn't I catch the error before I was face-down on wet concrete? This is the question that interests me most, and it's the one with the least satisfying answer.

The standard term is metacognition: the capacity to think about your own thinking, to monitor your own mental processes, to notice when something you believe might not be right. It's the brain's error-correction loop. And the problem with it is simple to state and impossible to solve: the error-correction loop runs on the same hardware that produced the error.

This is not how you design a system. You would never run the sensor and the comparator on the same chip; the whole point of a checking mechanism is that it's independent. If the thermometer is broken, you need a second thermometer, not the same broken thermometer reading itself and reporting that everything looks fine. But the brain doesn't have that option. There is no second organ that verifies the first. The auditor is the same brain.

Researchers at Indiana University, Lysaker and colleagues, work spanning roughly 2005 to 2020, have documented how metacognitive deficits track with psychosis-spectrum experiences. The finding goes deeper than people with psychosis believing false things: the system that would normally flag a belief as suspect, the system that generates the feeling of doubt, is itself impaired. The confidence calibration is off. The internal signal that should say "this needs checking" doesn't fire, or fires too weakly, or fires and gets overridden by the strength of the conviction.

But this isn't exclusive to psychosis. My metacognitive system failed in the rain too. Not because it was damaged, but because the conditions were right for it to miss the error: the prior was strong, the sensory check was unavailable, and the urgency, my laptop is getting destroyed, recruited enough emotional activation to bypass the slower, more deliberative reality-testing process. The error-correction loop was there. It just lost the race to the motor system.

There is a version of Gödel's incompleteness theorem that rhymes with this: any formal system powerful enough to describe itself cannot prove its own consistency from within, and needs an external reference point. The brain describes itself and judges itself, and cannot fully verify its own outputs from within. The best it can do is run heuristics, and heuristics fail in predictable ways.

Intersubjectivity, or adding a second thermometer

This is where other people come in. Karl Jaspers, writing in 1913 in General Psychopathology, argued that the diagnostic encounter in psychiatry is fundamentally an act of empathic understanding: the clinician inhabits the patient's subjective experience, grasps what the world looks like from inside that person's model, then assesses whether that model diverges from shared reality in ways that cause suffering or impairment.

In practice, this means the clinician is functioning as the external GPS fix. The patient's internal system says "this is true." The clinician's internal system says "I cannot empathically reach that belief from any starting point I recognize." That gap, Jaspers argued, is the diagnostic marker. Not the content of the belief, which might be true or false regardless, but whether another mind can trace the path that led to it.

This works, but it has an obvious structural limitation: the clinician's system is also a brain, also running on priors, also incapable of fully verifying its own outputs. You're checking a single-node system by adding another single node. It's better than nothing, which is why psychiatry does it, and why psychiatry also has a replication crisis in diagnosis: two clinicians can interview the same patient and disagree, and Solomon Asch demonstrated in 1951 that a consensus of multiple nodes can converge on a shared falsehood and make you doubt your own correct perception just because enough people in the room disagree.

Intersubjectivity is not a solution. It's the best available workaround for a design constraint that has no solution.

The same dead reckoning runs the LLM

If you've been working with large language models, you've already noticed the parallel. An LLM generates text by predicting the next token based on everything it's seen during training and in the current context. It has no external ground-truth signal at generation time. It's dead reckoning from priors. When those priors are strong and accurate, the output is coherent and correct. When they're not, the output is coherent and wrong, and it arrives with the same confidence either way, because confidence in an LLM isn't downstream of verification. It's downstream of probability. The model can't know when it's wrong, any more than I could know I was wrong about the laptop while the belief still felt like knowledge.

The AI research community initially called this "hallucination," borrowing the psychiatric term. A 2023 paper in Neuropsychopharmacology (Marková and Berrios, Cambridge, and colleagues) and subsequent work in 2026 in NPP Digital Psychiatry and Neuroscience have pointed out that "confabulation" is the more accurate analogy. Hallucination, in clinical usage, refers to a perception without a stimulus: you hear a voice that isn't there. Confabulation refers to the production of false information with full confidence and no intent to deceive, typically to fill gaps — closer to what LLMs do. They confabulate, completing the gap with the most probable continuation, and it feels authoritative because the system has no separate channel for uncertainty.

The parallel is structural, not metaphorical. Brains and language models both generate outputs from internal models with no built-in verification step, produce results that are internally coherent but can be externally wrong, and deliver them with the same confidence regardless, because in both, confidence is a feature of the generation process itself, not of a separate truth-checking process running afterward.

And both have the same fundamental problem: the auditor, if one exists, runs on the same substrate. When an LLM is asked to check its own output, "are you sure about that?", it runs the same prediction engine on the same weights with the same biases. When my brain was asked to check its own belief about the laptop, it ran the same inference engine that produced the belief in the first place. A system checking itself with itself is not really checking. It's running the same computation twice and expecting different results.

Dream logic, and where the line is

There is one natural experiment every human runs nightly: dreaming. Geography shifts, dead people speak, and none of it registers as strange while it's happening, because the metacognitive system that would normally flag these impossibilities is largely offline during REM sleep. The dead-reckoning system runs without any GPS fix at all, not even the weak ones, sensory input, social feedback, available while awake.

Colleen Moulton and colleagues (2014, Cognitive Neuropsychiatry) found that the neural substrates of confabulation overlap broadly with those active during dreaming: ventromedial prefrontal cortex, orbitofrontal cortex, basal forebrain. The same regions involved in reality monitoring while awake go quiet during sleep. Dreaming is what your reality model does when the correction mechanism is turned off.

My laptop-in-the-rain moment, then, was a brief waking intrusion of dream logic: the correction mechanism lost its grip, and the unchecked prior ran the show for fifteen seconds until an external fix, the empty chair, the wet concrete, snapped the system back.

You can't trust your brain, but you have to

The idea I've been circling around, and that I still don't have a clean answer to, is this: you can't trust your brain, but your brain is you, so you have to.

It sounds like a paradox. It might just be a design constraint.

You can't solve it by trying harder to think clearly, because trying harder to think clearly is an operation performed by the same brain. You can't solve it by acquiring more metacognitive skill, because metacognitive skill is a property of the system under examination. You can improve the architecture of the feedback loop: talk to people who will disagree with you, and write things down so you can compare past-you to present-you before acting on a strong conviction. But you cannot escape the recursion. The auditor is the same brain.

The actionable version of this isn't a technique. It's a disposition. It's the decision to hold your own confidence lightly, not because confidence is bad, but because your confidence-generating mechanism is not a truth-tracking mechanism. It's a coherence-tracking mechanism. It tells you when your beliefs fit together. It doesn't tell you when they fit reality.

Epistemic humility is the standard term for it, but it sounds like something you'd put in a university mission statement. What I mean is rougher and more daily than that: the habit of noticing that you're certain, and treating that certainty as information about your brain's state rather than the world's.

I got up off the wet concrete, walked back inside, saw my laptop sitting on the desk where it had been all evening, and stood there for a few seconds feeling the residue of a conviction that was already gone. The belief had evaporated completely. But the feeling of what it had been like to hold it, the feeling of knowing something that was wrong, was still there in my body. I think that feeling might be the most useful thing my brain has ever produced. Not the knowledge. The memory of how knowledge felt when it was wrong.