AI handles patterns with ease.

Humans feel the weight of being wrong.

That’s where the divide becomes clear and important.

A recent study from the Icahn School of Medicine at Mount Sinai evaluated a health-focused version of ChatGPT using 60 clinician-authored patient scenarios. The findings were important and cast a light on the utility of large language models in medicine. Here's the key takeaway: In 52% of cases that physicians unanimously judged to require emergency care, ChatGPT did not recommend it. It performed well in routine complaints and in textbook emergencies where the pattern was clear. But it stumbled in the gray zone, where clinical signs were subtle and the cost of being wrong carried real consequence.

Consider a patient with abdominal pain or fever of unknown origin. The presentation may not scream catastrophe. The vital signs are not yet extreme and the laboratory data are incomplete. The experienced clinician senses trajectory, not just snapshot. That sense often leads to escalation before certainty arrives.

The inverted U curve revealed by the study is more than a product critique. It's my sense that this is a cognitive signature of AI and certainly worth a closer look.

The Architecture Reveals Itself

Lets start there. The system did not fail because it was reckless; it behaved according to its design. LLMs are engines of computation. They aggregate patterns across vast data sets and generate responses that are statistically coherent within training data. The coldness of that sentence reflects the computational rigidity that is, in my opinion, functionally antithetical to human cognition. Simply put, when structure is clear, they excel. When ambiguity intersects with risk, we see a dangerous emergence of computation fragility.

Now, a clinician in that gray zone does not simply calculate likelihood but "leans toward" consequence. If the faint possibility of bad outcome exists, the prudent move is often escalation—an escalation that may not maximize a statistical perspective. But, and more importantly, it reflects a weighting of what happens if the judgment is wrong. Both physician and patient bear the outcome, and it's this shared exposure that shapes the decision.

Computation optimizes within parameters and a model's risk assessment can be tuned toward caution. Yet it doesn't stand inside the exam room or within the impact of consequence. It produces what best fits the distribution and doesn't experience the gravity of error. OpenAI describes its health tool as supportive rather than diagnostic, and that distinction matters. The study, however, underscores what can happen when coherence is mistaken for clinical authority.

Computation and Judgment

This brings us to a larger distinction. Computation follows rules, even when those rules are probabilistic. It works within frames defined by training data and objective functions. It can be exhaustive within its bounds. But here's the difference. Judgment operates differently, as it interprets situations that don't fit neatly inside predefined categories. Human judgement asks what matters here, not simply what is most likely, and it revisits the frame itself when necessary.

AI is certainly a powerful technology in medicine and will play an expanding role. However, it's essential that we recognize that we're dealing with two architectures of intelligence that are not identical. And these two modalities may create a type of parallax intelligence that can work synergistically. Today's AI represents a dramatic expansion of computational capacity. Human cognition, at its most distinctive, is judgmental in the classical sense and lives in both consequence and biography.

Fluency Is Not Prudence

When fluency feels like prudence, we risk assigning computational systems roles that depend on judgment. In medicine, that confusion can become visible quickly. In everyday life, it might be a bit more quiet. We consult algorithms for recommendations and rankings. And in this context, we begin to treat decisions as efficiency problems.

My contention has never been that AI is flawed or even malicious. It is different. What I have described as anti-intelligence is not the absence of intelligence but a structural inversion of it. It produces coherence without any interior life. It generates answers without inhabiting the world in which those answers live and die. That distinction matters most where ambiguity and consequence intersect.

The recent clinical findings offer an important reminder. The weakness didn't appear in obvious emergencies or trivial complaints. It emerged where interpretation and a sort of moral and clinical caution converged. That's not a glitch in AI, but a reflection of its architecture.

The Fork in Our Own Cognition

Computational intelligence will continue to expand and that expansion is valuable and transformative. It augments clinical care, research, and even patients communication. Alongside it, we need to remain clear about what judgment is and where it remains indispensable. Judgment isn't merely a threshold adjustment in training an LLM. It is orientation in a world where outcomes are lived.

For me, the deeper question pushes beyond whether machines will become more capable. That's an easy yes. The question is whether we—clinicians, teachers, parents—remain attentive to the difference between optimization and responsibility. If that distinction blurs, the technology will not be the only thing that changes. Our own cognition may begin to tilt toward what is computationally convenient.

The Mount Sinai study was a clinical evaluation, yet also reads as a philosophical one. It signaled to me that intelligence has more than one form and not just some amalgamation grabbed from the pages of a science fiction novel.

Ramaswamy, A., Tyagi, A., Hugo, H. et al. ChatGPT Health performance in a structured test of triage recommendations. Nat Med (2026). https://doi.org/10.1038/s41591-026-04297-7

© Psychology Today