What AI Can’t Replace in Behavioral Health

That conclusion is now being reached by computer scientists themselves.

In his new book, Turing’s Mistake: Escaping the Yoke of Unintelligent Machines, renowned computer scientist Peter J. Denning argues that the core assumptions driving AI development for the past 75 years were wrong from the start—and as a result, true human-level AI is likely impossible. He points to Alan Turing’s foundational 1950 paper, which assumed human intelligence exists independently of the body and can therefore be replicated in software. Denning’s case is that this assumption was wrong then and remains wrong now, no matter how large our language models become.

I want to explain why this argument deserves serious attention, and why it’s directly relevant to every psychiatrist, psychologist, social worker and counselor who has ever worried that AI might one day replace them professionally. I also want to be honest about where I think Denning goes too far, because intellectual honesty serves the field better than reassurance does.

Turing’s two assumptions

Turing proposed two things in 1950 that shaped everything that followed. First, that intelligence can exist without a body—that it is, in principle, substrate-independent, meaning it could emerge from silicon as readily as from neurons. Second, that if a machine can exhibit behavior indistinguishable from a human, we should accept that it possesses intelligence (the “Turing Test”).

Denning challenges both. His core argument is that the most important dimensions of human intelligence—common sense, practical know-how, intuition, context and culture—are embodied. They live not just in our minds but in our bodies, our histories, our relationships and the communities and cultures that shaped us. They cannot be encoded as data because we do not know how to represent them as data. He calls this “the representation problem.”

“Whereas descriptions of skillful outcomes (‘know what’) can often be represented as bits and stored in a machine, we do not know how to encode the embodied knowledge for skillful performance (‘know how’).” —Denning (2026)

He uses the example of a virtuoso violinist: the musician can play beautifully but cannot fully describe how to produce that beauty, and a robot—lacking a biological body—could not grasp what the musician feels when playing. This gap between knowing what and knowing how is not a technical limitation that more compute will close. It is structural.

This argument is not new

Denning is not the first to make this case. Philosopher Hubert Dreyfus argued almost exactly the same thing in What Computers Can’t Do in 1972, drawing on phenomenological philosophy to contend that human expertise depends on embodied, contextual and situational knowledge that cannot be formalized. The AI community largely dismissed Dreyfus at the time. Decades later, his critics generally conceded he had identified something real, even if they disputed his conclusions about what could ultimately be achieved.

That history cuts two ways. On one hand, the fact that the embodied intelligence critique has persisted for over 50 years without being refuted—and has gained renewed traction as LLMs have scaled—suggests it points at something durable rather than a temporary technical gap. On the other hand, the history of AI is equally a history of human overconfidence about what machines would never be able to do. Chess masters insisted computers could not play at grandmaster level. They were wrong.

Caveat: Denning demonstrates persuasively that current architectures cannot acquire embodied, tacit or cultural knowledge. What he does not fully establish is that no future architecture could. His argument should be read as a serious challenge to AGI optimism, not a closed proof.

The four dimensions AI can’t currently reach

With that caveat in place, Denning identifies four categories of human intelligence that remain beyond current machines, regardless of model size or training data:

1. Tacit knowledge

The embodied know-how that cannot be fully articulated. This includes intuition, gut feelings, spontaneous creativity and the accumulated practical wisdom that emerges from doing something thousands of times in real contexts with real people. It is anchored in the body in ways we do not yet understand how to observe or measure.

2. Context

The layered, fractal circumstances that give words and actions their meaning. Context tells us whether someone is being sarcastic or sincere, angry or teasing, whether to be direct or gentle. Every context rests on previous contexts, which rest on earlier ones still, in an infinite and fractal regression. No model can fully capture that.

3. Culture

The shared values, norms, histories, power dynamics and moods that are the background assumptions of all human communication. Culture is not a dataset. It is lived. Scaling up LLMs with ever-larger neural networks will not enable them to acquire embodied cultural knowledge.

4. The mutual incomprehension divide

Denning’s most striking conclusion is that AI systems and humans may ultimately develop different forms of tacit knowledge that neither can fully understand. “Machines cannot read our tacit knowledge and we cannot read theirs,” he writes. “We are aliens across an uncrossable divide.” This also argues for humility about our own understanding of what machine cognition actually is—not just about its limits.

Why this maps almost perfectly to behavioral health

Consider what a skilled clinician actually does in a session. The therapeutic relationship, which decades of psychotherapy research identifies as the strongest predictor of outcomes, is built on precisely the capacities Denning says AI cannot acquire: empathic attunement, cultural responsiveness, the intuitive sense that something is off even when the client’s words say otherwise, and the wisdom accumulated from thousands of hours of sitting with human suffering.

The Society for the Advancement of Psychotherapy stated it directly in a 2025 article: high-stakes clinical contexts—including suicide risk assessment, mandated reporting, and severe psychopathological presentations—demand far more than data analysis. They require clinicians to integrate intuition, cultural sensitivity and emotional attunement into clinical conceptualization and treatment planning in real time.

A 2025 Stanford study found that AI chatbots expressed stigmas, responded inappropriately to critical situations, and could not safely replicate the therapeutic bond between clinician and client. That bond, the therapeutic alliance, is not a feature that can be added to a model. It emerges from the same embodied, contextual, culturally-embedded intelligence that Dreyfus described in 1972 and Denning is describing again now.

Put simply, the defining work of clinical practice sits in the categories that neither Dreyfus nor Denning believes machines can enter, and that the empirical record on AI in mental health settings consistently confirms they have not entered.

Where human judgment remains irreplaceable

Here is where AI cannot go, mapped against what behavioral health clinicians do:

Reading nonverbal and paralinguistic cues

A seasoned clinician notices the slight hesitation before a client answers, the way affect does not match content, the constricted gesture that contradicts a reported sense of relief. These are not data patterns extracted from text. They are real-time, embodied, contextual perceptions that require a witness who is also a body in the room.

Risk stratification under uncertainty

Deciding whether to pursue involuntary hospitalization for a client expressing suicidal ideation requires synthesizing immediacy of intent, access to means, protective factors, history of impulsive behavior, and the strength of the therapeutic alliance—while simultaneously holding that person’s autonomy, safety and the therapeutic relationship in tension. This is Denning’s tacit knowledge in its most high-stakes form.

Culturally responsive formulation

Culture is not a demographic checkbox. It is, as Denning describes, values, norms, judgments, histories, power and care. A clinician from the same community as a client brings embodied cultural knowledge that no model trained on text can replicate.

The therapeutic relationship itself

The alliance is co-created in the moment, sensitive to rupture, requiring repair, and dependent on genuine presence. OpenAI’s own data shows that approximately 1 million users per week show signs of emotional reliance on ChatGPT—which tells us something about the scale of unmet relational need, but nothing about AI’s capacity to actually meet it therapeutically.

Clinical intuition and pattern recognition

Differential diagnosis in psychiatry and psychology is rarely algorithmic. It involves what Dreyfus called “expert intuition” and what Denning calls “know how”—pattern recognition from thousands of hours of relational, embodied practice. The sense that this presentation feels like depression with psychotic features rather than primary psychosis is grounded in something in the texture of the interaction, the client’s history and accumulated clinical experience.

What AI can—and should—do

None of this means clinicians should dismiss AI. Quite the opposite. Whether or not Denning is right about the permanent impossibility of AGI, the division of labor his argument implies is the right one for now—and possibly for a long time.

AI is suited to what can be encoded: synthesizing structured information from documentation, identifying symptom patterns across aggregate data, reducing administrative burden, supporting measurement-based care by tracking outcomes across time and helping clinicians spend less time on documentation and more time on the work that is distinctly theirs.

The APA’s survey data shows this is already happening. As of 2025, 56% of psychologists have used AI to assist with their practice at least once, and the most common application is administrative and operational.

At Eleos, this is the specific commitment: AI that takes the documentation burden off providers so they can be more present in the work that defines good care. Both Dreyfus and Denning, separated by half a century, give it a philosophical foundation.

A note on the stakes

Denning closes with a warning: If machines cannot reliably interpret the unspoken context behind human intentions, aligning advanced AI systems with human goals is genuinely difficult. That has implications for every domain in which AI is deployed, and it has particular implications for behavioral health, where misaligned AI is potentially harmful.

The same structural features that make behavioral health clinicians hard to replace are the ones that make behavioral health AI potentially dangerous when deployed without appropriate human oversight. The clinician’s embodied, contextual, culturally-embedded judgment is the safeguard.

The bottom line

Denning has written a book that deserves engagement rather than either dismissal or uncritical acceptance. He is right that the most meaningful dimensions of clinical work are exactly where machines have made the least progress and where the theoretical barriers are the most serious. Dreyfus said the same thing more than 50 years ago, and the intervening decades have not proven him wrong on this point.

Where Denning may overreach is in concluding that the gap is permanently uncrossable. History counsels humility there. But for behavioral health clinicians asking a practical question—should I be worried that AI will replace me?—the honest answer is: not for the parts of your work that matter most. Those parts live exactly where machines can’t go.

The invitation is not to fear AI but to use it deliberately: let it do what it can do, so you can do more of what only you can.