The Illusion of Competence: How AI Tools Can Mask the Erosion of Clinical Judgment

Providing psychotherapy and conducting psychological assessments are among the most cognitively and emotionally demanding activities in professional practice. A psychologist must hold a client’s history, attend to what is said and what is not, monitor the relational moment, integrate theoretical frameworks, and make real-time clinical judgments, often under conditions of genuine uncertainty. This work is not only technically complex; it is personally taxing in ways that accumulate across a career. It is understandable, then, that psychologists would welcome tools that ease the burden.

Artificial intelligence (AI) offers exactly that relief. Large language models (LLMs) can generate polished clinical notes, draft case formulations, and suggest treatment plans with speed and apparent coherence. The output looks competent. The clinical reasoning behind it may be thin. This is the risk the Dunning-Kruger effect helps name: people with limited engagement in the actual cognitive work of a domain can overestimate their competence, because the products they produce no longer require them to demonstrate it (Kruger & Dunning, 1999). AI tools do not reduce psychologists’ expertise deliberately. They can do so quietly, by making it possible to appear highly competent while gradually losing the capacities that genuine competence requires. This article examines how that happens, and what psychologists can do about it.

The Difference Between Knowing and Appearing to Know

To understand the risk, it helps to be clear about what LLMs actually do. They are trained on enormous collections of text and generate responses through probabilistic pattern recognition. This means they predict the sequence of words most statistically likely to follow the input they receive. The process is powerful, often accurate, and fundamentally different from what a psychologist does when synthesizing a clinical case.

Call this distinction abstract versus concrete synthesis. An LLM generating a case formulation draws on patterns distilled from thousands of written clinical texts. The output may be coherent, theoretically grounded, and clinically plausible. But it is constructed from symbolic representations of clinical reality. This refers to descriptive or conceptual writing about clients’ experiences, patterns, identities, or clinical presentations, rather than direct interaction or engagement with clients themselves. For example, the psychologist working in a real session is doing something categorically different. The clinician is synthesizing what the client just said with the three-second pause that preceded it, the slight forward lean in the chair, and the catch in the voice when the client mentioned his father. Moreover, the psychologist is interpreting the relational moment. All of this unfolds against a backdrop of theoretical training refined through years of practice. LLMs cannot read the room. They cannot modulate the pace and tone of a therapeutic response in real time because they have no access to the room, the relationship, or the moment. Concrete synthesis is embodied. Abstract synthesis, however fluent, is not.

A note on a related concern: LLMs are trained on text collections that reflect existing cultural biases and historical inequities in mental health care. When a psychologist uses these tools without critical attention to culture and context, the outputs can reproduce assumptions that do not fit the client’s lived experience. The risk is not only that the formulation may be inaccurate, but that the psychologist, trusting the output, may not notice what was left out. This concern is real and merits sustained attention. A full treatment of diversity, equity, and cultural competence in AI-assisted practice is beyond the scope of this article. The focus here remains on the cognitive and developmental risks that arise even when the output appears culturally appropriate and clinically sound.

Automation Bias and the Abdication of Judgment

Clinical expertise is not accumulated by exposure to information. It is built through the slow, iterative work of applying judgment, getting it wrong, revising, and applying again. A psychologist who works through a difficult differential formulation independently, tolerates the uncertainty of an incomplete clinical picture, and then revises that thinking in light of new session material is doing something developmentally significant, even when the written product looks no different from one generated in seconds by an LLM. The process is the point (Greengrass, 2026). When AI tools consistently undermine that process, the outputs may remain adequate while the underlying reasoning capacity quietly atrophies. This matters most in moments the tools cannot anticipate: a subtle shift in presentation, an unexpected disclosure, the kind of in-session recalibration that depends on a clinician’s own sustained attention to the human relationship. Recognizing what does not fit, adapting mid-session to information the intake never predicted, and reasoning within the human relationship itself are not tasks an AI tool can perform on the clinician’s behalf. They depend on a capacity that is preserved only through continued intentional and deliberate practice.

Compounding this risk is automation bias. Research consistently shows that people tend to over-rely on automated recommendations, especially when those recommendations arrive in fluent, confident language (Goddard et al., 2011; Romeo & Conti, 2025). Even experienced psychologists are not immune. Studies have found that clinicians can adopt algorithmically generated interpretations without sufficient critical appraisal (Farmer et al., 2025). The broader research on human-AI collaboration points in the same direction: consistent reliance on algorithmic assistance can erode the capacity to perform the same task well without that assistance (Choudhury & Chaudhry, 2024).

Consider Dr. Mendez, a psychologist who begins using an LLM occasionally, perhaps to suggest a treatment plan or refine a differential. The tool is useful. Over time, the convenience of a quick, well-organized output begins to shift the workflow. The psychologist begins querying the LLM earlier in the process, eventually consulting it before engaging in independent analysis. The LLM’s output becomes the starting point, the anchor. The psychologist’s own clinical judgment, once the generative engine of the work, becomes the revision process. The outputs still look competent. But the sequence has quietly reversed. This is not clinical reasoning augmented by AI. It is clinical reasoning gradually displaced by it.

Dr. Mendez may not notice the shift. The clients may not notice. The record may show no obvious decline. And yet something essential has been lost: the habit of trusting one’s own cognitive and emotional capacities to do the work first.

Documentation Is More Than an Administrative Task

One of the more seductive applications of AI in psychological practice is clinical documentation. Ambient scribes promise to reduce the burden of note-writing, generating structured summaries that free up clinician time. For psychologists managing high caseloads, this is a genuine benefit. But it carries a risk.

Clinical documentation is not simply an administrative task. Writing a progress note is a reflective process. It requires the psychologist to organize observations, integrate theory, and make explicit the clinical decisions that shaped the session. These are the skills through which durable clinical reasoning is built and maintained (Sweller, 1988). A psychologist’s note is evidence not only of what occurred in a session, but of how the psychologist understood it.

Research reinforces this concern. Castro et al. (2026) found that AI-generated psychiatric documentation overemphasized signs and symptoms while minimizing treatment strategies and intervention rationale. In essence, the note narrates a clear story about the frequency and intensity of the client’s anxiety, but much less detail about how the psychologist is intervening and why. A psychologist who habitually accepts AI-generated notes without meaningful revision is not simply saving time. As a result, the clinician is bypassing the reflective process that consolidates clinical judgment in favor of documentation that summarizes content rather than clinical decisions.

This matters ethically. High-quality care is reflected in the integrity of the clinical record. Documenting thoughtfully, capturing treatment rationale, clinical hypotheses, and attention to contextual factors is itself an ethical responsibility. It reflects the psychologist’s accountability to the client, the record, and the profession. When documentation is outsourced without meaningful review, that ethical accountability is not delegated. It is quietly abandoned. Ethics is not a separate domain from clinical competence, for it is embedded in the commitment to doing the hard work well, even when a faster alternative is available.

Practical Recommendations

None of this argues against using AI in psychological practice. LLMs offer genuine value as consultation resources, prompts for critical analysis, and tools for broadening the range of hypotheses a clinician considers. The argument is about sequencing and stance. Several principles follow.

1. AI-generated formulations should follow rather than precede independent clinical reasoning. The psychologist who develops her own differential formulation and then consults an LLM to examine what she may have missed is doing something different from the psychologist who queries the LLM first. The first sequence sharpens clinical thinking. The second quietly replaces it. This is a choice worth making consciously rather than letting convenience decide.
2. AI-generated outputs should be treated as objects of critical analysis, not as drafts to be refined. Before accepting an LLM’s formulation, ask what it assumed, what it excluded, and how it compares to your own reasoning. This turns an AI interaction into a reflective exercise rather than a shortcut.
3. When reviewing AI-generated outputs, ask what cultural assumptions are embedded in the formulation. Does the output account for the client’s cultural context, identity, and lived experience, or does it default to generalized clinical patterns and majority culture expectations? A formulation that reads as coherent may still be culturally misaligned. Noticing this requires the psychologist’s own cultural competence, not the LLM’s fluency.
4. Documentation skills should be maintained independently of AI tools. Even for experienced psychologists, the act of writing a note from the session material is a reflective practice that consolidates clinical judgment. AI-assisted documentation should supplement that process, not supplant it. The question is not whether to use these tools, but how the psychologist uses the tool.

For those who supervise or consult, modeling a transparent stance toward AI use matters. When colleagues see only the polished output, not the moments of revision and rejection, they learn that AI engagement is private and informal. Making one’s reasoning visible, including the places where the AI got it wrong or missed the point, provides a concrete model of critical, accountable use.

Conclusion

When integrated without deliberate attention to their consequences, AI tools can quietly widen the gap between apparent and genuine competence, producing practitioners who are fluent in the language of clinical reasoning without having done the work from which that reasoning emerges. The antidote is not to reject these tools. It is to insist that they remain in their proper place: as supplements to clinical judgment, not substitutes for it.

The stakes extend beyond any individual practitioner. Psychology’s obligation to the public rests on the assurance that the professionals it credentials have genuinely earned and continue to demonstrate the competencies they represent. An illusion of competence, however convincing it may appear, does not meet that standard. Sitting with a client, reading the room, and doing the difficult cognitive and emotional work of understanding another person’s inner life cannot be generated by any algorithm. It must be practiced, again and again, remaining genuinely one’s own.