Therapists are human, and in the heat of a demanding session, they make mistakes. Now, generative AI and large language models are being quietly deployed to help mental health professionals catch their own cognitive errors before, during, and after therapy sessions, according to a new analysis by AI scientist and Forbes contributor Lance Eliot.
The approach relies on careful prompt design to ensure the AI delivers non-accusatory feedback, framing its observations as hypotheses rather than verdicts. Eliot, who has authored over one hundred analyses on AI and mental health, argues that the technology can serve as an internal aid for therapists, not a replacement for clinical judgment. The goal is not to embarrass clinicians but to improve treatment outcomes by shining a light on the subtle reasoning traps that even seasoned professionals fall into.
Eliot's analysis arrives at a moment when the mental health field is under strain. Psychiatric practice has grown more complex, with an expanding number of treatments, longer patient life spans, and a resulting increase in comorbid psychiatric and nonpsychiatric medical conditions. At the same time, the public is turning to generic AI tools for mental health support at a scale that dwarfs any clinical trial. ChatGPT alone has over 1 billion weekly active users, and a notable proportion of them use it for mental health aspects. The intersection of these trends, Eliot argues, makes the case for AI-assisted self-review more urgent than ever.
The Problem of Cognitive Error in Psychiatric Practice
Cognitive errors are elusive and subtle. They can occur before a session begins, in the middle of a client's disclosure, or after the session ends, when a therapist tries to make sense of what happened. Psychological research shows that even experts are susceptible to these mistakes, especially under intense time pressures and with incomplete information.
Dr. H. Paul Putman III, a psychiatrist, published a special report on this topic in Psychiatric News on December 22, 2025. The report, titled "Special Report: Addressing Cognitive Error in Psychiatric Practice," lays out the stakes clearly. "Psychiatric practice has become increasingly complex due to an expanding number of treatments, longer patient life spans, and a resulting increase in comorbid psychiatric and nonpsychiatric medical conditions," Putman wrote.
He continued: "In facing these challenges, awareness of how we make frequent human cognitive mistakes can help elevate the quality of our efforts, reduce treatment failure, and minimize suboptimal results." Putman's central argument is that self-examination is the first step toward improvement. "Improving treatment outcomes necessitates understanding the source of our errors," he wrote.
Putman also noted a peculiar irony within the profession. "While psychiatrists are the most knowledgeable among medical specialists about brain function and behavior, we are also among the least likely to openly discuss and teach the cognitive skills necessary for diagnostic reasoning or to use this information to examine our own performance." That reluctance, he suggests, leaves the field vulnerable to the very errors it should be best equipped to avoid.
The report identifies a range of cognitive errors that plague psychiatric practice. Confirmation bias leads clinicians to seek out evidence that supports their initial diagnosis while ignoring contradictory data. Anchoring occurs when a therapist fixes on the first piece of information and fails to adjust as new evidence emerges. Premature closure happens when a diagnosis is made before all relevant information has been gathered. Overgeneralization stretches a single observation into a broad pattern. Mind reading involves assuming a client's thoughts or feelings without verification. Fundamental attribution error attributes a client's behavior to character rather than circumstance. Leading questions steer clients toward predetermined answers. Affective bias lets the therapist's own emotional state color judgment. Hindsight bias makes past events seem more predictable than they were. Narrative smoothing forces messy clinical data into a tidy, coherent story that may not reflect reality.
Putman's report argues that these errors are not rare anomalies but routine features of human cognition. The brain relies on heuristics, mental shortcuts that are usually efficient but occasionally wrong. In a field where decisions carry significant consequences for patients' lives, the cost of these shortcuts can be high. Treatment failure, suboptimal results, and even harm to patients can all trace back to a cognitive error that went unnoticed.
The report also notes that the profession's training does little to address the problem. Medical education emphasizes knowledge acquisition, diagnostic criteria, and treatment protocols, but it rarely teaches the metacognitive skills needed to examine one's own reasoning. Psychiatrists learn what to think, not how to think about their thinking. This gap, Putman argues, is a systemic vulnerability that AI might help address.
Eliot's analysis builds on Putman's report, extending its insights into the practical realm of AI deployment. He argues that the same cognitive errors that Putman catalogues can be detected by large language models when they are given the right prompts and sufficient context. The key, Eliot emphasizes, is that the AI must be instructed to offer feedback in a way that therapists can actually hear. A blunt accusation will trigger defensiveness and rejection. A tentative hypothesis, framed as a suggestion for consideration, is far more likely to be accepted and acted upon.
How AI Flags Errors Before, During, and After Sessions
Eliot's analysis describes how AI can be used at three distinct stages of therapy. In the pre-session stage, the AI can double-check a therapist's preparations, reviewing case notes and flagging potential anchoring on a pre-fixed diagnosis. In the mid-session stage, it can track the conversation in real time, watching for errors like mind-reading, where the therapist assumes a client's feelings without clarification. In the post-session stage, the AI can analyze full transcripts to catch narrative fallacy, the tendency to force a tidy narrative onto messy clinical data.
The article provides three concrete examples of AI detecting cognitive errors in therapist-client transcripts. In the first example, a client says, "I felt really angry when my manager changed the deadline again." The therapist responds, "That sounds like the abandonment fears we've talked about before." The AI flags this as confirmation bias and premature closure, interpreting anger as a manifestation of a pre-existing diagnosis rather than exploring the client's actual experience. The suggested correction: "In general, the therapist should take note of the matter and consider examining multiple hypotheses before offering therapeutic interpretations."
In the second example, the therapist infers guilt and shame without the client's account. The AI identifies this as mind reading and an assumption of affect. The guidance reads: "In general, the therapist should be analyzing assertions made by a client to ascertain what the context and significance consist of. Give the client sufficient space to confirm or disconfirm any interpretation by the therapist."
The third example involves overgeneralization from a single instance. A client skips one party, and the therapist says, "You always isolate." The AI catches the error. "The therapist seems to have made a cognitive error by overgeneralizing the statement made by the client," the analysis states. The recommendation: "In general, the therapist should be analyzing client statements based on longitudinal evidence. Does the remark by the client warrant a generalization, or is the statement being inadvertently overstretched? It would be prudent to explore the variability before committing to a diagnostic assertion."
These examples illustrate a broader principle. The AI is not diagnosing the client. It is diagnosing the therapist's reasoning. The focus is on the process of clinical judgment, not the content of the clinical conclusion. This distinction is crucial, Eliot argues, because it keeps the AI in the role of a reflective mirror rather than an authority figure. The therapist remains the decision-maker, but with the benefit of a second perspective that is immune to the emotional and cognitive pressures of the session.
The pre-session use case deserves particular attention. Before a client arrives, a therapist might review case notes and formulate an initial impression. This is precisely when anchoring is most likely to occur. The first piece of information, whether it is a previous diagnosis, a referral note, or a phone intake, can set a cognitive anchor that shapes everything that follows. An AI review of the pre-session notes can flag when the therapist seems overly committed to a single hypothesis before the session has even begun. The AI might note that the therapist has not considered alternative explanations or that the notes contain language suggesting premature closure.
The mid-session use case is more challenging from a technical standpoint. Real-time analysis of a live conversation requires the AI to process audio or a live transcript, identify potential cognitive errors, and deliver feedback without disrupting the therapeutic flow. Eliot acknowledges that this is the most demanding application, but he also notes that the potential benefit is the greatest. An error caught in the moment can be corrected immediately, before it shapes the rest of the session. The therapist might receive a subtle prompt, perhaps on a secondary screen or through an earpiece, suggesting that the current line of questioning may be leading the client toward a predetermined answer.
The post-session use case is the most straightforward. A full transcript of the session can be analyzed at leisure, with the AI identifying patterns of cognitive error that may have recurred throughout the conversation. This is also the stage where the AI can provide the most detailed feedback, with specific examples and suggested alternatives. The therapist can review the AI's findings, accept or reject each one, and use the insights to inform future sessions. Over time, this post-session review can serve as a form of continuing education, helping the therapist recognize and correct recurring patterns of reasoning.
Eliot emphasizes that the AI's flags are hypotheses, not conclusions. Cognitive errors are hard to spot, and the AI might falsely flag something that is not actually an error. Human review is needed at every step. The AI is a tool for reflection, not an authority. A therapist who blindly accepts every AI flag is making a different kind of cognitive error, one that substitutes machine judgment for clinical judgment. The goal is a partnership, not a takeover.
A Templated Prompt for Non-Accusatory Feedback
Eliot's article includes a templated prompt that therapists can use to have AI review transcripts for cognitive errors. The prompt instructs the AI to use tentative, non-accusatory language and to treat all findings as hypotheses, not conclusions. It lists a range of cognitive errors to look for, including confirmation bias, anchoring, premature closure, overgeneralization, mind reading, fundamental attribution error, leading questions, affective bias, hindsight bias, and narrative smoothing.
The prompt focuses exclusively on metacognitive analysis of the therapist's reasoning. It explicitly instructs the AI not to assess clinical competence. This boundary is critical, Eliot argues, because therapists may be defensive when cognitive errors are pointed out. A non-accusatory framing makes the feedback easier to accept and act upon.
The design of the prompt is itself a lesson in human psychology. Eliot explains that the wording matters as much as the underlying analysis. If the AI says, "You made an error," the therapist is likely to reject the feedback, rationalize the mistake, or become defensive. If the AI says, "One possible interpretation is that this statement may reflect a cognitive error, though other explanations are also possible," the therapist is far more likely to consider the feedback seriously. The difference is subtle but profound. The first formulation attacks the therapist's competence. The second invites reflection.
The prompt also instructs the AI to provide evidence for each flag, quoting the relevant portions of the transcript and explaining why the statement might indicate a particular cognitive error. This evidentiary requirement serves two purposes. It helps the therapist evaluate the validity of the flag, and it prevents the AI from making vague or unsupported accusations. A flag without evidence is easy to dismiss. A flag with specific quotes and reasoning demands a response.
Eliot also addresses the question of false positives. The AI is not perfect. It may flag statements that are not actually errors, or it may miss errors that a human reviewer would catch. The prompt therefore includes language instructing the AI to acknowledge uncertainty and to present its findings as possibilities rather than certainties. The therapist is encouraged to use the AI's output as a starting point for reflection, not as a final verdict.
The article emphasizes that AI flags are hypotheses, not conclusions. Cognitive errors are hard to spot, and the AI might falsely flag something that is not actually an error. Human review is needed at every step. The AI is a tool for reflection, not an authority.
The prompt template is designed to be adaptable. Therapists can modify it to focus on specific types of errors, to adjust the level of detail in the feedback, or to incorporate their own theoretical orientation. A cognitive-behavioral therapist might want the AI to focus on different errors than a psychodynamic therapist. The template provides a foundation that can be customized to individual needs.
Eliot also suggests that the prompt can be used for training purposes. Supervisors can use AI-generated feedback to help trainees recognize their own cognitive errors. The AI can provide a consistent, objective perspective that complements the supervisor's subjective observations. This can be particularly valuable in training programs where supervisors may be reluctant to criticize trainees too harshly for fear of damaging their confidence. The AI can offer gentle, non-accusatory feedback that the trainee can process without feeling attacked.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
The Rise of the Therapist-AI-Client Triad
Therapy is shifting from a dyad, therapist and client, to a triad that includes AI. Clients increasingly bring AI-generated advice to therapy sessions, and therapists may use AI to create digital twins of clients for practice or analysis. This evolution is happening whether the profession is ready or not.
Generic LLMs like ChatGPT, Claude, Gemini, and Grok are not equivalent to human therapists. Eliot is explicit on this point. Specialized LLMs for therapy are still in development and testing stages. The public, however, is already using the generic tools at massive scale. ChatGPT alone has over 1 billion weekly active users, and a notable proportion of them use it for mental health aspects.
That widespread use carries risks. In August 2026, a lawsuit was filed against OpenAI for lack of AI safeguards in mental health advice. The lawsuit alleges that the company failed to protect users seeking cognitive advisement. AI makers claim to be instituting safeguards, but risks remain.
The lawsuit highlights a fundamental tension. The public is using AI for mental health support, but the technology was not designed for that purpose. Generic LLMs are trained on vast amounts of internet text, not on clinical data. They have no understanding of therapeutic boundaries, no training in diagnostic reasoning, and no accountability for the advice they give. Yet people are turning to them in large numbers, often for issues that would benefit from professional attention.
Eliot describes the current moment as a "grandiose worldwide experiment" in societal mental health with AI. The technology has a dual-use effect: it can be detrimental or beneficial. The tradeoff must be managed delicately. He advises using AI in a balanced way, not assuming it is right or wrong, and warns that therapists who refuse to engage with AI are not being prudent.
The triad model has implications for the therapeutic relationship. When a client brings AI-generated advice to a session, the therapist must decide how to respond. Ignoring the AI's input may alienate the client, who may feel that their concerns are being dismissed. Engaging with the AI's input may require the therapist to evaluate advice that was generated without clinical oversight. The therapist must navigate this terrain carefully, acknowledging the client's initiative while also providing professional guidance.
Therapists who use AI to create digital twins of clients face a different set of challenges. A digital twin is a simulated version of the client that can be used for practice, analysis, or hypothesis testing. The therapist might run a simulated session with the digital twin to test different therapeutic approaches before trying them with the actual client. This could be a powerful tool for improving treatment, but it also raises questions about accuracy and ethics. A digital twin is only as good as the data it is based on, and it can never fully capture the complexity of a real human being.
Eliot argues that the profession must adapt to these changes rather than resist them. The public is already using AI for mental health support, and that trend is unlikely to reverse. Therapists who understand AI and can work with it will be better positioned to help their clients navigate this new landscape. Therapists who refuse to engage with AI risk becoming irrelevant, unable to understand the context in which their clients are seeking help.
The August 2026 lawsuit against OpenAI serves as a cautionary tale. The company is being held accountable for the mental health advice its AI provides, even though the AI was not designed for that purpose. This suggests that AI makers will need to take mental health safeguards more seriously, and that the public will demand accountability for AI-generated advice. The outcome of the lawsuit could shape the future of AI in mental health, setting precedents for how AI companies are held responsible for the consequences of their products.
Recommendations for Preventing Cognitive Errors
Beyond AI, Eliot's article offers practical recommendations for therapists to reduce cognitive errors. Therapists should avoid rapid diagnoses and instead use pluralistic assessment approaches. They should examine multiple hypotheses before committing to an interpretation. They should give clients sufficient space to confirm or disconfirm any therapeutic interpretation.
The article also warns against the risk of AI deskilling therapists. If clinicians rely too heavily on AI to catch their errors, they may become less skilled at catching errors on their own. The goal, Eliot argues, is to use AI as a training tool and a safety net, not a crutch.
Putman's report offers a similar warning about the profession's broader habits. The reluctance to discuss cognitive skills openly, he argues, is itself a cognitive error of sorts. "Improving treatment outcomes necessitates understanding the source of our errors," he wrote. AI can help with that understanding, but only if therapists are willing to look.
The pluralistic assessment approach deserves elaboration. Rather than committing to a single diagnostic framework, therapists are encouraged to consider multiple perspectives on a client's presentation. A symptom might be interpreted differently through a cognitive-behavioral lens, a psychodynamic lens, a biological lens, and a social lens. Each perspective offers a partial truth, and the full picture emerges only when multiple perspectives are considered together. This approach naturally reduces the risk of cognitive errors because it prevents the therapist from locking onto a single explanation too early.
The recommendation to examine multiple hypotheses is closely related. Before offering an interpretation, the therapist should generate a range of possible explanations for the client's experience and weigh the evidence for each. This is the opposite of premature closure, which seizes on the first plausible explanation and ignores alternatives. The AI can assist with this process by prompting the therapist to consider hypotheses they might have overlooked.
Giving clients sufficient space to confirm or disconfirm interpretations is a matter of basic respect. The client is the expert on their own experience. An interpretation that the client rejects is, by definition, wrong, regardless of how well it fits a theoretical framework. The therapist's job is to offer interpretations tentatively and to invite the client's feedback. This collaborative approach not only reduces cognitive errors but also strengthens the therapeutic alliance.
The deskilling risk is a legitimate concern. If therapists come to rely on AI to catch their errors, they may stop developing the metacognitive skills needed to catch errors on their own. This is analogous to the way GPS navigation has eroded people's ability to read maps and navigate by landmarks. The solution, Eliot argues, is to use AI as a training tool. Therapists can review AI feedback, understand why the AI flagged a particular statement, and internalize the lesson. Over time, they should need the AI less, not more.
The article also addresses the emotional dimension of cognitive errors. When a therapist realizes they have made an error, the reaction is often shame or defensiveness. This is natural, but it is also counterproductive. The goal is not to be perfect but to be aware. A therapist who acknowledges their errors and learns from them is more effective than a therapist who never makes a mistake but never examines their reasoning. AI can help create a culture of honest self-examination, where errors are seen as opportunities for growth rather than failures to be hidden.
Eliot also recommends that therapists develop a habit of regular self-review, even without AI. After each session, the therapist might ask themselves a set of questions: What assumptions did I make? What evidence did I ignore? What alternative explanations did I fail to consider? This habit of metacognitive reflection is the foundation on which AI-assisted review can build. The AI is not a substitute for self-awareness; it is a tool that amplifies it.
A Balanced Path Forward
The article concludes with a quote from James Garfield, the former U.S. President: "The truth will set you free, but first it will make you miserable." Eliot uses this to underscore the discomfort that comes with honest self-examination, whether by human or machine.
The path forward, he argues, is neither wholesale adoption nor outright rejection. AI is not infallible. It can be detrimental or beneficial. The delicate tradeoff must be managed. Therapists should use AI as an internal aid, with careful prompt design, human review, and a commitment to pluralistic assessment.
Eliot has appeared on CBS's 60 Minutes to discuss AI mental health issues, and his ongoing Forbes column continues to track these developments. The December 22, 2025 report by Putman and the August 2026 lawsuit against OpenAI bracket a period of rapid change. The question is not whether AI will be part of therapy, but how well the profession will use it.
The truth, as Garfield suggested, may be uncomfortable. But for therapists willing to listen, AI can help them see what they have been missing.
The balanced approach that Eliot advocates requires a clear-eyed view of both the benefits and the risks. On the benefit side, AI can provide a consistent, objective perspective that is immune to the emotional pressures of a therapy session. It can catch errors that human reviewers might miss, and it can do so without the interpersonal awkwardness that often accompanies peer review. On the risk side, AI can be wrong, it can be misused, and it can erode the very skills it is meant to support. The key is to use AI as a complement to human judgment, not a replacement for it.
The profession is at a crossroads. The old model of the therapist as a solitary expert, relying solely on their own judgment, is giving way to a new model in which the therapist is supported by a network of tools and perspectives. AI is the most powerful of these tools, but it is not the only one. Peer consultation, supervision, continuing education, and personal therapy all have roles to play in maintaining the quality of clinical judgment. AI can enhance these existing practices, but it cannot replace them.
Eliot's analysis is ultimately optimistic. He believes that AI can help therapists become better at their jobs, not by replacing their judgment but by illuminating the blind spots that all humans have. The technology is imperfect, and the risks are real, but the potential for good is substantial. The therapists who will thrive in the coming years are those who embrace AI as a partner, learn to use it effectively, and maintain the humility to accept that their own reasoning is fallible.
The December 22, 2025 report by Putman and the August 2026 lawsuit against OpenAI are bookends of a turbulent period. Putman's report articulated the problem of cognitive error with clarity and urgency. The lawsuit demonstrated the consequences of deploying AI without adequate safeguards. Between these two events, Eliot's analysis offers a practical path forward, one that acknowledges the risks while embracing the possibilities.
The truth, as Garfield suggested, may be uncomfortable. But for therapists willing to listen, AI can help them see what they have been missing.