How AI Feedback Is Closing the CBT-for-Psychosis Gap

For 20 years, researchers, trainers, and clinical leaders have been trying to get evidence-based psychological treatments for serious mental illness into community settings. And for 20 years, they have failed. Fewer than 1% of Americans with a psychotic-spectrum disorder ever receive Cognitive Behavioral Therapy for psychosis (CBTp), the intervention that clinical practice guidelines say should be the standard of care. At NatCon 2026, a Lyssn, University of Washington, and King County partnership showed a working piece of what a solution might look like.

Three things every clinical leader should know

1. The 2024 American Law Institute Restatement of Torts shifted the legal definition of "reasonable care" from customary practice to evidence-based practice. If your organization is not delivering evidence-based care and does not have a plan to train up to it, your liability exposure just increased.
2. AI's most valuable role in mental health right now is not chatbots-as-therapists. It is feedback. Practitioners in most settings receive almost no direct feedback on their skills after training. AI can close that gap.
3. The training model that scales is the embedded clinical coach model. Asynchronous training plus group consultation plus AI-powered individual practice, with human coaches embedded in the agency. Nothing else is fast enough.

The Legal Landscape Just Moved. Most Clinical Leaders Missed It.

Dr. Sarah Kopelovich, Associate Professor of Psychiatry and Behavioral Sciences at the University of Washington and the first person in the country to hold a Professorship in CBT for Psychosis, opened her section of the panel with a legal update that most of the audience was hearing for the first time.

In 2024, the American Law Institute (ALI) issued a Restatement of Torts that redefined the bar for medical malpractice. The concept of "reasonable care" used to be defined as customary practice. If a clinician had not been trained in CBTp and nobody in the facility had been trained in CBTp, that clinician could not be held negligent for not providing it.

Under the 2024 restatement, reasonable care is now defined as evidence-based care. Which means that a clinician or an institution can now be held liable for negligence when evidence-based treatment exists and is not being offered.

"Legal accountability for implementation gaps is our new norm. That is our new reality. What that means is that we actually now have a pragmatic or legal mechanism for enforcing clinical practice guidelines."

— Dr. Sarah Kopelovich, University of Washington

The ALI Restatement is not law in every state. During the Q&A, one audience member correctly pointed out that state adoption is uneven and will take years. But the direction of travel is clear. Plaintiffs' attorneys are already citing the new standard, and it is aligning with a parallel federal push: SAMHSA's 2020 CBT for Psychosis Implementation Guide, of which Kopelovich was the lead author, is the first federal document to specify that CBTp should be the standard of care in mental health settings and that CBTp-informed care should be implemented in forensic, correctional, educational, and primary-care settings.

The pressure is economic, too. Kopelovich cited JAMA Psychiatry 2025 data using 2024 numbers putting the annual economic burden of untreated or under-treated schizophrenia-spectrum disorders at approximately $367 billion in the United States, of which about $70 billion is direct healthcare cost.

Why the Workforce Is the Bottleneck

Kopelovich's national point-prevalence estimate found that there are approximately 15 trained CBTp practitioners for every 10,000 Americans with a psychotic disorder. Fewer than 1% of people who need this treatment will ever get it. Meanwhile, more than 55% of U.S. counties have no psychologist, psychiatrist, or social worker at all.

"We simply don't have enough people to train. We cannot approach this problem through additive solutions. We need solutions that will function as force multipliers."

— Dr. Sarah Kopelovich

Kopelovich's approach at UW's SPIRIT Center is to think about two questions differently than the field traditionally has. Who are we training? and How are we training them? Instead of training only licensed doctoral-level clinicians (a group that shows heavy attrition through the training cycle), the SPIRIT model extends training to peer specialists, care case managers, care coordinators, and psychiatrists in CBTp-informed principles.

The "how" question then leads to learning science. The gold standard for skill acquisition is deliberate practice: repeated attempts at a specific skill with proximal, specific feedback. Kopelovich quoted Bruce Lee: "I fear not the man who has practiced 10,000 kicks. I fear the man who has practiced one kick 10,000 times." Which is exactly what an in-person trainer with a caseload cannot deliver at scale.

The Real AI Use Case: Feedback, Not Replacement

Dr. Zac Imel, Co-founder and Chief Science Officer of Lyssn and Professor at the University of Utah, framed the AI question at the panel bluntly. Read any article about AI in mental health and you will see one of two framings. Either the machine is coming to replace the therapist, or the machine is going to write the notes and stay in the background. Imel argues both framings miss the point.

"How can we build tools that support the expertise and human-centered work of mental health? How can we take tools that can scale certain parts of the mental health care system, like feedback for a therapist, that's reliable and scalable in a way that just would be almost impossible to do given the resources we have right now?"

— Dr. Zac Imel, Chief Science Officer, Lyssn

Lyssn's approach uses foundational open-source LLMs fine-tuned on more than 5,000 responses to simulated patient prompts, rated by clinical experts, trained clinicians, and lay people. This is important architecture context: Lyssn does not depend on a single frontier model, does not ship PHI out to consumer AI services, and can validate its own algorithms against expert human raters.

The Lyssn team set an inter-rater agreement threshold of 0.8 for its scoring algorithms. That is higher than the 0.77 typical benchmark for two expert human raters. In the CBTp trial data, the algorithms exceeded that bar across every skill module.

What the Training Tool Actually Does

The tool that Kopelovich's team developed with Lyssn is called CBTpro. It has four sections.

• Waiting room. Learners meet four simulated patients, review clinical presentations, chief complaints, diagnoses, and recovery goals.
• Chatbot section. Practice basic CBT concepts (collaborative empiricism, Guided Discovery) with a low-stakes chatbot before working with the standardized patient videos.
• Resource repository. Every client-facing handout developed for the tool is downloadable, including single-page psychoeducation on hearing voices.
• Skill modules. Seven modules corresponding to a course of CBTp, from psychoeducation and normalization through wellness planning. Each module includes a short explainer video, technical instruction on the skill, and then practice with the standardized patients.

The standardized patients are real actors, not AI renderings. They were filmed in collaboration with Kopelovich's clinical team to portray a range of clinical presentations including thought disorder, paranoia, and command-type auditory hallucinations.

"The first time that they're sitting with someone in the room with psychosis or schizophrenia, it's not with a real person where they say, 'Oh man, I hope this person's therapist comes in quickly, because I don't know if I can do this.' The first time that they're doing that is with a video where they can practice as many times as they want."

— The Lyssn moderator

The supervisor dashboard is the other half of the value. Supervisors can see, per learner, per program, or per agency, which skill modules have been attempted, which have been completed, how many attempts each took, and where learners are struggling.

The Data From the Randomized Trial

The CBTpro randomized controlled trial (NCT05127837), whose results were posted publicly in February 2026, enrolled roughly 100 practitioners across 64 clinical programs in five states and studied downstream outcomes in 300 of their clients.

Three findings from the trial stood out:

• Practitioners in the AI-enhanced training group demonstrated an improved working alliance with their clients compared to those who received only the open online course.
• Clients whose practitioners had the AI-enhanced training reported better self-perceived recovery at three months, sustained at six months, even though the practitioners only had active access to the tool for three months.
• The AI-trained cohort's clients showed reductions in persecutory ideation above and beyond the treatment-as-usual group.

Additional peer-reviewed evidence comes from the Kopelovich et al. 2025 field trial in Psychotherapy, which established the technical reliability of the AI feedback across eight CBT skills, and a 2026 JMIR Medical Education study on voluntary engagement patterns with virtual-patient CBT training tools.

The Embedded Clinical Coach Model: How King County Is Scaling

Dr. Brian Allender, Chief Medical Officer of the Behavioral Health Recovery Division of King County, Washington, laid out the operational implementation side. King County serves roughly 527,000 Medicaid recipients across 2.4 million residents and works with a network of over 50 outpatient behavioral health providers employing more than 4,000 client-facing staff.

King County uses the embedded clinical coach model, developed at UW's SPIRIT Center. The framework has three tiers:

1. Asynchronous online training and recorded didactics train the initial cohort of clinicians in an intervention.
2. Group consultation provides ongoing peer consultation, case presentations, and coaching in a group setting.
3. Embedded clinical coaches ("CBTp champions") are identified from clinicians who demonstrate skill and enthusiasm, and go on to run future rounds of training.

At the time of the panel, King County had trained roughly 150 clinicians in foundational CBTp, 40 in group CBTp, 40 in individual CBTp, and identified seven CBTp champions across seven agencies. The bottleneck the champions face is exactly the one Lyssn addresses: they cannot review sessions and give individual feedback to every trainee.

Frequently Asked Questions From the Audience

Q: Are you using ChatGPT or Gemini behind the scenes?

No. Imel emphasized that Lyssn started this work before frontier models existed. The company uses foundational open-source LLMs that it fine-tunes on its own annotated data. PHI does not leave the platform for consumer AI services.

Q: How does the tool decide whether the learner did well?

Currently, the tool provides a numeric score and pre-written qualitative feedback that is monotonically related to the score. Imel described a follow-on NIH trial that is developing more sophisticated feedback modeled on how learning science teaches math skills, with scaffolding tailored to learner performance.

Q: We're compliance officers. What are supervisors actually saying?

Kopelovich noted that as a trainer, the hardest thing to get clinicians to do is a role-play in a workshop. The tool makes practice private and low-stakes, which reduces avoidance. But there is still real avoidance around pressing "record," which is why the supervisor still matters. Human coaches check in when learners have not completed their modules.

Q: Does the training generalize to more severe presentations?

The simulated patients include thought disorder, paranoia, and command auditory hallucinations, but as Kopelovich noted, deliberate practice for cognitive-behavioral skills works best when the simulated patient is verbally engaged and organized enough for CBT. For clients who are severely thought disordered, the appropriate intervention is engagement (MI, befriending), not CBTp, so the training does not start there.

The Bottom Line

The picture from NatCon 2026 is coherent for the first time in years. Economic pressure, legal accountability, federal policy alignment, and AI capability are converging on the same answer: evidence-based psychotherapy for serious mental illness has to become a core competency of the behavioral health workforce. AI is not going to make therapists obsolete. It is going to make it possible, finally, to train them at the scale the country's mental health crisis requires.