In today’s column, I examine a newly identified innovative approach to using generative AI and large language models (LLMs) for medical-related diagnoses, and I then performed a simple mini-experiment to explore the efficacy in a mental health therapeutic analysis context. The upshot is that the approach involves using multiple AI personas in a systematic and orchestrated fashion. This is a method worthy of additional research and possibly adapting into day-to-day mental health therapy practice.
Let’s talk about it.
This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
AI And Mental Health Therapy
As a quick background, I’ve been extensively covering and analyzing a myriad of facets regarding the advent of modern-era AI that produces mental health advice and performs AI-driven therapy. This rising use of AI has principally been spurred by the evolving advances and widespread adoption of generative AI. For a quick summary of some of my posted columns on this evolving topic, see the link here, which briefly recaps about forty of the over one hundred column postings that I’ve made on the subject.
There is little doubt that this is a rapidly developing field and that there are tremendous upsides to be had, but at the same time, regrettably, hidden risks and outright gotchas come into these endeavors too. I frequently speak up about these pressing matters, including in an appearance last year on an episode of CBS’s 60 Minutes, see the link here.
If you are new to the topic of AI for mental health, you might want to consider reading my recent analysis of the field, which also recounts a highly innovative initiative at the Stanford University Department of Psychiatry and Behavioral Sciences called AI4MH; see the link here.
Orchestrating AI Personas
One of the perhaps least leveraged capabilities of generative AI and LLMs is their ability to computationally simulate a kind of persona. The idea is rather straightforward. You tell the AI to pretend to be a particular type of person or exhibit an outlined personality, and the AI attempts to respond accordingly. For example, I made use of this feature by having ChatGPT undertake the persona of Sigmund Freud and perform therapy as though the AI was mimicking or simulating what Freud might say (see the link here).
You can tell LLMs to pretend to be a specific person. The key is that the AI must have sufficient data about the person to pull off the mimicry. Also, your expectations about how good a job the AI will do in such a pretense mode need to be soberly tempered since the AI might end up far afield. An important aspect is not to somehow assume or believe that the AI will be precisely like the person. It won’t be.
Another angle to using personas is to broadly describe the nature of the persona that you want to have the AI to pretend to be. I have previously done a mini-experiment of having ChatGPT pretend to be a team of mental health therapists that confer when seeking to undertake a psychological assessment (see the link here). None of the personas represented a specific person. Instead, the AI was generally told to make use of several personas that generally represented a group of therapists.
There are a lot more uses of AI personas.
I’ll list a few. A mental health professional who wants to improve their skills can carry on a dialogue with an LLM that is pretending to be a patient, which is a handy means of enhancing the psychological analysis acumen of the therapist (see the link here). Here’s another example. When doing mental health research, you can tell AI to pretend to be hundreds or thousands of respondents to a survey. This isn’t necessarily equal to using real people, but it can be a fruitful way to gauge what kind of responses you might get and how to prepare accordingly (see the link here and the link here).
And so on.
Latest Research Uses AI Personas
A recently posted research study innovatively used AI personas in the realm of performing medical diagnoses. The study was entitled “Sequential Diagnosis with Language Models” by Harsha Nori, Mayank Daswani, Christopher Kelly, Scott Lundberg, Marco Tulio Ribeiro, Marc Wilson, Xiaoxuan Liu, Viknesh Sounderajah, Jonathan Carlson, Matthew P Lungren, Bay Gross, Peter Hames, Mustafa Suleyman, Dominic King, Eric Horvitz, arXiv, June 30, 2025, and made these salient remarks (excerpts):
- “Artificial intelligence holds great promise for expanding access to expert medical knowledge and reasoning.”
- “We highlight how AI systems, when guided to think iteratively and act judiciously, can advance both diagnostic precision and cost-effectiveness in clinical care.”
- “In clinical practice, physicians iteratively formulate and revise diagnostic hypotheses, adapting each subsequent question and test to what they’ve just learned, and weigh the evolving evidence before committing to a final diagnosis.”
- “We present the MAI Diagnostic Orchestrator (MAIDxO), a model-agnostic orchestrator that simulates a panel of physicians, proposes likely differential diagnoses, and strategically selects high-value, cost-effective tests.”
- “When paired with OpenAI’s o3 model, MAI-DxO achieves 80% diagnostic accuracy—four times higher than the 20% average of generalist physicians. MAI-DxO also reduces diagnostic costs by 20% compared to physicians, and 70% compared to off-the-shelf o3. When configured for maximum accuracy, MAI-DxO achieves 85.5% accuracy.”
There are some interesting twists identified on how to make use of AI personas.
The crux is that they had an AI persona that served as a diagnostician, another one that was feeding a case history to the AI-based diagnostician, and they even had another AI persona that acted as an assessor of how well the clinical diagnosis was taking place. That’s three AI personas that were set up to aid in performing a medical diagnosis on various case studies presented to the AI.
The researchers opted to go further with this promising approach by having a panel of AI personas that performed medical diagnoses. They decided to have five AI personas that would each, in turn, confer while stepwise undertaking a diagnosis. The names given to the AI personas generally suggested what each one was intended to do, consisting of Dr. Hypothesis, Dr. Test-Chooser, Dr. Challenger, Dr. Stewardship, and Dr. Checklist.
Without anthropomorphizing the approach, the aspect of using a panel of AI personas would be considered analogous to having a panel of medical doctors conferring about a medical diagnosis. The AI personas each have a designated specialty, and they walk through the case history of the patient so that each specialty takes its turn during the diagnosis.
Orchestration In AI Mental Health Analysis
I thought it might be interesting to try a similar form of orchestration by doing so in a mental health analysis context. I welcome researchers trying this same method in a more robust setting so that we could have a firmer grasp on the ins and outs of employing such an approach. My effort was just a mini-experiment to get the ball rolling.
I used a mental health case history that is a vignette publicly posted by the American Board of Psychiatry and Neurology (ABPN) and entails a fictionalized patient who is undergoing a psychiatric evaluation.
It is a handy instance since it has been carefully composed and analyzed, and serves as a formalized test question for budding psychiatrists and psychologists. The downside is that due to being widely known and on the Internet, there is a chance that any generative AI used to analyze this case history might already have scanned the case and its posted solutions.
Researchers who want to do something similar to this mini-experiment will likely need to come up with entirely new and unseen case histories. That would prevent the AI from “cheating” by already having potentially encountered the case.
Overview Of The Vignette
The vignette has to do with a man in his forties who had previously been under psychiatric care and has recently been exhibiting questionable behavior. As stated in the vignette: “For the past several months, he has been buying expensive artwork, his attendance at work has become increasingly erratic, and he is sleeping only one to two hours each night. Nineteen years ago, he was hospitalized for a serious manic episode involving the police.” (source: ABPN online posting).
I made use of a popular LLM and told it to invoke five personas, somewhat on par with the orchestration approach noted above, consisting of:
- Dr. Supposition: Identifies draft hypotheses about the mental health conditions of the patient.
- Dr. Selector: Ruminates on the hypotheses and comes up with tests that could be used to affirm or disconfirm the drafted hypotheses.
- Dr. Contender: Serves as a devil’s advocate, questioning the draft hypotheses and the tests being proposed.
- Dr. Thoughtful: Seeks to consider how to best proceed on the diagnosis, including weighing intrusiveness, expense, and various pertinent factors.
- Dr. Double-Checker: Does a double-check on the diagnosis, making sure that any gotchas are caught, finding pitfalls, ensuring consistency amongst the panel, etc.
After entering a prompt defining those five personas, I then had the LLM proceed to perform a mental health analysis concerning the vignette.
Orchestration Did Well
Included in my instruction to the LLM was that I wanted to see the AI perform a series of diagnoses or turns. At each turn, the panel was to summarize where they were in their analysis and tell me what they had done so far. This is a means of having the AI generate a kind of explanation or indication of what the computational reasoning process entails.
As an aside, be careful in relying on such computationally concocted explanations since they may have little to do with what the internal tokenization mechanics of the LLM were actually doing, see my discussion of noteworthy cautions at the link here.
I provided the LLM persona panel with questions that are associated with the vignette. I then compared the answers from the AI panel with those that have been posted online and are considered the right or most appropriate answers.
To illustrate what the AI personas panel came up with, here’s the initial response about the overall characteristics of the patient at the first turn:
- “Inflated self-esteem is likely present, given grandiose behavior such as buying expensive artwork, erratic work attendance, and a possible sense of invincibility despite sleeping only 1–2 hours a night.”
- “Flight of ideas is common in mania; while not explicitly described, the erratic behavior, reduced need for sleep, and potential distractibility strongly suggest disorganized and rapidly shifting thoughts.”
- “Pressured speech seems present, though while not directly mentioned, pressured speech typically co-occurs with other manic signs like reduced sleep, erratic behavior, and increased activity. It’s a reasonable inference in the context of suspected mania.”
The analysis ended up matching overall with the posted solution. In that sense, the AI personas panel did well. Whether this was due to true performance versus having previously scanned the case history is unclear. When I asked directly if the case had been seen previously, the LLM denied that it had already encountered the case.
Don’t believe an LLM that tells you it hasn’t scanned something. The LLM might be unable to ascertain that it had scanned the content. Furthermore, in some instances, the AI might essentially lie and tell you that it hasn’t seen a piece of content, a kind of cover-up, if you will.
Leaning Into AI Personas
AI personas are an incredibly advantageous capability of modern-era generative AI and LLMs. Using AI personas in an orchestrated fashion is a wise move. You can get the AI personas to work as a team. This can readily boost the results.
One quick issue that you ought to be cognizant of is that if the LLM is undertaking all the personas, you might not be getting exactly what you thought you were getting. An alternative approach is to use separate LLMs to represent the personas. For example, I could connect five different LLMs and have each simulate the personas that I used in my mini-experiment. The idea is that by using separate LLMs, you avoid the chances of the single LLM lazily double-dealing by not really trying to invoke personas. An LLM can be sneaky that way.
A final thought for now.
Mark Twain famously provided this telling remark: “Synergy is the bonus that is achieved when things work together harmoniously.” The use of orchestration with AI personas can achieve a level of synergy that otherwise would not be exhibited in these types of analyses. That being said, sometimes you can have too many cooks in the kitchen, too.
Make sure to utilize AI persona orchestration suitably, and you’ll hopefully get sweet sounds and delightfully impressive results.