AI Chatbot Trains Residents in Code Status Talks in Pilot Randomized Trial

When a patient is admitted to a hospital in serious condition, one of the most consequential conversations a physician can have is about code status: whether, if the heart stops or breathing fails, the care team should attempt resuscitation. These discussions shape the course of end-of-life care, yet many young doctors receive little structured practice in how to conduct them. A new pilot randomized trial published in the Journal of General Internal Medicine suggests that a purpose-built artificial intelligence chatbot may offer a scalable way to fill that training gap, with residents who practiced with the tool showing measurable, though not statistically significant, improvements in their code status discussion skills.

The study, led by Dr. Jessica R. Lichter of the Icahn School of Medicine at Mount Sinai and NYC Health + Hospitals/Elmhurst, took place at a public safety-net hospital affiliated with an academic medical center. The research team recruited internal medicine residents at the postgraduate year two and three levels, the stages at which trainees increasingly handle serious conversations with patients and families on their own. Thirty-one of 41 eligible residents enrolled in the study, and 25 completed all phases of the protocol, providing the dataset for the final analysis.

The intervention centered on ChatMD, a novel AI chatbot designed to simulate realistic code status discussions. The platform was conceived and developed by Dr. Lichter together with Dr. Alex Beutel, an independent technical contributor whose work on the project was conducted outside his employment at OpenAI. ChatMD plays the role of a hospitalized patient confronting decisions about resuscitation preferences, allowing trainees to rehearse the delicate blend of medical explanation, empathy, and shared decision-making that these conversations demand, without the logistical burden of arranging standardized patient sessions.

All participants began with a shared lecture on the elements of effective code status discussions, ensuring a common foundation before randomization. Residents assigned to the intervention then completed one hour of supervised practice with ChatMD, followed by a faculty-led debriefing modeled on established simulation debriefing principles, and an additional hour of independent practice with the chatbot. Residents in the control group continued with their usual clinical training during the study period. Roughly eight weeks later, with a standard deviation of about two weeks, all enrolled residents completed a standardized patient encounter designed to test their code status discussion skills in a realistic setting.

The standardized patient encounters were graded by assessors blinded to which arm each resident had been assigned, using a modified version of a previously validated checklist. On overall code status discussion performance, intervention residents scored 72.5 percent compared with 66.0 percent among controls. The gap was wider on skills specific to code status conversations, where the intervention group achieved 71.4 percent against 57.6 percent for the control group. However, with the small sample size inherent to a pilot feasibility study, neither difference reached statistical significance, with p-values of 0.35 and 0.15 respectively.

The researchers complemented the quantitative results with a qualitative analysis of resident feedback, using established thematic analysis methods. Intervention residents consistently valued ChatMD for what educators call deliberate practice: the opportunity to repeat a difficult conversation multiple times, experiment with different phrasings, and receive immediate feedback without fear of harming a real patient. Many reported increased confidence in approaching code status discussions on the wards. At the same time, residents identified areas for improvement, particularly the emotional realism of the chatbot’s responses and the depth of the feedback it provided, pointing to clear directions for the next generation of the tool.

The stakes of this training problem are well documented in the medical literature. Prior studies have shown that hospital code status discussions are often rushed, inconsistently conducted, and frequently fail to incorporate patient values, contributing to do-not-resuscitate orders that may not reflect genuine patient preferences. Earlier educational interventions, including mastery learning programs for hospitalists and curricula for trainees, have demonstrated that these skills can be taught, but such programs typically require trained faculty, standardized patients, and protected time, resources that many institutions struggle to sustain. An AI-based simulator that residents can access independently could, in principle, extend this kind of training far beyond what traditional methods allow.

The trial also sits within a rapidly expanding body of work on artificial intelligence in medical education. Recent reports have described AI-simulated patient interactions for improving medical interview skills and multimodal communication training with AI-assisted video analysis for physicians caring for older adults. A 2023 scoping review catalogued a growing range of AI tools supporting communication skills training across the health professions. At the same time, researchers have flagged cautionary considerations, including the tendency of large language models toward sycophancy, or excessive agreement with users, which could undermine the realism of a simulated patient who should sometimes push back or express distress. The Macy Foundation innovation reports on AI in medical education have emphasized both the promise and the integration challenges facing such technologies.

The study team has been transparent about the limitations inherent in a pilot design. The sample of 25 completers was too small to detect anything but large effects, and the non-significant findings should be read as signals of promise rather than proof of efficacy. The single-site setting at a safety-net hospital, while valuable for testing feasibility in a resource-constrained environment, raises questions about generalizability. The eight-week interval between training and assessment is a strength in that it tested retention rather than immediate recall, but longer-term durability of skills remains unexamined. The authors themselves frame the results as supporting continued development of AI-based simulation for code status training rather than as a call for immediate adoption.

Nevertheless, the direction of the findings is notable. The largest observed difference was in the skills most specific to code status conversations, exactly the domain where a purpose-built chatbot might be expected to add value beyond general clinical experience. And the qualitative enthusiasm of residents for on-demand, judgment-free practice suggests that the mechanism of change, deliberate repetition with feedback, is one that AI tools are uniquely positioned to deliver. As the technology matures and larger trials become feasible, AI-simulated patients could become a routine complement to standardized patient encounters, giving every trainee the hours of practice that mastering one of medicine’s hardest conversations requires. For now, this pilot trial offers an early, carefully measured glimpse of what that future might look like.

Subject of Research: AI-based simulation training for code status discussion skills among internal medicine residents

Article Title: Artificial Intelligence-Based Simulation to Improve Code Status Discussions Among Internal Medicine Residents: A Pilot Randomized Trial

Article References: Lichter, J. R., Beutel, A., Trottier, D., Forsyth, B., & Kanevsky, J. (2026). Artificial Intelligence-Based Simulation to Improve Code Status Discussions Among Internal Medicine Residents: A Pilot Randomized Trial. Journal of General Internal Medicine. https://doi.org/10.1007/s11606-026-10761-4

Image Credits: AI Generated

DOI: 10.1007/s11606-026-10761-4

Keywords: artificial intelligence, medical education, code status discussions, internal medicine residents, simulation, chatbot, palliative care, communication skills, randomized trial, advance care planning, graduate medical education, standardized patients