When AI Drafts and Doctors Edit: A New Model for ICU Family Conversations Outperforms Both Alone
Japanese study finds physician-edited AI responses scored highest on empathy and medical accuracy in simulated end-of-life scenarios, suggesting a collaborative future for critical care communication.

In the emotionally charged environment of intensive care units, where families face life-and-death decisions about loved ones, the quality of physician communication can profoundly shape outcomes. Now, research from Japan suggests an unexpected solution: artificial intelligence working as a drafting partner for doctors, rather than replacing them or being sidelined entirely.
A cross-sectional study conducted at Sapporo Medical University School of Medicine found that written explanations for ICU end-of-life scenarios achieved their highest quality ratings when physicians edited AI-generated drafts. This collaborative approach outperformed both physician-only and AI-only responses across measures of medical validity and empathy, according to findings from evaluations conducted between September 1 and November 30, 2025.
The Human-AI Partnership Advantage
Forty-five evaluators—equally divided among physicians, nurses, and community members—assessed responses to 15 simulated critical care family inquiries under blinded conditions. The questions emerged from three realistic ICU vignettes: septic shock requiring VA-ECMO support, respiratory failure in a patient with malignancy, and postcardiac arrest syndrome.
Seven board-certified intensivists, each with more than 10 years of critical care experience, authored the physician-only responses and edited the AI drafts. The AI platform used was OpenAI GPT-5.0 via ChatGPT Plus interface, accessed during August 1–15, 2025, using Japanese-language prompts in Temporary Chat mode without iterative refinement.
The collaborative physician-AI texts demonstrated statistically superior performance compared to single-source approaches across all three primary endpoints (P<0.001), with particularly pronounced advantages in empathy ratings. Evaluators used five-point Likert scales to assess responses, with empathy measured using a framework based on the Japanese translation of the Consultation and Relational Empathy (CARE) measure.
Implications for Clinical Practice
The study, which received institutional review board approval (identifier 7-1-14), achieved complete data collection—90 evaluation scores per healthcare assessor and 45 empathy ratings per community evaluator. Statistical analysis employed Friedman tests with Bonferroni-adjusted significance thresholds (P<0.0167).
For healthcare providers, these findings support implementing AI-assisted drafting workflows for high-stakes ICU family communications. The approach could help standardize communication quality across clinicians with variable interpersonal skills while maintaining clinical accuracy—addressing documented gaps in formal communication training within postgraduate medical education.
For health AI developers, the research validates a human-in-the-loop model rather than autonomous generation. The methodology demonstrates that GPT-5.0 serves optimally as a drafting tool requiring physician refinement, particularly for acute care contexts demanding both technical precision and relational sensitivity.
Real-World Applicability
The study's design deliberately tested practical implementation: single-prompt AI entries without refinement, experienced intensivists as editors, and diverse evaluator perspectives including nonhealthcare community members who represent patient families.
Improved communication quality in ICU settings may enhance surrogate decision-makers' prognostic understanding, reduce emotional distress during family conferences, and better align life-sustaining treatment decisions with patient values. These outcomes address a critical vulnerability point where communication failures frequently create persistent comprehension discrepancies between medical teams and families.
The findings suggest that the future of clinical AI may lie not in replacement or isolation, but in thoughtful collaboration—where technology drafts and human expertise refines, combining the strengths of both for conversations that matter most.
Sources
Follow Health AI Central
Keep up with health AI where you already read
On LinkedIn
The day’s strongest health AI stories in your feed, one or two each weekday.
On Google
Choose Health AI Central as a preferred source, and Google is more likely to show our reporting in Top Stories.
Reporting & accountability
How we maintain accuracy
Health AI Central separates reported facts, analysis, and public-source data. Readers can review our standards, request a correction, or consult the relevant research methodology before relying on a claim.