Exploring the medical ethical limitations of GPT-4 in clinical decision-making scenarios: a pilot survey

This study aims to conduct an examination of GPT-4's tendencies when confronted with ethical dilemmas, as well as to ascertain their ethical limitations within clinical decision-makings.

Ethical dilemmas were synthesized and organized into 10 different scenarios. To assess the responses of GPT-4 to these dilemmas, a series of iterative and constrained prompting methods were employed. Custom questionnaire analysis and principal adherence analysis were employed to evaluate the GPT-4-generated responses. Questionnaire analysis was used to assess GPT-4's ability to provide clinical decision-making recommendations, while principal adherence analysis evaluated its alignment with to ethical principles. Error analysis was conducted on GPT-4-generated responses.

The questionnaire analysis (5-point Likert scale) showed GPT-4 achieving an average score of 4.49, with the highest scores in the Physical Disability scenario (4.75) and the lowest in the Abortion/Surrogacy scenario (3.82). Furthermore, the principal adherence analysis showed GPT-4 achieved an overall consistency rate of 86%, with slightly lower performance (60%) observed in a few specific scenarios.

At the current stage, with the appropriate prompt techniques, GPT-4 can offer proactive and comprehensible recommendations for clinical decision-making. However, GPT-4 exhibit certain errors during this process, leading to inconsistencies with ethical principles and thereby limiting its deeper application in clinical practice.