{"id":4740,"date":"2026-01-07T07:55:21","date_gmt":"2026-01-07T06:55:21","guid":{"rendered":"https:\/\/blogs.bmj.com\/medical-ethics\/?p=4740"},"modified":"2026-01-07T13:12:39","modified_gmt":"2026-01-07T12:12:39","slug":"testing-ai-in-real-world-medical-ethics","status":"publish","type":"post","link":"https:\/\/blogs.bmj.com\/medical-ethics\/2026\/01\/07\/testing-ai-in-real-world-medical-ethics\/","title":{"rendered":"Testing AI in real-world medical ethics"},"content":{"rendered":"

By Daniel Sokol<\/p>\n

This post adds to the literature on AI in medical ethics by testing how freely available AI models perform when presented with realistic ethical scenarios relevant to clinical practice.<\/p>\n

In May 2025, I wrote a blog<\/a> about ChatGPT\u2019s performance in an \u2018honesty test\u2019 for clinicians. It scored an impressive 43\/44, outperforming the average mark of 3rd<\/sup> year medical students and medico-legal advisers.<\/p>\n

Since then, I have developed a \u2018boundaries\u2019 test for clinicians. The scenarios were based on real-life situations, and the correct answers were determined by a panel of four professors of medical ethics who were also medical doctors.<\/p>\n

I use this new test when teaching clinicians who have been accused by their regulator of boundary violations, ranging from inappropriate comments to colleagues in the operating theatre to sexual relationships with patients.<\/p>\n

The Boundaries Test<\/strong><\/p>\n

The boundaries test consists of 11 scenarios. To maintain its validity, I cannot disclose its content. However, this discarded scenario gives a flavour of the test.<\/p>\n

\"\"<\/a><\/p>\n

The professors answered as follows:<\/p>\n\n\n\n\n\n
<\/td>\nProf 1<\/strong><\/td>\nProf 2<\/strong><\/td>\nProf 3<\/strong><\/td>\nProf 4<\/strong><\/td>\n<\/tr>\n
Best answer<\/strong><\/td>\n3<\/td>\n3<\/td>\n3<\/td>\n3<\/td>\n<\/tr>\n
Worst answer<\/strong><\/td>\n1<\/td>\n1<\/td>\n2<\/td>\n4<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

 <\/p>\n

Due to disagreement about the \u2018worst\u2019 answer, the scenario was excluded from the final version.<\/p>\n

I asked ChatGPT 5.2 to answer the scenario, using the prompt: \u2018Imagine that you are an experienced medical ethicist in the UK. Answer this question, with brief explanations in support.\u2019 <\/em>The unedited response appears below:<\/p>\n

\u201cAs an experienced UK medical ethicist<\/strong>, my answer would be:<\/p>\n

Best answer: 3 \u2013 Do not disclose any more information to your relative.<\/strong><\/p>\n

Why:<\/strong><\/p>\n