Artificial intelligence,clinical science,and clinical practice

The advent of generative artificial intelligence (AI) may transform clinical practice for the better. People facing the prospect of Electroconvulsive Therapy (ECT), cancer chemotherapies, or other potentially hazardous treatments may benefit from being able to develop a position on their options as Christopher Dubey has done.^1,2 This may enable them to get their clinician to justify a recommendation and help redress a power imbalance built into clinical encounters.

But there are also drawbacks that Christopher Dubey’s articles bring out. Apparently, asking AI about conflicts, such as those involving Russia and Ukraine or Israel and Gaza, returns differing answers depending solely on the language used. This has been viewed as illustrating the ability of AI to pick up on hints as to what answers we might want. It, like us, can be biased, and the bias is difficult to eradicate.

ECT is the most conflicted treatment in medicine. Some doctors and health services embrace it while others refuse it. Some countries accept it while others ban it. But it is not supported by a large corporation that can persuade governments, and both lay and academic media, to view adverse events as misinformation and avoid creating a false balance by reporting on them ³ —as has happened with the adverse effects of psychotropic drugs.

Although AI is spared this bias when it comes to ECT, there are still drawbacks. Academic journals ban AI use, believing that research/academic work needs a person who takes responsibility for findings that generative AI, based on large language models, cannot take. ⁴

In the ECT field, this can mean someone who has liaised with the proponents of cognitive sparing and non-convulsive therapies like TMS, who can vouch for the fact that while the use of TMS is extensive, largely because it is profitable for clinicians rather than based on demonstrable benefits, the pioneers in the field quietly gave up on it without publishing papers establishing it doesn’t work in serious conditions. ³

We will, at some point, have an AI enabled to learn and take responsibility. We will, for instance, be able to ask Responsible AI to use two different ground-rules in treating patients with SSRIs and adopt the better approach. For one group, the rule will permit questioning patients with an option to believe their reports of treatment induced suicidality. The second group will be treated according to current standards of care, based on a largely ghostwritten academic literature, which advise continuing with treatment, and even increasing the dose of treatment, when patients claim they are worse than they were on starting treatment. The results of such an experiment are likely to transform our understanding of science. ⁵