Several reasons
A recent meta-analysis examined when it is useful to combine human intelligence with AI (1). Following a systematic review of 106 experimental studies, the authors concluded that human–AI systems perform worse on average than either humans or AI on their own. This is particularly true for tasks where AI performs better than humans. However, synergies were observed where humans initially outperformed machines.
There may be several reasons why combining humans and AI does not always produce better results. Doctors may either trust AI too much and overlook obvious errors, or not trust it enough and ignore useful suggestions. AI tools that are not seamlessly integrated into clinical practice may disrupt workflow and distract clinicians rather than assist them. AI is often better than humans at routine and structured tasks, such as image recognition and systematic categorisation, whereas humans are better at contextual assessment and communication. When humans and machines try to outperform each other, conflicts may arise, preventing effective collaboration. Consequently, their combined performance may be worse than, for example, AI on its own .
Doctors may either trust AI too much and overlook obvious errors, or not trust it enough and ignore useful suggestions
Synergies are not a given (1); they require clearly defined, complementary roles for AI and humans, and tools that are adapted to everyday clinical practice.
One study shows that synergy is possible (2). A network of primary care clinics in Nairobi used a clinical decision support tool that served as a 'safety net' for doctors. The tool identifies potential documentation errors and clinical decision-making errors but is only activated when necessary. It is thus intended to support, rather than replace, the doctor's expertise.
The results are promising. There were 16 % fewer diagnostic errors and 13 % fewer treatment errors among doctors who used the tool compared with those who did not have access to it. In absolute terms, the tool could have averted diagnostic errors in 22,000 consultations and treatment errors in 29,000 consultations annually at these clinics alone. All doctors reported that the tool improved the quality of care, and 75 % described the effect as 'substantial'.
The Nairobi study was different because AI was evaluated under real, dynamic clinical conditions. Furthermore, the tool was not an 'extra layer' but an integral part of the electronic health record. It provided suggestions for the doctor, who always had ultimate responsibility. Any usability issues were addressed on an ongoing basis by the programmers.