Two black boxes
Pair human and machine so each covers the other's blind spots.
The penultimate chapter sets human and AI decision-making side by side across generalisability, scale, fatigue, bias, interpretability and accountability, and argues the prize is a well-designed team rather than replacement. It cools the hype, noting that most tools sit at the level of data and information while the clinician supplies judgement. Bias runs both ways: models can inherit human prejudice, as with atypical cardiac presentations in women, yet rethinking how knee arthritis is graded narrowed disparities for underserved patients. The authors debunk common myths about interpretability and show that explanations often fail to stop clinicians following bad AI advice, a form of automation bias. They also weigh liability, trust and the self-driving car analogy.
Three takeaways
- AI scales and stays consistent but generalises poorly across tasks
- Explanations do not reliably stop clinicians following unsafe advice
- The order of AI input, before or after the clinician, changes accuracy
Reflection
Questions for this chapter
Prompts to consider on your own, in a journal club or in a teaching session. All questions are optional, so you can read them without submitting anything. If you choose to submit answers, they are collected through Google Forms.