In defence of post-hoc explanations in medical AI

Post-hoc explanations don't need to be perfect to be useful: The paper argues that even though post-hoc explanations don't replicate the exact reasoning of AI systems, they can still meaningfully improve how clinicians understand and work with medical AI tools—challenging the assumption that explanations must be completely accurate to have value.

Post-hoc explanations can improve team performance: The authors present evidence that post-hoc explanations increase the accuracy of clinician-AI teams, suggesting that providing some explanation to doctors leads to better combined human-AI decision-making in medical settings than using AI without explanation.

Explanations help clinicians justify decisions to patients: Beyond just understanding the AI, post-hoc explanations enable clinicians to better explain and defend AI-informed medical decisions to patients, which is important for informed consent and accountability in healthcare.

Post-hoc explanations are a practical solution, not a perfect one: The paper acknowledges that post-hoc explanations are not a complete solution to the "black box problem" in medical AI, but argues they remain a pragmatic and valuable strategy for making medical AI systems more transparent and trustworthy in real-world clinical practice.

The debate over explanation methods shouldn't delay implementation: By defending post-hoc explanations against recent criticism, the authors suggest that waiting for perfect explanation methods shouldn't prevent the adoption of explainability tools that can already provide meaningful benefits to clinicians and patients today.