The Interpretability Trade-off: Ethical Implications of Deploying Deep Learning Models in Diagnostic Healthcare

This paper critically examines the ethical paradox inherent in deploying deep learning diagnostic systems in healthcare settings. While these "black box" algorithms deliver superior diagnostic accuracy across multiple clinical domains, current deployment practices systematically violate the foundational principle of informed consent through two largely unexamined mechanisms that existing AI ethics literature has consistently overlooked. Primary Contributions: Primary Contributions: The first overlooked issue is pervasive non-disclosure of AI involvement to patients. Despite widespread deployment of AI-assisted diagnostic tools, most jurisdictions impose no legal requirement to inform patients when artificial intelligence contributes materially to their medical diagnoses or treatment decisions (Petrie-Flom Center, 2025; Duffourc, 2025). This transforms informed consent from a meaningful ethical safeguard into what scholars describe as a "symbolic" checkbox ritual (Petrie-Flom Center, 2025). Empirical case studies document this failure: Harvard's Edmond & Lily Safra Center for Ethics (2025) analyzed Epic's MyChart AI-generated patient communications system and found that approximately 6% of AI communications contained factual hallucinations, 7% of AI advice posed risk of severe patient harm if followed, and critically, most hospital organizations systematically failed to disclose that communications were AI-generated. The second overlooked issue concerns structural misalignment of responsibility and capability, creating what philosophers term "responsibility gaps" (Lang, 2023). Current practice holds clinicians fully legally and professionally accountable for AI-assisted diagnostic decisions, yet clinicians lack the understanding, control, or anticipatory capability to preempt, explain, or correct algorithmic errors (Lang, 2023; Frimpong, 2025). This misalignment represents what scholars describe as "forced supererogation"-an impossible moral position where clinicians must either use more accurate AI and accept responsibility for unpredictable errors they cannot control, or refuse AI and accept professional blame for suboptimal care (Lang, 2023).