Better Algorithms, Busier Radiologists

Radiology’s abundant AI breakthroughs have not displaced radiologists because models struggle outside benchmarks, face regulatory and insurance hurdles to autonomy, and replace only a slice of a radiologist’s job. Past experience with mammography CAD shows that clinical performance and human behavior can negate lab gains, and malpractice risk further slows full automation. As efficiency improves, demand for imaging grows, so AI is increasing radiologists’ workloads rather than eliminating their roles.

Key Points

• Benchmark success has not translated to clinical reliability: most radiology AIs are single-task, trained on narrow datasets, and suffer significant out-of-distribution performance drops, especially on messy, real-world images and underrepresented populations.
• History warns about automation bias and clinical gaps: mammography CAD increased callbacks and biopsies without improving detection, while double reading by two clinicians performed better.
• Regulatory and insurance barriers deter autonomy: FDA sets a high bar for autonomous tools to self-detect out-of-scope cases, retraining triggers new approvals, and malpractice carriers often exclude AI-only diagnoses.
• AI displaces only part of radiologists’ work: image interpretation is a minority of their tasks, with substantial time spent on protocoling, supervision, communication, and teaching.
• Efficiency boosts increase demand: like prior digitization, better, faster AI is likely to expand imaging volume and create new use cases rather than reduce radiologist employment.

Sentiment

Hacker News largely agrees with the article's thesis. The most upvoted and substantive comments come from practicing radiologists who validate the claims about the gap between benchmark performance and clinical reality, the liability barriers, and the breadth of radiologist work beyond image reading. While a subset of techno-optimist commenters argue that replacement is inevitable but delayed, the dominant sentiment treats current AI capabilities in radiology with measured skepticism and acknowledges the deep institutional, legal, and practical barriers to autonomous AI diagnosis.

In Agreement

• Practicing radiologists confirm that AI products available today are non-contributory and nowhere near replacing human interpretation, with one noting that diagnosis requires genuine reasoning and synthesis beyond pattern matching
• Liability and malpractice barriers are real and durable — malpractice insurers explicitly exclude AI, and the correlated failure risk creates class action exposure fundamentally different from individual physician errors
• The self-driving car analogy actually supports the article's thesis, as even Waymo operates only in constrained environments with human backup after decades of development
• Automation bias — humans uncritically accepting AI output — can make human-plus-AI performance worse than humans alone, as the mammography CAD experience demonstrated
• Radiologists spend most of their time on non-diagnostic tasks AI cannot automate: consulting with clinicians, communicating with patients, protocoling studies, and navigating political dimensions of report interpretation
• The Jevons paradox is already at work — AI making imaging faster or cheaper increases demand for imaging rather than reducing the need for radiologists

Opposed

• Current medical AI products use outdated architectures from the mid-2010s; purpose-built transformer models could represent an enormous capability leap that hasn't yet been attempted
• AI has demonstrated superhuman capabilities in specific domains like mammography screening and can detect patterns such as patient race from chest X-rays that human radiologists cannot
• The liability barrier is a regulatory and legal choice rather than a fundamental constraint — laws can change, especially under cost pressure from Medicare and Medicaid
• Rising radiologist salaries create stronger economic incentives to find AI-based alternatives, similar to how rising teamster wages accelerated adoption of automobiles
• The last-mile problem in AI is being progressively solved in driving and could follow a similar trajectory in radiology