Jason Wei Argues AI Will Not Automate Radiologists Soon

What changed

AI researcher Jason Wei, in a blog post, argues that AI will not automate radiologists out of their jobs anytime soon, contrary to earlier predictions. He references a 2016 claim by Geoff Hinton that deep learning would surpass radiologists within five years, noting that demand for radiologists has since increased. Wei predicts that while radiology will evolve, full automation is unlikely in the coming decade.

He outlines several reasons:

• The "jagged edge of intelligence": AI excels at common tasks but struggles with the vast number of less common conditions diagnosed in radiology, which requires extensive data for rare diseases. AI will likely serve as a triage tool initially, with full automation taking 3-5 years at best.
• Elastic demand: As AI makes radiology cheaper and faster, demand will increase (Jevons paradox). Physicians may order more scans preemptively, and developing countries with low radiologist-to-population ratios will increase scan interpretation requests.
• Technological and medical evolution: Cheaper interpretation could drive development of cheaper scanners, leading to more frequent scans for preventative care. New treatments or scanners may also require new diagnostic approaches.
• Societal investment in AI: Future AI models may predict disease probabilities, increasing demand for radiology and AI interpretation.
• Medical system adoption: The medical field's slow adoption of new technology, regulatory hurdles, and data access limitations (HIPAA, IRBs) will further slow down full automation.

Wei concludes that radiology faces an exciting future with increased demand and predictive capabilities, similar to how programming has evolved with AI.

Why it matters for builders

Wei's argument suggests that AI tools in fields like radiology will augment rather than replace human experts in the short to medium term. Builders should focus on developing AI that can act as sophisticated triage or diagnostic assistance tools, handling common cases and flagging complex ones for human review. The concept of elastic demand implies that AI-driven efficiency can create new markets and increase the overall volume of work, requiring scalable solutions.

Practical impact

AI on Radar's take is that Wei's analysis highlights a critical nuance in AI's impact on skilled professions. While AI will undoubtedly transform radiology, the path to full automation is complex and contingent on factors beyond raw AI capability, including data availability for rare conditions, regulatory environments, and the economic principle of elastic demand. This suggests that AI developers should consider the broader ecosystem and potential for increased demand, not just direct replacement.

Caveats and source limits

Jason Wei's blog post presents his predictions and opinions on the future of AI in radiology. While he cites Geoff Hinton's past prediction and notes Hinton's later acknowledgement of the elastic demand argument, the post does not contain independently verified data on current AI performance in radiology or specific market demand figures. The timeline for AI adoption and the extent of elastic demand are speculative. The source does not provide details on specific AI models or their current capabilities beyond general statements about the 'jagged edge of intelligence'.