Demand for human radiologists is at an all-time high

AI capabilities vs real‑world performance

• Commenters repeatedly note a gap between AI’s benchmark performance and hospital use: models degrade on out‑of‑sample data, struggle with rare conditions, and can latch onto spurious cues (e.g., hospital‑specific artifacts) rather than pathology.
• Radiology AIs are generally narrow: good at a few common findings or specific tasks (e.g., triage, certain cancers), not at full-case reasoning.
• Several radiologists and clinicians say current tools are helpful but “nowhere close” to replacing human interpretation, especially on edge cases and complex cross‑sectional imaging.

What radiologists actually do

• There is disagreement about how much time radiologists spend talking to patients and other clinicians.
  • Some report teleradiology setups where rads only read images and dictate reports.
  • Others (especially hospital‑based and interventional radiologists) describe frequent consultations, procedure work, and collaborative planning.
• One theme: the job is not just “spot the nodule” but building a mental model of anatomy, integrating clinical context, and handling ambiguous or conflicting data.

Liability, regulation, and incentives

• Legal risk is seen as the dominant barrier to full automation. Malpractice systems are built around a licensed human who can be sued; vendors avoid being that “throat to choke.”
• Even if AI outperforms humans statistically, insurers and regulators currently insist on a human sign‑off; malpractice policies often exclude AI‑only workflows.
• Some expect earlier fully automated deployment in low‑resource settings or narrow, tightly validated niches (e.g., diabetic retinopathy).

Augmentation vs replacement

• Widely shared view: AI will mostly augment radiologists—triaging worklists, drafting reports, flagging rare conditions, and modestly reducing reading time.
• A minority argue that if AI becomes clearly superior and law changes, hospitals will eventually skip human readers for many studies; others counter that existential stakes and patient expectations will preserve a human role.

Workforce, training, and hype

• Demand for radiologists is very high; groups report difficulty hiring despite generous offers.
• Some suggest vacancies are partly because trainees fear long‑term automation risk and choose other specialties.
• Hinton’s 2016 “stop training radiologists” line is debated as emblematic of over‑optimistic AI timelines, likened to self‑driving car hype: impressive demos, but thousands of dangerous edge cases slow real deployment.