The Curriculum Gap That Is Quietly Reshaping Radiology's Future Workforce

In 2016, Geoffrey Hinton, Nobel laureate and godfather of deep learning, stood at a Toronto seminar and said:

"People should stop training radiologists now. It's just completely obvious that within five years deep learning is going to do better than radiologists... We've got plenty of radiologists already."

That statement didn't end radiology. But it did something arguably more damaging: it embedded a narrative of displacement into the minds of a generation of future physicians, a narrative that new data confirms is still shaping career decisions today.

What the research actually shows

Yilmaz et al. (2026), published in Academic Radiology, surveyed 401 Canadian medical students and residents across all 17 Canadian medical schools, producing what is one of the most current and methodologically rigorous national assessments of how AI is shaping radiology specialty preference.

The headline finding is sobering: 32.9% of students already interested in radiology reported feeling discouraged from pursuing the specialty because of AI. Students who believed AI would diminish the necessity for radiologists had 50% lower odds of being interested in diagnostic radiology (OR=0.50, 95% CI: 0.36–0.68). A companion meta-analysis of 21 studies cited in the paper found that 31.9% of medical students globally are less likely to choose radiology due to AI-related concerns, a pattern that is widespread and persistent.

As a field, radiology is hemorrhaging potential radiologists to a fear that is largely driven by misinformation.

But here is what I find most important, and most hopeful

The paper introduces a concept the authors call "informed realism," and it re-frames everything.

AI-savvy students, those with genuine AI knowledge and formal exposure, were significantly more likely to view AI as a tool that enhances radiologist efficiency, not replaces them. They were more likely to advocate for radiologist-industry partnerships and to see AI as an amplifier of clinical capability. Students who agreed that AI would amplify radiologists' capabilities had double the odds of being objectively AI-savvy (OR=2.00).

Yet paradoxically, 45.5% of AI-savvy students still reported discouragement from pursuing radiology, compared to only 28.4% of their less knowledgeable peers (p<0.001). More knowledge, more nuance, but also more awareness of genuine workforce disruption.

This is not a failure of education. This is what honest AI literacy looks like. And it points directly to what we need to do next.

The knowledge gap is stark, and actionable

Perhaps the most striking data point in the paper: 75.4% of students who subjectively believed they were not tech-savvy thought AI would reduce radiologist demand. Among objectively AI-savvy peers? Only 18.9% (OR=13.19, 95% CI: 6.71–25.91). The perception of displacement is overwhelmingly concentrated in students who simply don't know enough about AI to evaluate it accurately.

This is a curriculum design problem, and it's entirely solvable.

The paper also found that students interested in radiology were disproportionately influenced by formal sources: local radiologists, journals, and conferences, nearly 47.2% vs. 18.7% of uninterested students.

What I believe we should do

At Luxsonic Healthcare we work at the intersection of radiology workflow and AI every single day. What we see in the industry mirrors what this paper confirms, the gap is not in the technology.

It's in how we talk about it, and when.

The paper is clear in its conclusion: "These findings support the need for early, evidence-based AI education that not only addresses misconceptions, but also encourages current healthcare trainees to embrace AI in a technology-enhanced healthcare system."

Here is what that could look like in practice:

1. Integrate AI literacy into pre-clinical medical education. Students forming specialty impressions in Years 1 and 2 should understand what clinical AI actually does, and what it cannot do, before media narratives and Hinton-era soundbites fill that vacuum.

2. Teach the "informed realist" framing. The goal is not to tell students AI is harmless. It is to give them the knowledge to evaluate it accurately. AI-savvy students were more discouraged AND more likely to choose radiology, because they saw the full picture.

3. Make expert voices accessible earlier. The data shows students trust local radiologists above all other sources. Departments and professional societies need structured programs that bring working radiologists into pre-clinical and clerkship conversations about AI, before students opt out of the specialty.

4. Address the equity risk explicitly. The paper raises something rarely discussed: "without formal AI education, the field will remain open to technology-inclined students and risks losing students who could diversify radiology with expertise in global health, sustainability, and ethics." AI illiteracy in medical education is not a neutral gap. It is a filter that selects for a particular kind of trainee and excludes others.

5. Don't confuse enthusiasm for education with readiness. The paper found that believing AI improves medical education was not associated with interest in radiology. Job security anxiety overrides abstract support for AI in curriculum. We need education that directly addresses workforce realities, not just AI capabilities.

Hinton's prediction did not come true. Since 2016, Mayo Clinic alone grew its radiology department by 55%. The American College of Radiology projects specialty supply growth of 26% over the next 30 years. We are facing a radiologist shortage, not a surplus.

But if we fail to act on what this paper is telling us, we may manufacture one.

I'd genuinely like to hear from radiology education leaders: Is AI literacy formally embedded in your curriculum? What are the barriers?

Reference: Yilmaz E, Atalla M, Gaisinsky A, et al. The Impact of Artificial Intelligence on Radiology Specialty Preferences Among Canadian Medical Students and Residents. Academic Radiology. 2026. https://doi.org/10.1016/j.acra.2026.04.039

At Luxsonic, we're committed to building AI-enabled radiology workflows that expand what radiologists can do, not reduce who they are.