There’s a radiologist I know at a major Sydney hospital. Their department installed an AI system last year that reviews chest X-rays and flags potential abnormalities. The AI is pretty good—peer-reviewed studies show it catches things human readers miss.

She told me she ignores it about 60% of the time.

Not because it’s usually wrong. It’s not. She ignores it because it doesn’t fit her workflow, the explanations aren’t helpful, and she doesn’t trust it enough to change her clinical judgment based on what it says.

This is the AI adoption problem in healthcare that nobody wants to talk about. We’ve built increasingly sophisticated clinical AI systems. And clinicians are politely declining to use them.

The Trust Problem

Let’s start with the fundamental issue: trust.

Clinicians are trained to understand the reasoning behind their decisions. They learn pathophysiology, diagnostic criteria, evidence-based guidelines. When they make a clinical call, they can explain why.

AI systems—particularly deep learning models—often can’t. They produce a recommendation or a probability score without explaining the underlying reasoning in clinically meaningful terms. “This has an 87% probability of being pneumonia” doesn’t tell a doctor which features of the image led to that conclusion.

This is called the black box problem, and it’s a massive barrier to adoption. If you don’t know why the AI reached a conclusion, how do you evaluate whether to trust it?

Clinicians don’t trust blindly. Nor should they. Medical training emphasises critical thinking and scepticism. “Trust the algorithm because it’s usually right” fundamentally conflicts with how doctors are taught to practice medicine.

The Integration Failure

Even when clinicians are willing to consider AI recommendations, the systems often aren’t integrated in a way that’s actually usable.

An AI alert that requires leaving the EMR, opening a separate application, logging in again, and finding the relevant patient is probably not going to get used. Yet that’s how many early clinical AI systems were implemented—as standalone tools rather than embedded workflows.

Radiologists reading dozens of scans per hour can’t afford a 30-second context switch for each one. Emergency physicians managing multiple patients simultaneously don’t have time to check a separate AI dashboard. If it’s not seamlessly integrated into existing workflow, it’s friction, not assistance.

The systems that are actually getting used tend to be the ones that present recommendations directly within the tools clinicians are already using. An AI alert that appears in the radiology PACS software as the scan is being reviewed? Much higher adoption than one requiring a separate system.

Alert Fatigue Is Real

Healthcare technology has a history of crying wolf. EMR systems generate alerts constantly—drug interaction warnings, duplicate order notifications, documentation reminders. Clinicians learn to ignore them because most aren’t clinically significant.

Now we’re adding AI-generated alerts to that noise. If the AI flags something concerning on 30% of chest X-rays, but 90% of those flags are false positives or findings the clinician had already identified, people stop paying attention.

The threshold problem is hard to solve. Set the sensitivity too high and you catch everything but overwhelm clinicians with false positives. Set it too low and you miss the rare-but-critical cases that are the whole point of the system.

Some systems are trying to learn individual clinician preferences—adjusting alert thresholds based on which recommendations a specific doctor tends to act on. That’s promising, but it requires sophisticated personalisation that most current systems don’t have.

The Liability Grey Zone

Here’s something clinicians worry about: If the AI recommends something and I ignore it, am I liable if it turns out the AI was right?

Conversely, if I follow the AI’s recommendation and it’s wrong, am I liable for not exercising independent clinical judgment?

These questions don’t have clear answers yet. The legal framework around clinical AI is still evolving. Medical indemnity insurers are still figuring out how to handle this. Until there’s clarity, that uncertainty makes clinicians cautious about relying on AI recommendations.

Some clinicians treat AI as a second opinion—useful input, but not determinative. Others view it as a safety net for catching rare oversights but not something to actively incorporate into routine decision-making. Very few treat it as authoritative guidance.

What Actually Works

Despite these barriers, some clinical AI implementations are genuinely succeeding. What do they have in common?

Excellent integration. They work within existing systems, not as separate tools. Recommendations appear at the point of decision-making without requiring extra steps.

Explainability. They provide clinically meaningful reasoning, not just a probability score. “Increased opacity in right lower lobe with air bronchograms consistent with consolidation” is more useful than “82% probability of pneumonia.”

High precision, even if it costs recall. It’s better to flag fewer cases but be right when you do flag something than to catch everything but drown clinicians in false positives. Specificity matters more than sensitivity for most clinical AI applications.

Collaborative design. The systems that work were built with clinicians involved from the start, not developed in isolation and dropped into clinical environments expecting adoption.

Performance transparency. Clinicians can see how the AI performs—false positive rate, false negative rate, performance on different patient populations. Treating the system as a black box breeds suspicion. Transparency builds trust.

The Cultural Shift Required

There’s also a cultural component that’s easy to underestimate. Medicine is a profession built on expertise and autonomy. Introducing AI recommendations can feel like external second-guessing of clinical judgment.

Framing matters. AI positioned as a tool to help clinicians do their job better gets a different reception than AI positioned as quality assurance checking their work.

The most successful implementations I’ve seen involve clinicians championing the technology—peers advocating for it based on their own experience, not administrators mandating it from above.

Training is also critical. Not just “here’s how to use the system,” but “here’s the evidence for how it improves outcomes, here’s how it performs, here’s when you should trust it and when you should be sceptical.” Treating clinicians as informed partners rather than end-users who just need to follow recommendations makes a huge difference.

What Needs to Change

If we want clinical AI to actually improve care rather than sitting unused, a few things need to happen.

Better explainability research. We need AI systems that can articulate their reasoning in clinically meaningful ways. This is an active research area, but it needs more investment and urgency.

Workflow-integrated design. Stop building standalone AI applications. Build AI capabilities into the tools clinicians already use daily.

Regulatory clarity on liability. Clinicians need to know what’s expected of them regarding AI recommendations. Clear guidelines from medical boards and indemnity providers would help.

Evidence-based implementation standards. Not every AI tool with decent accuracy should be deployed clinically. We need rigorous evaluation of real-world performance and clinical impact, not just algorithm performance on test datasets.

Honest conversations about limitations. AI vendors need to be clearer about what their systems can’t do, not just what they can. Overselling creates backlash when reality doesn’t match promises.

The Path Forward

Clinical AI has genuine potential to improve patient outcomes. Systems that help catch diagnostic misses, flag deteriorating patients earlier, or identify high-risk cases that need closer attention can save lives.

But realising that potential requires solving the adoption problem. Building technically sophisticated AI isn’t enough if clinicians don’t trust it or can’t use it effectively.

The solutions aren’t purely technical. They’re sociotechnical—requiring changes in how we design systems, how we integrate them into clinical practice, how we train clinicians to work with them, and how we build trust through transparency and evidence.

We’re getting there. Slowly. The systems being deployed in 2026 are better than the ones from 2024. But there’s still a gap between what AI can theoretically do and what clinicians are actually willing to let it do in practice.

Closing that gap is the real challenge. And it’s as much about understanding clinical culture and workflow as it is about improving algorithms.