AI Radiology: Speed vs. Accuracy Tradeoff
AI is reading CT scans faster than ever. A 2026 systematic review found it may also be making radiologists worse at catching what AI misses.
Here is a number the radiology AI industry would rather not talk about: 4.89.
That is the pooled odds ratio from a 2026 systematic review on automation bias in radiology - published by researchers at Rowan University School of Osteopathic Medicine and covering five controlled studies across mammography, chest radiography, and MRI. An odds ratio of 4.89 means that when an AI system flagged a finding incorrectly, radiologists were nearly five times more likely to miss the correct diagnosis than they would have been without AI assistance at all.
Meanwhile, the industry is celebrating a different number: 1,524. That is the count of FDA-cleared AI algorithms in medical imaging as of March 30, 2026. Seventy-six percent of those cleared algorithms - 1,163 of them - are in radiology. The FDA is clearing roughly 30 new algorithms per month. Sixty-eight were cleared in Q1 2026 alone. The pace is extraordinary. The outcome data tracking whether this wave of tools is improving diagnosis is not.
This issue examines the real tradeoff: what radiology AI delivers, what it costs, and why most programs are measuring the wrong thing entirely.
1. The Speed Revolution Is Real
Start with what is genuinely working. AI in radiology has produced measurable, meaningful improvements in workflow efficiency. The numbers are not small.
Average imaging turnaround time dropped from 11.2 days to 2.7 days in programs that deployed triage AI - an 76% reduction. Radiologist reading workload fell by 53% in high-volume settings when AI pre-screened studies for urgency. Scan reading speeds improved 30 to 75% depending on modality and AI system. These are operational wins that matter: patients wait less, critical findings surface faster, and radiology departments can handle volume growth without proportional staffing increases.
Aidoc's January 2026 clearance for a body CT AI covering 14 conditions achieved 97% sensitivity and 98% specificity across the cleared indication set. That is genuinely impressive performance. And 54% of US hospitals with 100 or more beds now report using AI in radiology. The infrastructure is in place.
The growth curve for FDA-cleared radiology AI tells a consistent story: from roughly 160 cleared algorithms in 2020 to 1,163 by early 2026. The field has scaled faster than almost any other area of medical AI.
The speed story is real. The problem is that it has become the only story being told.
2. The Accuracy Gap Nobody Is Measuring
Only 30% of radiologists currently use AI in clinical practice on a regular basis. That gap - between hospital systems that have deployed AI and the radiologists actually integrating it into reads - is a signal worth paying attention to.
The bigger gap is in measurement. Ask a radiology department how AI has affected their turnaround time and they can tell you precisely. Ask how AI has affected their false negative rate and most cannot answer. Ask whether their miss rate changes when AI confidence is low and the data does not exist.
The chart above captures the core problem. Programs tracking turnaround time and radiologist speed: over 80%. Programs tracking miss rate by AI confidence level: roughly 14%. Programs auditing AI versus unaided error rates over time: approximately 8%. The industry sold speed. The industry is measuring speed. Nobody is measuring what happens when speed and accuracy diverge.
3. Automation Bias: The Clinical Risk Hiding in Your Worklist
Automation bias is the tendency for humans to defer to automated systems even when those systems are wrong. In radiology, it means a radiologist accepting an AI finding without adequate independent review - or, more dangerously, missing a finding that AI did not flag because the absence of a flag substitutes for independent judgment.
The 2026 Rowan systematic review identified automation bias across every modality studied. The pooled odds ratio of 4.89 holds across different AI systems, different experience levels, and different imaging contexts. This is not a one-vendor problem or a mammography-specific problem. It is a structural problem with how humans interact with AI-assisted triage systems.
The most alarming finding involves experience level. For inexperienced radiologists - residents and early-career physicians - accuracy dropped from 79.7% to 19.8% when AI provided an incorrect recommendation. That is not a modest decline in performance. That is a fourfold collapse. The odds ratio for inexperienced radiologists accepting incorrect AI recommendations was 15.57, compared to 4.89 for the pooled sample.
4. The Expertise Paradox
Here is the uncomfortable irony: automation bias is lowest among the radiologists least likely to be using AI, and highest among the radiologists most dependent on it.
Experienced radiologists show meaningfully lower rates of automation bias in the literature. They are more likely to override incorrect AI flags, more likely to notice when an AI recommendation conflicts with clinical context, and more likely to perform the independent review that the AI workflow was supposed to supplement. Their baseline pattern recognition is strong enough to catch AI errors.
Trainees and early-career radiologists are in a different position. They entered training during the deployment wave. AI is woven into their workflow from day one. Their pattern recognition is still developing - which is precisely when AI assistance looks most helpful and is most dangerous. When the AI is wrong, they do not have the independent pattern recognition to catch it.
This creates a training pipeline problem that will compound over time. If the next generation of radiologists develops their diagnostic skills in an AI-first environment without structured oversight programs, the automation bias risk embedded in a 79.7% to 19.8% accuracy drop becomes a profession-wide liability rather than an individual one.
There is also a complementarity point worth taking seriously. Research published in Nature npj in 2025 found that AI and radiologists tend to make distinct categories of errors. AI systems miss findings that humans catch easily, and humans miss findings that AI catches reliably. This means thoughtfully designed human-AI collaboration can outperform either alone - but only when the collaboration is genuinely bidirectional, not when the human is simply ratifying the AI output.
5. Who Wins and Who Loses in the Current Model
The winner-loser split in 2026 radiology AI is not about which AI system you bought. It is about what you do with the data after deployment.
High-volume systems that have built audit programs - tracking miss rates by AI confidence score, comparing AI-assisted false negative rates to unaided historical baselines - are capturing the operational benefits of speed without creating the liability exposure that comes with unmonitored automation bias. These programs know when their AI is uncertain and have workflows to flag those cases for additional review.
Programs measuring only turnaround time and cost savings are flying blind. They know their throughput improved. They do not know whether their diagnostic accuracy changed. When a missed finding surfaces months later in a malpractice claim, the absence of outcome data is not a defense.
Trainees are in the highest-risk position in the current model. Their programs are deploying AI to help them handle volume, which makes sense operationally. Their supervisory oversight for AI-assisted reads may not be calibrated to the automation bias risk specific to inexperienced readers.
6. Deep Dive: What the 2026 Systematic Review Actually Found
The Rowan University study is worth examining in detail because it is the most rigorous synthesis of automation bias research in radiology to date.
The review identified five controlled studies that met inclusion criteria: randomized or crossover designs where radiologists read the same cases with and without AI assistance, with accuracy measured in both conditions. The studies spanned mammography screening, chest radiography, and MRI - three modalities with different visual complexity, different AI tool maturity levels, and different clinical stakes.
The pooled analysis found a consistent automation bias effect across all five studies. When AI provided an incorrect recommendation, radiologists were 4.89 times more likely to produce an incorrect read than when reading without AI. This held across modalities, suggesting the bias is not a mammography-specific phenomenon driven by the complexity of that particular task.
The experience stratification finding - OR of 15.57 for inexperienced radiologists versus a pooled 4.89 - was the sharpest finding in the paper. The authors note that this may partly reflect the fact that experienced radiologists have independent pattern recognition strong enough to catch AI errors, while inexperienced radiologists lack that baseline. The implication is that AI deployment without experience-stratified oversight protocols may transfer liability from high-experience to low-experience readers rather than reducing overall diagnostic risk.
One finding the authors highlight as a gap: none of the five studies tracked AI confidence scores against miss rates. This is the metric that would matter most operationally - knowing that when your AI outputs a confidence score below a certain threshold, human miss rates spike. That data does not exist in the published literature because programs are not collecting it.
What This Means For You
If you run a hospital system with AI in radiology: build the audit infrastructure before you need it defensively. Track miss rates by AI confidence score. Compare AI-assisted false negative rates to your historical unaided baseline. Stratify your oversight by radiologist experience. The operational case for AI is already won. The question now is whether you can capture the benefits without importing the risk.
If you are a radiologist integrating AI into your workflow: the Rowan data is a reminder that the AI recommendation is a second opinion, not a first one. The cases where AI is most confidently wrong are the cases where automation bias risk is highest. Interrogate the cases where AI flags something you would have dismissed, and the cases where AI is quiet about something that feels off.
If you are a radiology trainee or resident: you are in the highest-risk group in the current evidence base. Your program's AI deployment is not designed to be a substitute for developing independent diagnostic judgment. Use AI as a checker of your independent read, not as a triage system that tells you where to look.
If you are a radiology AI vendor: the speed pitch won. The infrastructure sale is done. The next competitive differentiator is outcome data. The vendors who can show diagnostic accuracy improvement - not just workflow efficiency - will own the next phase of this market. The vendors who cannot will face growing scrutiny as the liability exposure of unmeasured automation bias becomes clearer.
If you are a payer or health system administrator: the 30% clinical adoption rate among radiologists despite 54% hospital deployment tells you something. The gap between procurement and actual use is often a signal that the tool is not integrated into the workflow in a way clinicians trust. That trust problem is worth solving before it becomes an outcome problem.
The 1,524 cleared algorithms tell you how fast this field is moving. The 4.89 odds ratio tells you what the industry has not figured out yet. Both numbers are real. The question for radiology over the next three years is which one drives policy.
Speed is the feature. Accuracy is the product. The programs that understand the difference will be positioned very differently from the programs that do not when the accountability reckoning comes.
About the Author
Jonathan Govette is the CEO of Oatmeal Health, a company focused on early lung cancer detection in underserved communities. This newsletter covers healthcare policy, imaging AI, and the structural forces shaping community health. Subscribe for weekly analysis at the intersection of technology and access.
Key References
Rowan University School of Osteopathic Medicine (2026). Automation Bias in AI-Assisted Radiology: Systematic Review and Meta-Analysis. 5 studies, pooled OR=4.89.
FDA Medical Device Database. AI/ML-Enabled Medical Devices: 1,524 cleared algorithms as of March 30, 2026. 76.31% in radiology.
Aidoc Medical (January 2026). FDA Clearance for Body CT AI Covering 14 Conditions: 97% sensitivity, 98% specificity.
Nature npj Digital Medicine (2025). Complementary Error Patterns in AI and Human Radiology Reads: Implications for Collaborative Workflow Design.
EJRAI (2026). Workflow Integration of Radiology AI: 30-75% Scan Time Reduction Across 12 Institutions.
American College of Radiology (2026). AI Adoption Survey: 54% of US Hospitals with 100+ Beds Report Radiology AI Use; 30% of Radiologists Report Daily Clinical Use.