What automation bias is — and why it's the AI risk that gets less airtime
Most coverage of AI in radiology asks how accurate the algorithm is. Automation bias asks a harder question: what happens to the human once the algorithm is in the room? It's the tendency for a radiologist using AI to over-rely on it — accepting the AI's read and scrutinizing the images less than they would unassisted. It's a recognized drawback of AI in medicine, but until recently it was thinly studied.
A paper published in RSNA's flagship journal Radiology puts numbers to it. Titled "Automation Bias in Action," the study used eye-tracking cameras to watch what radiologists actually looked at while reading screening mammograms, with and without an AI tool's prompts.
The study: eye-tracking 10 readers, with and without AI
A team led by Yan Chen, PhD, at the University of Nottingham had 10 breast radiologists interpret two-view screening mammograms in two rounds, six weeks apart. In round one they read unaided; in round two they read the same kind of cases with a commercially available AI tool as decision support, its prompts overlaid on the images. Eye-tracking cameras recorded where and how long each reader's gaze settled. The test set of 60 cases was deliberately seeded with a mix of AI suggestions: 26 true-positive, 14 false-negative, 14 false-positive and 6 true-negative.
The headline finding is the one that should give any imaging leader pause. When the AI produced a false-negative suggestion — i.e. it missed a cancer — the readers' median sensitivity for those cases fell from 71% unaided to 39% with AI (p = 0.002), a 32-percentage-point drop, as reported by AuntMinnie. The eye-tracking data explained why: on those missed-cancer cases, readers fixated less often (0.44 vs 0.47 fixations per second; p = 0.03) than when reading unassisted. The AI's silence was, in effect, telling them not to look.
The picture wasn't uniformly bad. When the AI raised a false-positive — flagged something that wasn't there — readers largely dismissed it: specificity on those cases actually rose (39% vs 21%; p = 0.004). In other words, readers were better at ignoring the AI's false alarms than at catching what it silently missed. That asymmetry is the crux of the automation-bias problem.
A greater number of visible prompts also lengthened reading: median read time rose from 25 seconds with no prompts to 34 seconds with four or more (p = 0.001) — a reminder that safeguards and efficiency pull against each other. The authors framed the takeaway for vendors bluntly: prioritize algorithms and recall thresholds that minimize false-negative suggestions, "as human readers may dismiss false-positive suggestions."
One caveat worth stating plainly: this was a small pilot — 10 readers, mammography, a controlled test set — and the authors say the results "should be interpreted with consideration" and used to design larger studies. It is a warning signal, not a settled effect size. But the direction is consistent with what other work on AI over-reliance has found, and the mechanism is general enough to matter for any modality, including CT.
Experience helps — but it isn't a plan
In an accompanying editorial, Paola Clauser, MD, PhD, of the Medical University of Vienna, cautioned that the efficiency benefits of AI in reporting have to be weighed against the risk of automation bias. She noted that current evidence suggests less experienced readers are more prone to it, while more experienced readers are better at spotting wrong suggestions — and urged particular care in preparing residents and young radiologists.
A second editorial, from Nooshin Abbasi, MD, and Catherine Giess, MD, of Brigham and Women's Hospital and Harvard Medical School, argued that evaluating AI in imaging must evolve "from algorithms alone to complete diagnostic systems." That is the operative phrase for buyers: the thing you are purchasing is not a model, it is a human-plus-model system — and the system is where the risk lives.
The honest implication for AI reporting workflows
It would be easy — and wrong — to read this study as "so keep a human in the loop and you're safe." The uncomfortable point is that the humans in this study were in the loop, and automation bias affected them anyway. Any workflow where a person reviews AI output, including draft-then-sign reporting, inherits some exposure to over-reliance. The question is not whether a workflow eliminates automation bias — none does — but whether it is designed to blunt it.
Two design choices matter most. First, AI must never produce a final report autonomously. No AI system today is approved for autonomous final reporting without a radiologist, and the whole point of automation bias is that a rubber-stamp step is barely better than no step. Second, the review should be substantive and, ideally, layered — more than one qualified set of eyes, with the final reader independently accountable for what they sign.
That is the model xAID is built on. The AI produces a structured CT report draft; xAID's in-house radiologist reviews every preliminary; and the report is delivered ready-to-sign, so the client's reading radiologist signs the final. Two independent radiologists sit between the algorithm and the patient's chart — the opposite of autonomous sign-off. It doesn't make anyone immune to automation bias, but it removes the single-point-of-failure the study warns about.
How to evaluate a vendor's safeguards against automation bias
If you're assessing an AI reporting vendor, automation bias belongs on the due-diligence checklist alongside accuracy and turnaround. Use these questions as a starting point:
| Ask the vendor | What a strong answer looks like |
|---|---|
| Does the AI ever issue a final report on its own? | No. A qualified radiologist is always between the AI and the final report; the AI output is a draft, never a signed deliverable. |
| Who reviews the AI's output, and how independent is that review? | A named, qualified radiologist reviews every case; the final signer is independently accountable, not rubber-stamping a pre-filled draft. |
| How does the workflow surface disagreement between AI and human? | Discordance is flagged and resolved rather than silently overwritten; the human can add findings the AI omitted. |
| How is performance monitored over time? | Ongoing QA tracks accuracy and reader behavior, so deterioration — the study's core warning — is caught early. |
| Is the AI tuned to minimize false negatives? | Thresholds are set to avoid silent misses, which the study shows humans are least able to catch, without flooding readers with alerts. |
None of this is exotic. It's the difference between buying an algorithm and buying a diagnostic system with the human factors engineered in. For a broader framework, the AI radiology reporting buyer's guide and this breakdown of AI reporting accuracy cover the ground — accuracy and oversight are two halves of the same purchase.
Frequently asked questions
What is automation bias in radiology?
Automation bias is the tendency for a radiologist using an AI tool to over-rely on it — accepting the AI's interpretation and scrutinizing the images less than they would unassisted. It is a recognized risk of AI in medicine, and it matters most when the AI is wrong, because the reader may accept an incorrect suggestion or overlook a finding the AI missed.
What did the 2026 RSNA Radiology mammography study find about AI and automation bias?
In an eye-tracking study of 10 breast radiologists reading screening mammograms with and without a commercial AI tool, median reader sensitivity for cancer cases the AI missed (false-negative suggestions) fell from 71% without AI to 39% with AI. Readers also fixated less and for shorter durations on those missed-cancer cases, consistent with automation bias.
Are experienced radiologists less prone to automation bias?
An editorial accompanying the study noted that current evidence suggests less experienced readers are more prone to automation bias, while more experienced readers seem more capable of identifying wrong AI suggestions. The editorial called for particular care in preparing residents and young radiologists for the risk.
How should buyers evaluate an AI radiology vendor's safeguards against automation bias?
Ask whether the AI ever produces a final report autonomously (it should not), whether a qualified radiologist independently reviews AI output, how the workflow flags disagreement between the AI and the human reader, and what quality-assurance and monitoring processes track performance over time. Automation bias is a reason to require layered human oversight, not autonomous sign-off.
Source: Taib AG, Chen Y, et al. "Automation Bias in Action: Eye Tracking of Humans Reading Screening Mammograms with and without AI Prompts," Radiology (2026), DOI 10.1148/radiol.252590; accompanying editorials by Clauser and Abbasi & Giess. Coverage via AuntMinnie and Radiology Business. Figures rounded as reported.