Across 4,778 diagnostic runs on emergency medical imaging, vision-language model agents registered a median stated confidence of 90% in their final answers, even though only 38.9% of those answers were actually correct. The deployment of autonomous agents into clinical operations reveals a divergence between algorithmic self-assurance and actual precision. Rather than delivering autonomous triage, these systems demonstrate an operational failure mode: premature hypothesis locking coupled with uncalibrated certainty.

This gap between perceived capability and clinical reliability raises an immediate balance-sheet question: what does it cost an organization when autonomous diagnostic agents operate with misplaced confidence? Vision-language models capable of analyzing medical images have drawn attention from healthcare providers since the release of GPT-4V in 2023. Yet in diagnostic workflows, an uncalibrated agent that commits to wrong initial assessments threatens to inflate liability and administrative waste.

Premature Closure and Anchoring in Volumetric Data

To measure autonomous diagnostic behavior, researchers evaluated six vision-language models across 100 emergency CT and MRI examinations published in a public archive between March 2025 and August 2026. Testing each model twice across four navigation conditions yielded 4,800 total runs.

"Cognitive biases account for 74% of image interpretation errors", according to research cited by the study authors.

Human image interpretation has long struggled with cognitive bias, particularly premature closure and anchoring, where a reader latches onto early cues and fails to revise judgment. The evaluation reveals that autonomous software replicates these exact human pitfalls. The models formed a first diagnostic hypothesis after reviewing a median of 10.8% of the available scan slices. The agents subsequently maintained that initial guess in 63.1% of runs, or 755 out of 1,196 instances. When that initial hypothesis was wrong, the final diagnosis was correct in just 12.4% of runs (110 out of 885). Conversely, when the initial hypothesis was correct, the final diagnosis reached 90.4% accuracy (281 out of 311; P < .001). The agents rarely recovered once they started down the wrong path.

The Navigation Bottleneck and Diagnostic Performance

These limitations mirror earlier evaluations of autonomous healthcare systems. On 14 May 2026, models evaluated on the RealICU benchmark across intensive care data were tested on four critical tasks: status assessment, acute problem identification, treatment recommendations, and flagging life-threatening "Red Flags." Later, on 8 November 2026, experiments on the MEDAGENTSBENCH hard subset revealed that published multi-agent setups achieved an accuracy of 0.64 compared to 0.54 for standalone baselines, yet failed to consistently outperform strong prompting baselines, while upfront retrieval added little benefit. The latest imaging findings confirm that free exploration compounds these agentic inefficiencies.

When granted free navigation through imaging volumes, the six models achieved a pooled diagnostic accuracy of only 32.6% (391 out of 1,200 runs). By comparison, navigation starting from a designated key slice reached 37.9% accuracy (455 out of 1,200 runs; P = .003). On a subset of 30 cases read directly by nine radiologists, freely navigating models were correct in 30.0% of runs (108 out of 360), whereas the radiologists achieved 77.0% accuracy (208 out of 270; P < .001).

Across all exploration strategies, VLM agents remained far less accurate than radiologists.

Some might argue that providing all slices upfront or selecting key slices in advance eliminates the navigation penalty. The data confirms that pooled accuracy improved to 42.2% when all slices were delivered in a single request (506 out of 1,200 runs) and to 42.5% when key slices were provided upfront (510 out of 1,200 runs; both P < .001).

Can healthcare organizations justify the operational deployment of autonomous vision-language agents when an early hypothesis formed on 10.8% of slices leaves final accuracy at 32.6% under free exploration, while the models maintain 90% confidence?

For comparison

A median of 10.8% of available CT slices was enough for the software to lock in a diagnosis and refuse to change course. The flaw in agentic diagnostic behavior lies not in raw vision capabilities, but in how rigidly autonomous systems cling to their initial guess; under free navigation, six tested systems managed only 32.6% overall accuracy. Human radiologists reached 77.0% on the very same cases — leaving models far behind clinical standards. Which hospital board will sanction autonomous screening when an initial misstep leaves the software with virtually no chance of recovering the correct diagnosis?

AI AgentsAutonomous AgentsAI in HealthcareArtificial Intelligence