The Medical Moravec's Paradox: Why Vision-Language Models Fail at Clinical Triage

Introduction

Modern Vision-Language Models (VLMs) can generate radiology reports that appear clinically coherent, answer complex diagnostic questions, and engage in multi-step reasoning over medical images. Yet these same systems frequently fail to notice when a chest X-ray has been swapped for a brain MRI, when an image is flipped upside down, or when a panel contains obvious modality mismatches. This discrepancy between high-level linguistic fluency and basic visual competence constitutes what Khan et al. term the Medical Moravec's Paradox in their paper MedObvious: Exposing the Medical Moravec's Paradox in VLMs via Clinical Triage.

The concept extends Moravec's original observation that machines often struggle with perceptual and motor tasks that humans find effortless, even while mastering complex cognitive challenges. In medical imaging, this paradox carries severe consequences. Clinical interpretation does not begin with diagnosis; it begins with pre-diagnostic triage. Radiologists first verify that the input is valid: correct modality, appropriate anatomy, proper orientation, and basic image integrity. Only after these sanity checks do they proceed to pathological interpretation. Current VLMs, however, lack reliable input validation capabilities. They generate plausible diagnostic narratives on invalid inputs, creating a dangerous illusion of competence that could propagate errors through clinical workflows and AI agentic systems.

The Benchmark Design: Isolating Set-Level Consistency

The MedObvious benchmark represents a methodological advancement in medical AI evaluation by treating input validation as a distinct, safety-critical capability. Rather than assessing final diagnostic accuracy, MedObvious asks a preceding question: Is the input coherent and appropriate to interpret?

The benchmark comprises 1,880 tasks organized into five progressive tiers (T1 through T5) of increasing clinical specificity and difficulty. T1 addresses basic orientation and modality mismatches, such as detecting flipped images or CT scans masquerading as ultrasounds. T2 through T4 progress to anatomy/viewpoint verification and clinically motivated inconsistencies, while T5 focuses on high-saliency triage failures requiring detection of gross pathology or critical mismatches. This tiered structure allows precise localization of capability boundaries.

MedObvious employs a multi-panel grid format (2×2 or 3×3) that mirrors real clinical workflows involving multi-view ultrasound, multi-slice CT/MRI series, or multi-panel layouts in viewers such as 3D Slicer or ITK-SNAP. Models must perform set-level consistency checking, identifying whether any panel violates expected coherence or confirming that all panels are consistent. Crucially, the benchmark includes explicit negative controls, where all panels are valid and the correct response is that no outlier exists. This design directly measures false alarm rates, distinguishing between random guessing and genuine visual grounding.

The evaluation spans two axes. The Clinical Safety axis tests real failure modes that should trigger rejection before diagnosis, including wrong body parts, orientation errors, and viewpoint/anatomy mismatches. The Visual Grounding axis employs synthetic inconsistencies, such as inserting modality-incompatible textures into images, to test whether models actually examine visual inputs rather than relying on language priors. Additionally, MedObvious incorporates five evaluation formats, ranging from visual referring questions to multiple-choice and open-ended variants, enabling robustness testing across interface designs.

Empirical Findings: The Fragility of Visual Sanity Checking

The empirical results from evaluating 17 VLMs, including 7 general-purpose models, 4 medical adaptations, and 6 proprietary systems, reveal systematic deficiencies in pre-diagnostic verification. The best performing model achieved only 63.2% mean accuracy, indicating that sanity checking remains largely unsolved even among current advanced systems.

Several specific failure modes emerge as particularly concerning. First, models exhibit hallucination on negative controls, falsely claiming anomalies exist in clean, consistent inputs. This propensity for false alarms suggests that VLMs may lack reliable confidence calibration for visual consistency tasks, potentially leading to unnecessary clinical workflow interruptions or alert fatigue if deployed in screening contexts.

Second, performance degrades significantly when scaling from 2×2 to 3×3 grids, indicating that set-level consistency checking does not scale gracefully with input complexity. This has direct implications for AI agents operating over full imaging studies, which may need to process dozens of slices or series simultaneously.

Third, the benchmark exposes substantial format sensitivity. The same underlying visual capability yields dramatically different accuracy scores depending on whether the task is presented as multiple-choice or open-ended questioning. This volatility suggests that measured performance may reflect prompt engineering artifacts or linguistic biases rather than robust visual understanding. For instance, some models correctly identify outliers in visual referring tasks but fail when the same images are presented with multiple-choice options, or vice versa.

Notably, medical-specific VLMs did not demonstrate consistent superiority over general-purpose models on these pre-diagnostic tasks. This finding suggests that domain adaptation through fine-tuning on medical text and images may enhance diagnostic knowledge without necessarily improving the fundamental visual grounding required for input validation.

Your Take: The Architecture of Blindness

These results illuminate a fundamental architectural bias in current VLMs toward language-centric reasoning. The training objectives that produce fluent diagnostic text appear to decouple from the perceptual mechanisms necessary for basic visual verification. When a model generates a detailed report on a mismatched modality, it reveals that the system treats visual input as a trigger for language generation rather than as a constraint upon it.

This disconnect stems partly from benchmark design incentives. The medical AI community has prioritized end-task accuracy, VQA performance, and report generation metrics while assuming that input validation is solved or trivial. MedObvious demonstrates that this assumption is unsafe. In multi-image or agentic workflows, where VLMs navigate imaging software, select series, or trigger downstream analytical tools, the ability to detect obvious panel-level inconsistencies becomes safety-critical. An agent that cannot recognize it is viewing a corrupted slice or wrong anatomy may trigger inappropriate interventions or generate misleading reports.

The clinical workflow implications are significant. Radiologists perform pre-diagnostic triage rapidly and automatically, filtering out technical errors before conscious diagnostic reasoning begins. Current VLMs lack this cognitive architecture. They proceed directly to narrative generation, effectively bypassing the quality control layer that protects clinical decision-making from invalid inputs.

Furthermore, the format sensitivity observed in MedObvious suggests that these systems may be exploiting linguistic shortcuts or positional biases rather than developing robust visual representations. When performance swings wildly between multiple-choice and open-ended variants of identical visual tasks, the model is likely relying on textual priors or answer distribution statistics rather than genuine visual analysis.

Conclusion

The Medical Moravec's Paradox exposes a critical gap in the development pipeline for clinical AI. We cannot assume that models capable of sophisticated diagnostic reasoning possess the basic visual competence required for safe deployment. Input validation must be treated as a first-class capability, with dedicated benchmarks, explicit training objectives, and architectural components that enforce visual grounding before language generation.

Moving forward, several open questions demand attention. How can we architect VLMs to perform explicit pre-diagnostic verification as a gating mechanism for downstream tasks? Can we develop multimodal alignment techniques that bind linguistic outputs more tightly to visual constraints? And how should regulatory frameworks evaluate input validation capabilities alongside diagnostic accuracy?

The MedObvious benchmark provides the tools to measure progress on these questions. Until VLMs can reliably detect that a chest X-ray is not a brain MRI, or that an image is upside down, we risk building clinical AI systems that are articulate but blind, fluent but unsafe. The path to trustworthy medical AI requires us to solve the obvious before we claim mastery of the complex.