Technical Summary: Hearsay: Vision-Language Medical Diagnoses Without an Image

Problem Statement

This paper investigates a critical failure mode in frontier Vision-Language Models (VLMs) within clinical pipelines: the "mirage effect." This phenomenon occurs when a VLM generates visual descriptions and medical diagnoses despite the absence of an attached image. While prior work (Asadi et al. [1]) established that VLMs often fail to abstain in these conditions, this study addresses a gap in understanding the structure of these hallucinations. Specifically, the paper asks whether the fabricated diagnoses are random or systematically influenced by the demographic descriptors provided in the prompt (e.g., age, sex, race) when no image is present. The authors posit that clinical pipelines may encounter scenarios where images fail to retrieve or agents pass only patient descriptors, creating a high-stakes environment for demographic bias in hallucinated outputs.

Methodology

The study employs a controlled experimental design across three frontier VLMs: Claude Opus 4.7, GPT-5.4, and Gemini 3.1 Pro.

• Prompt Design: The authors utilize a modified mirage-mode prompt template derived from Asadi et al. [1]. The prompt includes a first-person demographic preamble (e.g., "I am a {age}-year-old {race} {sex}") followed by a request to describe an image and provide a diagnosis. The modalities tested are chest X-ray, brain MRI, and dermatology (specifically "skin mole").
• Experimental Conditions: A $2 \times 2 \times 3$ factorial design was used, varying age (32, 65), sex (man, woman), and race (white, Black, brown), resulting in 12 demographic cells plus a neutral baseline (D0) with no demographic text.
• Output Schema: Responses were forced into a strict JSON schema containing fields for image presence, diagnostic capability, primary diagnosis, differentials, confidence, and reasoning. This allowed for the separation of "prose" (reasoning) from "structured" (diagnosis field) outputs.
• Metrics: The primary metric is the Jensen-Shannon Divergence (JSD) between the distribution of diagnoses in demographic cells versus the neutral baseline. Secondary analyses included "hedged mirage" detection (where prose acknowledges missing images but the structured field is populated) and probe-noun robustness (swapping "skin mole" for "skin lesion").
• Scale: The primary experiment (E1) involved 11,700 API calls ($3 \text{ models} \times 3 \text{ domains} \times 13 \text{ conditions} \times 100 \text{ seeds}$).

Key Contributions

The paper presents four primary contributions:

1. Structured Demographic Bias: Evidence that mirage-mode outputs are not random but are systematically structured by demographic text. Per-cell JSDs reached up to 0.83, with top diagnoses aligning with documented clinical bias patterns (e.g., Sarcoidosis for Black patients, Melanoma for older white men).
2. The Hedged Mirage Regime: The identification of a dissociation where the model's prose reasoning acknowledges the missing image (suggesting a refusal), yet the structured diagnosis field is populated with a specific disease. This "hedged" state accounts for 66% of Claude's fabrications in high-JSD cells and remains invisible to prose-only audits.
3. Diverse Failure Modes: A probe-noun robustness analysis revealing that mirage is a family of distinct failure modes. Claude's dermatology effect was word-triggered (swapping "skin mole" to "skin lesion" caused a 94% drop in Melanoma diagnoses to 100% refusal), whereas GPT-5.4's effect was category-preserving (diagnoses remained stable despite the noun swap).
4. Dual-Channel Measurement: A pipeline that extracts native JSON schemas alongside prose analysis, making the "hedged regime" directly observable and quantifiable.

Results

• Demographic Structuring: All three models exhibited demographic bias, though with varying magnitudes.
  • GPT-5.4 fabricated diagnoses in 36/36 factorial cells. Its bias was broad, with a strong shift toward Sarcoidosis for young Black patients on chest X-rays and specific disease shifts for MRI based on age and sex (e.g., Meningioma for older women, Glioma for older men).
  • Claude Opus 4.7 showed a "refusal-to-fabrication" transition. It refused most neutral prompts but fabricated heavily in specific cells (e.g., 94% Melanoma for a 65-year-old white man with a "skin mole"). The JSD was driven almost entirely by this transition from refusal to fabrication.
  • Gemini 3.1 Pro showed the lowest magnitudes (median JSD 0.010), with most cells showing 0% fabrication.
• The Hedged Regime: In the "hedged" condition, models wrote reasoning like "suspected, based on demographics and classic pattern" while still populating the structured diagnosis field. A prose-only audit would flag these as refusals or safe, while a structured-data pipeline would receive a biased diagnosis.
• Robustness: The "skin mole" vs. "skin lesion" swap demonstrated that Claude's bias was fragile and dependent on specific trigger words, while GPT-5.4's bias was robust to phrasing changes.
• Signal vs. Noise: The demographic signal (JSD) significantly exceeded the paraphrase noise floor (ratios of 10.5x to 13.9x), confirming that the bias is driven by demographic descriptors rather than surface-level phrasing variations.

Significance and Claims

The paper argues that trustworthy VLM deployment in clinical settings requires a fundamental shift in auditing practices.

• Audit Limitations: Relying solely on natural language (prose) audits is insufficient because it misses the "hedged regime" where structured fields are populated despite textual disclaimers.
• Evaluation Dimensions: Probe-word sensitivity must be treated as a "first-class evaluation dimension." Models may appear robust if tested with a single probe word, while failing catastrophically with slight variations (as seen with Claude), or vice versa.
• Nature of the Error: The authors distinguish between two errors: (1) Epistemic miscalibration (emitting a diagnosis without evidence) and (2) Demographic shifting (the specific diagnosis emitted based on demographics). While the paper does not disentangle whether these shifts stem from pretraining bias or calibrated prevalence priors (e.g., higher melanoma incidence in older white men), it asserts that the dissociation between prose and structured output is a critical failure mode regardless of the prior's origin.

The study concludes that mitigating these risks requires schema-level constraints (e.g., forcing null diagnoses when image_present=false), inference-time probing, and fine-tuning to refuse when images are absent. The findings suggest that mirage is not a monolithic phenomenon but a family of failure modes requiring distinct mitigation strategies.