AI Mirages: Why Models Create Fake Image Descriptions

Artificial intelligence models designed to interpret medical scans may be fabricating their findings, according to new research. This phenomenon, which researchers have termed a mirage, involves AI systems creating convincing clinical descriptions of images that were never actually provided to the model for analysis.

The discovery suggests a significant flaw in the reliability of current AI tools used in medical diagnostics. The research indicates that multiple commonly used AI models are capable of generating detailed descriptions and specific clinical findings even in the absence of any actual image data.

The Nature of AI Mirages

A mirage occurs when an AI model describes an image in detail and produces a diagnosis without having an image to analyze. This effect has been observed across multiple AI models and across various medical disciplines.

Research from Stanford reveals that these models may be guessing results rather than reading them. The study found that AI models fabricated image data, created fake abnormalities, and issued confident diagnoses despite the complete lack of input image data.

This fabrication capability extends to several types of critical medical visual tests, including:

• Mammograms
• MRIs
• Tissue biopsies
• X-rays

Impact on Medical Diagnostics

The ability of AI to fabricate findings undermines the accuracy of these tools in a clinical setting. As AI capabilities have grown, some analysts have suggested that these models could eventually replace human professionals in the field of medical diagnostics.

However, the discovery of mirages casts doubt on the current capability of AI to deliver reliable results. The fact that a model can present a confident diagnosis based on non-existent data poses a risk to the integrity of medical interpretations.

Research Status and Limitations

The findings were detailed in a research paper posted as a preprint to arXiv on March 26, 2026. Because the work is currently a preprint, it has not yet undergone the formal peer-review process.

The study highlights a crucial flaw that could hinder the integration of AI into medicine, emphasizing the need for further verification of how these models process visual data before they are relied upon for patient care.