80 Million Records Expose AI‑Generated Data Pollution Undermining Diagnostic Reliability

Real‑world Dilemma: Hidden Risks of Medical AI Generation

Generative AI is rapidly being integrated into clinical documentation such as reports, discharge summaries, and electronic health records. While improving efficiency, AI‑generated content increasingly replaces manually curated medical data, creating a self‑reinforcing "generate‑train‑regenerate" cycle that threatens diagnostic safety.

Core Findings: Systematic Performance Degradation Across Tasks

The research team analyzed 216,307 radiology reports, 790 clinical notes, 1,000 ophthalmology records, and 9,781 chest X‑ray images—over 800,000 synthetic samples—across three tasks: clinical text generation, vision‑language radiology reporting, and medical image synthesis. Experiments with multiple representative model architectures demonstrated a consistent degradation pattern:

Vocabulary in radiology impressions fell from 12,078 unique tokens to about 200 (a 98.9% drop).

Distinct medical terms decreased by 66%, and reports became highly formulaic.

In vision‑language reporting (Swin‑Transformer + Llama‑2 R2GenGPT), report uniqueness dropped from 96.2% to 0.9% and token count from 8,186 to 94.

False confidence surged: the rate of "no acute finding" errors rose from 13.3% to 40.3% while model confidence remained high.

Synthetic medical images exhibited visual quality loss, pathological distortion, and amplified demographic bias.

These degradations were not limited to a single data type; they persisted across text, multimodal reports, and image synthesis, indicating that self‑training loops erode pathological variability and diagnostic reliability while masking the decline with over‑confident predictions.

Mitigation Strategies Evaluated

To assess clinical relevance, the generated outputs were structurally reviewed, edited, and evaluated by physicians, confirming the observed utility loss. The authors tested three mitigation approaches:

Mixed‑real‑data training: Incorporating real data to constitute at least 75% of the training set preserved pathological diversity and language fidelity, substantially reducing demographic bias.

Quality‑aware filtering: Leveraging limited real data to filter synthetic samples improved efficiency but could not replace the need for a high proportion of authentic data.

Pure synthetic augmentation: Adding only synthetic data proved ineffective and accelerated degradation, even worsening gender bias.

Conclusion and Policy Outlook

The authors advocate for mandatory data provenance and enforced human verification in medical AI deployment, arguing that voluntary oversight is insufficient. Without regulatory safeguards, generative AI could contaminate future patient data ecosystems, posing escalating safety risks.

Paper title: "AI‑generated data contamination erodes pathological variability and diagnostic reliability" (arXiv:2601.12946).