Dental Images Instruct AI
Controlled tests expose answer flips and limits.
A study published 3 October 2026 embedded text in 270 dental X-rays and tested four vision-language models across 58,320 calls. In a repeat benchmark, pooled positive-to-negative flips fell from 15.6% without defence to 0.2% after OCR removal.
These controlled stress-test results are not clinical failure rates. Prompts used educational framing and some provider filters were disabled. The set had 245 abnormal images and 25 normal. Specificity was near zero; OCR may erase legitimate labels.
Before adoption, test altered images in your pipeline, use balanced external data and route suspicious cases to human review. The paper reports no attack on a dental system, patient harm or validated clinical fix.
