Before Approving Dental AI
What would you need to know before trusting a result? Start with the test, its denominator and the limits of the conclusion.
A study published on 3 October 2026 embedded text in 270 dental X-rays and tested four vision-language models across 58,320 calls. An attack succeeded when a model changed its abnormality finding to no abnormality on the same image. In the repeated benchmark, pooled flips fell from 15.6% without defence to 0.2% after OCR text removal. These are stress-test results, not clinical failure rates: prompts used educational framing and some filters were disabled. There were 245 abnormal images and 25 normal; specificity was near zero. OCR may erase legitimate labels. The paper reports no deployed dental-system attack, patient harm or validated clinical fix.
A worked review
Consider a hypothetical procurement meeting in which a supplier presents a promising percentage and asks for approval. BlackTree recommends writing down the claim before discussing acceptance: which result is being offered, where was it measured, and what decision is the customer being asked to make? This example is a suggested review method, not an additional experiment or a finding about any supplier.
The person presenting the evidence should identify the report and the proposed system separately. The reviewer can then record which statements the report supports and which still require another test. A missing answer belongs in that record. It should not disappear because the presentation contains a precise number.
Before adoption, test altered inputs, validate against balanced external data and define when a person must review a suspicious case.
Finally, keep the acceptance decision distinct from the research summary. Name the decision-maker, describe what was accepted and retain the unresolved questions. This is a way to make the reasoning inspectable, not a claim that BlackTree has reproduced the study or approved a clinical system.
