NIST’s AITE launch proposed tests on data kept from model developers

NIST announced Artificial Intelligence Technology Evaluation, or AITE, on July 27, 2026. The program offered a sequestered environment for evaluating models on data withheld from developers, addressing the risk that test material has already entered a model’s training data. ^[1]

Its initial tasks concerned image analysis with large vision-language models in quantum science, genomics and public safety. NIST said the program would supply shared data, metrics and scoring so results could be compared on the same tasks. ^[1]

The announcement described two participation routes. Data providers could supply an otherwise inaccessible dataset and a meaningful task. Model providers could submit systems for testing across those datasets and learn how their performance compared under common measures. ^[1]

This was a program launch, not a published ranking or a demonstration that any named model was dependable in everyday use. The archived notice explains an evaluation design: control access to test data and make comparisons consistent. The strength and limits of later findings would depend on the specific tasks and results reported. ^[1]

About this item: Retrospective source brief. No first-hand reporting or independent testing is claimed. Corrections policy.