# Are AI-model review labels treated like hand labels (saved and provided on request)?

**URL:** <https://community.drivendata.org/t/are-ai-model-review-labels-treated-like-hand-labels-saved-and-provided-on-request/11609>\
**Category:** GEMS Prize Challenge\
**Created:** [November 15, 2026, 7:22pm UTC](https://community.drivendata.org/t/are-ai-model-review-labels-treated-like-hand-labels-saved-and-provided-on-request/11609 "2026-11-15T19:22:33Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![cohezion](https://avatars.discourse-cdn.com/v4/letter/c/87869e/32.png) [@cohezion](https://community.drivendata.org/u/cohezion)\
**Post date:** [November 15, 2026, 7:22pm UTC](https://community.drivendata.org/t/are-ai-model-review-labels-treated-like-hand-labels-saved-and-provided-on-request/11609/1 "2026-11-15T19:22:33Z")

</div>

Hi GEMS team,

In thread 11543 you confirmed that hand-labeling is permitted, provided the labels are saved and made available on request.

Does the same apply when the reviewer is a generative-AI model rather than a person? For example: candidate fault traces are produced by code from public lidar, and a vision-capable AI model reviews rendered images of each candidate and decides which ones to keep. The model’s per-candidate decisions are saved, would be provided on request, and the generative-AI use would be described in the narrative as section 3.2 requires.

Given that model outputs are not perfectly reproducible on re-run, would the saved decisions be enough for reproducibility at finalist verification, or is a re-runnable pipeline expected for that step?

Thanks!
