OpenAI slammed over training-data opacity: either no traceability or refusing to disclose

davidmanheim · x · 2026-09-09

Safety researcher davidmanheim presses OpenAI on its statement that it 'cannot rule out' de-identified usage data helped improve its models: why won't it say what the ground truth is?
His dilemma: either OpenAI lacks tooling to trace what it trains on (unacceptable), or it has it but won't disclose (damning).
He also notes the newer model OpenAI began training on Aug 28 could have ingested the data, and argues the influence is more plausible than skeptics claim since training continued through a later date on the model in question.

More from Companies & People