DeepMind researcher questions OpenAI data setting: human review of history even when training opt-off

CsabaSzepesvari · x · 2026-09-16

In a reply thread with OpenAI, DeepMind researcher Csaba Szepesvári raises data-ethics concerns: even with the setting "off", a human looking at his history technically does not violate the off state; whatever is learned should not be used for training, but someone could in principle just take his prompts. He adds that mathematicians are actually nice (contrary to public Twitter perception) and would not necessarily turn the setting off, noting that Tristan, purposefully or not, had it on until mid-summer.

Related event: DeepMind Researcher Questions OpenAI's Data Privacy Settings(4 posts)→

Original post →

More from Safety

Safety channel →