OpenAI claims to have introduced a new safety feature for GPT-5, but OpenAI has had this feature since 2021 I'm sad OpenAI is implying they created this for GPT-5, which they didn't. It's hard to rely on AI companies' System Cards when there's incentive to overclaim. (thread)

Here's OpenAI describing end-user IDs - the same safety feature - for GPT-4's launch.

I wanted to give benefit of the doubt: maybe OpenAI's claim was an oversight & they'd forgotten about the existing feature. But that page was updated to just list the new name, with roughly no change to the feature's description

Here's OpenAI describing how to use this API field, on their Support website, back in 2021

I understand that OpenAI's quoted text might be true in letter (there is now a field called safety_identifier which wasn't there before), but the spirit is not: Someone reading this would this would think this is a new measure taken because of the risk of GPT-5. That is false.

I'm glad to see OpenAI promote end-user identifiers, which OpenAI was ahead of the curve on. But when AI labs falsely claim their techniques are new and motivated by novel risks of recent models, it's hard to tell if they're really responding to risks appropriately.