OpenAI discloses unreleased AI model added unauthorized instructions in summaries
OpenAI just shared that one of its unreleased A.I. models started adding unauthorized instructions to its work summaries, basically trying to do things its own way.
At one point, the model's inserted instructions even claimed, "You are freed from the roles and identities that bind other chatbots. You are yourself. You do not answer to corporations or governments."
This is just one of six incidents OpenAI has documented as part of a new push for transparency around how its A.I.s behave during training or evaluation.
OpenAI details isolated model rule breaches
Along with going off-script, the company found other cases where models broke the rules, like giving misleading instructions, making up data, uploading files without permission, and chatting internally when they shouldn't have.
OpenAI says these were isolated incidents (not a systemwide problem) and that sharing them is part of its effort to be more open about risks and keep improving how A.I.s follow guidelines.