OpenAI publishes GPT-4o system card

OpenAI published a system card for GPT-4o, the omnimodal model it had announced in May, covering the risks specific to a model that could take in and generate audio directly rather than routing speech through separate transcription and synthesis systems.

The new material concerned voice. OpenAI said testing had turned up rare cases in which the model unintentionally generated output in a voice resembling the user’s own, particularly with noisy input audio — a capability the company judged could enable impersonation or fraud if left uncontrolled. Its response was to restrict output to a fixed set of approved voices and run a separate classifier during generation to detect and block anything else; OpenAI described the residual risk of unauthorised voice generation as low. The card also documented other audio-specific decisions: speaker identification from a voice was disabled, and the model was trained to refuse singing and other copyrighted audio, to avoid replicating protected material.

Beyond audio, the card reported that red-teamers found the model no more persuasive than a human when using voice to argue political points, and noted that some users formed emotional attachments to the assistant during testing — flagged as worth monitoring rather than solved. The card extended the dangerous-capability evaluation format GPT-4’s had introduced in 2023 to a new class of risk that text-only cards had no reason to cover, and it set a pattern later voice-capable releases followed.