GPT-4o System Card (PDF)

This document is the GPT-4o System Card published by OpenAI, a technical safety and governance disclosure describing the capabilities, risks, and mitigation measures associated with the GPT-4o model prior to its public release. The card states that OpenAI conducted external red teaming, frontier risk evaluations under its Preparedness Framework, and implemented behavioral mitigations across audio, vision, and text modalities to address identified risk areas including CBRN uplift, persuasion, model autonomy, and voice-based harms. The document discloses that GPT-4o introduces a native audio-visual modality that creates operationally distinct risk surfaces compared to text-only predecessors, including the potential for voice cloning, unauthorized speaker identification, real-time emotion inference, and generation of content that could enable social engineering or influence operations. The card engages the EU AI Act's high-risk and general-purpose AI classification frameworks, the FTC's consumer protection authority over deceptive AI-generated content, and broader voluntary AI safety commitments made by OpenAI to the US government. The document acknowledges residual risks across evaluated categories and states that GPT-4o received a 'medium' risk rating under the Preparedness Framework for CBRN and cybersecurity, with deployment authorized subject to ongoing monitoring and post-deployment mitigation updates.