GPT-4o System Card (PDF)

This is OpenAI's System Card for GPT-4o, a multimodal AI model that accepts text, audio, image, and video inputs and generates text, audio, and image outputs, published alongside OpenAI's Preparedness Framework scorecard to document safety evaluations conducted prior to deployment. The document discloses that GPT-4o's persuasion capability scored borderline medium risk on the Preparedness Framework, that a real-time streaming classifier blocks audio output deviating from a preset list of approved voices, and that post-training mitigations were applied to reduce speaker identification, ungrounded inference about speaker traits such as race, intelligence, and criminal history, and generation of erotic or violent speech. The document also acknowledges unresolved limitations including reduced safety robustness under noisy or interrupted audio conditions, instances of non-native accent generation in non-English speech output, and ongoing misinformation risks that may be more persuasive when delivered through voice.

This document is the GPT-4o System Card published by OpenAI on August 8, 2024, governing the safety evaluation and deployment disclosure of the GPT-4o multimodal model, consistent with OpenAI's voluntary commitments to the White House and its internal Preparedness Framework. The card states that GPT-4o was evaluated across cybersecurity, biological threats, persuasion, and model autonomy risk categories, with post-mitigation scores of low for three categories and medium for persuasion, and the document asserts that only models scoring medium or below post-mitigation may be deployed. Notable provisions include the deployment of a real-time streaming voice output classifier that the document states catches 100% of meaningful deviations from approved system voices, post-training mitigations against speaker identification, ungrounded inference, and sensitive trait attribution, and an explicit acknowledgment of residual risks including audio robustness degradation under noise or interruption and unresolved misinformation generation through audio. The document engages with the EU AI Act's high-risk AI system disclosure expectations, FTC Act principles regarding deceptive or unfair practices, and broader AI governance voluntary commitment frameworks; applicability of specific regulatory obligations depends on jurisdiction and the evolving regulatory landscape for foundation models. Material compliance considerations include the adequacy of disclosed residual risks for downstream API operators, the scope of data filtering practices including CSAM and personal information reduction, and the treatment of third-party voice actor data used in approved voice generation.

The document establishes that GPT-4o's voice output is restricted to a preset list of approved voices generated in collaboration with voice actors, enforced by a real-time streaming classifier that the card states catches 100% of meaningful deviations from approved voices. Under these terms, the model is post-trained to refuse requests to identify speakers by voice, to decline ungrounded inferences about speaker traits such as race, intelligence, or political attributes, and to block erotic or violent speech output by running a moderation classifier over audio transcriptions. The document acknowledges residual risks including reduced safety robustness under low-quality or interrupted audio and ongoing susceptibility to misinformation generation through audio prompting, which OpenAI states are areas of active mitigation development.