GPT-4o System Card (PDF)

This document is the GPT-4o System Card published by OpenAI, a pre-deployment safety disclosure governing the release of the GPT-4o multimodal AI model (text, audio, and image inputs/outputs); it functions as an internal and public accountability instrument under OpenAI's Preparedness Framework rather than a legally binding consumer contract. The most significant obligations it documents are OpenAI's self-imposed safety evaluation procedures, including external red teaming, Preparedness Framework frontier risk scoring across CBRN, cybersecurity, persuasion, and model autonomy domains, and the deployment of content classifiers and system-level mitigations prior to public release. Notable provisions that deviate from industry standard include explicit scoring of GPT-4o's uplift potential for chemical, biological, radiological, and nuclear weapons (rated 'medium' risk), acknowledgment of emotional manipulation and over-reliance risks from the audio modality's expressive voice capabilities, and the disclosure of a 'shallow' character consistency problem where the model's persona can be destabilized by adversarial prompting. The document engages the EU AI Act (particularly high-risk AI system classification obligations and transparency requirements under Articles 13 and 52), FTC Act Section 5 unfair or deceptive practices standards, and emerging NIST AI RMF guidance; material compliance considerations include whether the Preparedness Framework's self-certification model satisfies forthcoming mandatory third-party audit requirements under the EU AI Act, and whether disclosed residual risks in CSAM detection bypass and voice cloning constitute adequate consumer disclosure under FTC standards.

REGULATORY EXPOSURE: This document engages the EU AI Act (Regulation 2024/1689), particularly transparency obligations under Art. 13 (technical documentation), Art. 52 (disclosure of AI-generated content), and potential high-risk classification under Annex III; the FTC Act Section 5 (unfair or dece…