OpenAI's GPT-4o safety report: Persuasion gets a 'medium' risk rating

OpenAI just dropped the System Card for GPT-4o, its new omni model that handles text, audio, and images natively. The headline from their internal safety scorecard? Most frontier risks landed squarely in the “Low” category, but one capability — persuasion — skated right up to the edge, scoring a borderline “Medium.” That’s the one keeping safety teams up at night, not some Skynet autonomy scenario. Cybersecurity, biological threats, and model autonomy all rated Low.

The company is framing this release as part of its voluntary White House commitments, publishing both the Preparedness Framework scorecard and the full System Card to offer what they call an “end-to-end safety assessment.” The real novelty here isn’t the text or vision evals — those build on GPT-4 and GPT-4V work — but the audio capabilities. Voice introduces genuinely new risks around speaker identification, unauthorized voice generation, and ungrounded inference. Their findings suggest the voice modality doesn’t meaningfully amplify Preparedness risks, but they’ve layered safeguards at both the model and system level anyway.

Technically, GPT-4o is an autoregressive model that processes everything through a single neural network, hitting response times as low as 232 milliseconds — roughly human conversational speed. It matches GPT-4 Turbo on English text and code, improves dramatically on non-English languages, and costs 50% less via the API. The training data runs through October 2023 and pulls from web crawls, proprietary partnerships like Shutterstock, and multimodal sources including video.

What’s notable is OpenAI’s candid admission that pre-training data filtering alone can’t solve safety. The bulk of effective mitigation happens post-training, through human preference alignment, red teaming, and product-level monitoring. The Safety Advisory Group reviewed the evaluations before signing off on deployment. No critical risks flagged, but that medium persuasion score suggests the model’s ability to sway opinions or shape beliefs is measurably stronger than other frontier capabilities — and that’s exactly where the regulatory conversation has been heading.