A research document detailing the safety assessments and risk evaluations conducted before the model's release. GPT-4o, which launched in May 2024, underwent scrutiny by external security experts (referred to as red teamers) to identify potential risks, such as unauthorized voice cloning, inappropriate content generation, and reproduction of copyrighted audio.

The assessment categorized GPT-4o as having a "medium" risk level, primarily due to concerns in the persuasion category, where some generated content showed a higher potential for influencing opinions than human-written text. This evaluation was part of a broader framework that also examined cybersecurity, biological threats, and model autonomy, all of which were rated as low risk.

Lindsay McCallum Rémy, an OpenAI spokesperson, emphasized that the evaluations were conducted by both internal teams and external testers from organizations like Model Evaluation and Threat Research (METR) and Apollo Research.

The release of the GPT-4o System Card comes at a critical time for OpenAI, which has faced mounting criticism over its safety practices. Just before the system card's release, an open letter from Sen. Elizabeth Warren and Rep. Lori Trahan questioned the company's handling of whistleblower reports and safety reviews. This scrutiny follows previous controversies, including the brief ousting of OpenAI's CEO, Sam Altman, and the departure of a safety executive over concerns about the company's priorities.

The introduction of GPT-4o also coincides with the lead-up to a U.S. presidential election, raising concerns about the potential misuse of the model to spread misinformation. In response, there have been calls for greater transparency from OpenAI, especially regarding its training data and safety testing processes. Additionally, California state Sen. Scott Wiener is advocating for legislation that would regulate large language models, including imposing legal responsibilities on companies if their AI systems are used harmfully.

Despite these external evaluations, the responsibility for assessing GPT-4o's risks still largely falls on OpenAI itself.