OpenAI says its latest GPT-4o model is ‘medium’ risk

OpenAI has released its GPT-4o system carda research paper describing the security measures and risk assessments conducted by a startup before releasing its latest model.

GPT-4o went public in May of this year. Before its debut, OpenAI used an external group of red teams, or security experts, to try to find weaknesses in the system to find key risks in the model (which is pretty standard practice). They investigated risks like the possibility that GPT-4o would create unauthorized clones of someone’s voice, sexual and violent content, or fragments of reproduced copyrighted audio. Now the results are being published.

According to OpenAI’s own framework, the researchers deemed GPT-4o “medium” risk. The overall risk level was taken from the highest risk ratings of four broad categories: cybersecurity, biological hazardspersuasion and model autonomy. All of these were considered low risk, except for persuasion, where the researchers found that some GPT-4o writing samples could be better at influencing readers’ opinions than human-written text—although the GPT-4o samples were not more persuasive overall.

OpenAI spokeswoman Lindsay McCallum Rémy said: Edge that the system card contains readiness assessments created by the internal team with the participation of external testers listed on the OpenAI website such as Model Evaluation and Threat Research (METR) and Apollo Research, which evaluate AI systems.

What’s more, the company is releasing a highly effective multi-modal model just ahead of the US presidential election. There’s a clear potential risk that the model will inadvertently spread misinformation or be hijacked by malicious actors — even if OpenAI is keen to emphasize that the company is testing real-world scenarios to prevent misuse.

Many have called for OpenAI to be more crystal clear, not only about its model training data (is it trained on YouTube?) but also about its security testing. In California, home to OpenAI and many other leading AI labs, state Senator Scott Wiener is working to pass legislation to regulate immense language models, including restrictions that would hold the companies legally liable if their AI is used in a harmful way. If the bill passes, OpenAI’s frontier models would have to meet state-mandated risk assessment requirements before the models are made available for public apply. But the biggest takeaway from the GPT-4o system charter is that despite a group of external red teams and testers, much of it relies on OpenAI’s assessment.