OpenAI has published its GPT-4o datasheet, a document detailing the security measures and risk assessments the company performed ahead of the release of its latest AI model. GPT-4o was launched publicly in May of this year. Before its release, OpenAI engaged an outside group of security experts, known as red teamers, to identify vulnerabilities in the model, a fairly common practice in the industry. The experts examined risks such as the possibility that GPT-4o could create unauthorized clones of someone’s voice, create erotic and violent content, or create segments of copyrighted audio. Now, the results of those assessments have been made public.
According to OpenAI’s scoreboard, researchers rated GPT-4o as “medium risk.” The overall risk level was determined by the highest risk among four main categories: cybersecurity, biological threats, persuasion, and model autonomy. All of these categories were rated low risk, except for persuasion, where some text samples produced by GPT-4o were found to be more effective in influencing readers’ opinions than human-written texts, though overall the model’s samples were no more persuasive.
There have been calls for OpenAI to be more transparent, not just about its model training data, but also about its safety testing. In California, where OpenAI and several other AI labs are based, state Sen. Scott Wiener is trying to pass a law to regulate large language models, including restrictions that would make companies legally liable if their AI is used in harmful ways. If the law passes, OpenAI’s AI models will have to comply with state-mandated risk assessments before being made available to the public.
#ChatGpt #OpenAis #medium #risk
