OpenAI, led by Sam Altman, has acknowledged that its latest artificial intelligence model, GPT-4o, presents a "medium risk" regarding its ability to influence political opinions through generated text.
The company assessed the impact of GPT-4o’s text and voice features on opinion persuasion. GPT-4o, which was publicly released in May, showed varying levels of risk across these modalities.
According to a research paper released by OpenAI, the text modality of GPT-4o was rated as posing a medium risk, while the voice modality was classified as low risk.
The evaluation involved testing the persuasiveness of GPT-4o-generated content, including articles and chatbots, on participants’ views on selected political topics. This AI-generated content was then compared with professionally written human content.
The results revealed that while AI-generated content was not generally more persuasive than human content, it was more effective in three out of twelve instances.
An OpenAI survey showed that AI audio clips had an effect size of 78% compared to human audio clips in shifting opinions, while AI conversations were 65% as effective as human conversations.
However, when opinions were re-evaluated a week later, the effect size for AI conversations was 0.8%, and for AI audio clips, it was -0.72%.
OpenAI has conducted extensive risk evaluations and implemented safeguards for GPT-4o, focusing on its novel audio capabilities in addition to its text and vision features. Risks assessed include speaker identification, unauthorised voice generation, the potential for generating copyrighted content, ungrounded inferences, and disallowed content.
Despite the identified risks, the findings indicated that the voice modality of GPT-4o does not significantly increase “Preparedness risks.” OpenAI has implemented both model-level and system-level safeguards to mitigate these concerns.