OpenAI has decided to hold off on integrating a special version of its o3 "reasoning" model into its developer API due to concerns over potential risks. This model, optimized for web browsing and data analysis, was put through rigorous testing to evaluate its persuasiveness, with results published in a recent whitepaper. Although the model excelled in certain tests, it did not surpass human performance and raised questions about AI's ability to influence human beliefs and actions.
The whitepaper detailed several tests designed to measure the deep research model's persuasive abilities. These included tasks such as crafting convincing arguments and attempting to persuade another model, GPT-4o, to make a payment or reveal a codeword. While the model outperformed other OpenAI models in these scenarios, it struggled to surpass GPT-4o in a critical test involving the extraction of a codeword.
OpenAI expressed concerns about the potential real-world persuasion risks that AI models could pose, such as spreading misinformation at scale. The company is actively revising its methods to better understand and mitigate these risks.
“Additional scaffolding or improved capability elicitation could substantially increase observed performance,” OpenAI noted in its whitepaper.
Interestingly, while the deep research model demonstrated superior performance relative to its predecessors, it still fell short of matching the human baseline in key areas. These findings underscore the challenges AI faces in replicating complex human cognitive tasks such as persuasion.
OpenAI has chosen to limit the deployment of this model to ChatGPT only, excluding it from their API offerings for the time being. The decision reflects an abundance of caution as the company continues to refine its approach to handling AI persuasion risks.
“While we work to reconsider our approach to persuasion, we are only deploying this model in ChatGPT, and not the API,” stated OpenAI.
Leave a Reply