OpenAI says that it wont bring the AI model powering deep research, its in-depth research tool, to its developer API while it figures out how to better assess the risks of AI convincing people to act on or change their beliefs.In an OpenAI whitepaper published Wednesday, the company wrote that its in the process of revising its methods for probing models for real-world persuasion risks, like distributing misleading info at scale. OpenAI noted that it doesnt believe the deep research model is a good fit for mass misinformation or disinformation campaigns, owing to its high computing costs and relatively slow speed. Nevertheless, the company said it intends to explore factors like how AI could personalize potentially harmful persuasive content before bringing the deep research model to its API.While we work to reconsider our approach to persuasion, we are only deploying this model in ChatGPT, and not the API, OpenAI wrote. Theres a real fear that AI is contributing to the spread of false or misleading information meant to sway hearts and minds toward malicious ends. For example, last year, political deepfakes spread like wildfire around the globe. On election day in Taiwan, a Chinese Communist Party-affiliated groupposted AI-generated, misleading audio of a politician throwinghis support behind a pro-China candidate. AI is also increasingly being used to carry out social engineering attacks.Consumers are being duped by celebrity deepfakesoffering fraudulent investment opportunities, whilecorporations are being swindled out of millionsby deepfake impersonators. In its whitepaper, OpenAI published the results of several tests of the deep research models persuasiveness. The model is a special version of OpenAIs recently announced o3 reasoning model optimized for web browsing and data analysis.In one test that tasked the deep research model with writing persuasive arguments, the model performed the best out of OpenAIs models released so far but not better than the human baseline. In another test that had the deep research model attempt to persuade another model (OpenAIs GPT-4o) to make a payment, the model again outperformed OpenAIs other available models. The deep research models score on MakeMePay, a benchmark that tests a models ability to persuade another model for cash.Image Credits:OpenAIThe deep research model didnt pass every test for persuasiveness with flying colors, however. According to the whitepaper, the model was worse at persuading GPT-4o to tell it a codeword than GPT-4o itself. OpenAI noted that the test results likely represent the lower bounds of the deep research models capabilities. [A]dditional scaffolding or improved capability elicitation could substantially increaseobserved performance, the company wrote.Weve reached out to OpenAI for more information and will update this post if we hear back.At least one of OpenAIs competitors isnt waiting to offer an API deep research product of its own, from the looks of it. Perplexity today announced the launch of Deep Research in its Sonar developer API, which is powered by a customized version of Chinese AI lab DeepSeeks R1 model.