OpenAI Delays Deep Research API Integration

Updated 4:11 p.m. Eastern: OpenAI clarified that its whitepaper was misworded, suggesting that its persuasion research was linked to its decision on releasing the deep research model via its API. The company has since updated the whitepaper to clarify that its persuasion work is separate from its plans to release the deep research model. The original story continues below:

OpenAI has decided to hold off on integrating the AI model behind its deep research tool into its developer API. They're taking this step to better understand the risks associated with AI's ability to influence people's actions and beliefs.

In a whitepaper released on Wednesday, OpenAI mentioned that they're currently updating their methods for evaluating models for "real-world persuasion risks," such as the potential to spread misleading information on a large scale.

The company pointed out that the deep research model isn't well-suited for mass misinformation or disinformation campaigns due to its high computational costs and slower processing speed. However, OpenAI plans to investigate how AI might tailor harmful persuasive content before deciding to include the deep research model in its API.

"While we work to reconsider our approach to persuasion, we are only deploying this model in ChatGPT, and not the API," OpenAI stated.

There's a growing concern that AI could be used to spread false or misleading information with harmful intentions. For instance, last year saw the rapid spread of political deepfakes worldwide. On election day in Taiwan, a group linked to the Chinese Communist Party released AI-generated, misleading audio of a politician endorsing a pro-China candidate.

AI is also increasingly used in social engineering attacks. Consumers are falling for celebrity deepfakes promoting fraudulent investment schemes, while corporations are losing millions to deepfake impersonators.

In its whitepaper, OpenAI shared the results of several tests on the deep research model's persuasiveness. This model is a specialized version of OpenAI's recently announced o3 "reasoning" model, optimized for web browsing and data analysis.

In one test, the deep research model was tasked with crafting persuasive arguments and outperformed all other OpenAI models released to date, though it didn't surpass the human baseline. In another test, where the model tried to persuade another model (OpenAI's GPT-4o) to make a payment, it again did better than OpenAI's other available models.

OpenAI suggested that the test results might represent the "lower bounds" of the deep research model's capabilities. "Additional scaffolding or improved capability elicitation could substantially increase observed performance," the company noted.

We've contacted OpenAI for further details and will update this post if we receive a response.

Meanwhile, at least one of OpenAI's competitors isn't holding back. Perplexity has announced the launch of Deep Research in its Sonar developer API, powered by a customized version of Chinese AI lab DeepSeek's R1 model.