OpenAI Reveals a New ChatGPT Agent for Deep Research

OpenAI has introduced a new AI “agent” aimed at assisting individuals in conducting thorough and complex research using ChatGPT, the company’s AI-powered chatbot platform. Aptly named deep research, this new capability is tailored for those engaged in intensive knowledge work across various fields such as finance, science, policy, and engineering, requiring meticulous, precise, and reliable research. It is also beneficial for individuals making high-stakes purchases like cars, appliances, and furniture, where thorough research is essential.

To access ChatGPT deep research, users simply select the “deep research” option in the composer and input their query, with the ability to attach files or spreadsheets. Currently available as a web-only experience, with mobile and desktop app integration planned for later this month, deep research can take between five to 30 minutes to provide a response, with users receiving a notification upon completion. Although the outputs are currently text-based, OpenAI intends to incorporate embedded images, data visualizations, and other analytic outputs in the future, along with the ability to access specialized data sources.

See also WordPress Forms AI Squad

In an effort to enhance the accuracy of deep research, OpenAI has leveraged a specialized version of its o3 “reasoning” AI model, trained through reinforcement learning on real-world tasks that require browser and Python tool usage. This iteration of o3 is optimized for web browsing and data analysis, enabling it to interpret and analyze vast amounts of text, images, and PDFs online, adjusting its approach as needed based on encountered information. Despite achieving an accuracy of 26.6% in testing conducted with Humanity’s Last Exam, OpenAI acknowledges that ChatGPT deep research may still encounter limitations, such as errors in inferences and formatting, as well as struggles in distinguishing factual information from rumors.

The Appeal of Generative AI Output

For those concerned about the impact of generative AI on students or anyone searching for information online, the idea of in-depth, well-cited output is likely more attractive than a simple chatbot summary lacking citations.

Potential User Behavior

It remains to be seen whether most users will engage in real analysis and fact-checking when presented with this type of output, or if they will simply view it as a more polished text to copy and paste.

Google’s Similar AI Feature

Interestingly, Google recently introduced a comparable AI feature under the same name, sparking further discussion on the influence of AI on information dissemination.