OpenAI has announced a new AI agent that will help people conduct complex, in-depth research using ChatGPT. This is the company’s AI powered chatbot platform. It’s appropriately called deep research. OpenAI stated in a blog posted Sunday that this capability was designed for people who do intensive knowledge-based work in areas such as finance, science and policy and engineering, and need thorough, accurate and reliable research. It could also prove useful for those making “purchases which require careful research like cars, appliances and furniture.” OpenAI announced that it is making deep research available today to ChatGPT users, with a limit of 100 queries per month. Support for Plus, Team, and Enterprise users will follow. OpenAI said it would roll out Plus in a month and that the query limits of paid users will be “significantly increased” soon. OpenAI did not have a release date for ChatGPT users in the U.K. or Switzerland.
To utilize ChatGPT deep search, you will select “deep research”in the composer, and then enter your query. You can also attach files or spreadsheets. It’s currently only available on the web, but mobile and desktop apps will be integrated later this month. You’ll receive a notification once the search is complete. Deep research can take anywhere between 5 and 30 minutes.
At the moment, ChatGPT’s deep research outputs are only text. OpenAI has said it will add embedded images, data visualisations, and “analytical” outputs in the near future. OpenAI said that the ability to connect with “more specialized sources” including “subscription based” and internal resources is also on the roadmap.
How precise is ChatGPT’s deep research? AI is not perfect. It is prone to hallucinations, and other types errors that can be harmful in “deep research” scenarios. OpenAI has said that every ChatGPT deep-research output will be “fully document, with clear citations, and a summary [the] of thinking, making
it easy to reference, and verify the information.” The jury is still out on whether these mitigations will be enough to combat AI errors. OpenAI’s AI powered web search feature, ChatGPT Search in ChatGPT is not without gaffes. It often gives incorrect answers to questions. TechCrunch found that ChatGPT Search returned less useful results for certain queries than Google Search.
OpenAI uses a special version o3’reasoning’ AI model, which was trained via reinforcement learning to achieve a goal. As the model approaches the goal, it is rewarded with virtual “rewards,” which, in theory, will make it more adept at the task. OpenAI’sOpenai/
OpenAI claimed that it tested ChatGPT Deep Research using Humanity’s Last Examan evaluation that included more than 3,000 questions at the expert level in a variety academic fields. The o3 deep research model achieved an accuracy of 26,6%. This might seem like a low grade, but Humanity’s last Exam was designed to be more difficult than other benchmarks in order to keep up with model advancements. OpenAI claims that the o3 deep research model was far ahead of Gemini Thinking (6.2%), Grok-2 (3.8%), as well as OpenAI’s GPT-4o (3.3%).
OpenAI still notes that ChatGPT’s deep research has limitations. It can sometimes make mistakes and incorrect conclusions. The company stated that deep research can have difficulty distinguishing between rumors and authoritative information. It also has a hard time expressing when it is uncertain about something.
This type of well-cited, in-depth output is likely to be more appealing to anyone concerned about the impact of generative AI, whether it’s on students or anyone searching for information online. We’ll see if users actually double-check and analyze the output, or simply copy-paste it.
If this sounds familiar, Google announced a similar AI tool with the exact name less than two month ago.
