OpenAI has introduced a new feature in ChatGPT called Deep Research. This feature acts as an independent AI agent within the chatbot, designed to solve complex tasks quickly. It integrates current information from the internet and presents research data in a visually appealing manner.
Deep Research is based on a version of the upcoming AI model, o3. This AI is capable of analyzing vast amounts of text, images, and PDFs from the internet and responding dynamically to the information it finds. OpenAI has trained the AI to plan the task-solving process and then address it step-by-step.
For example, OpenAI developers asked Deep Research for a report on changes in retail over the past three years. The AI was instructed to organize the report with bullet points and tables. However, before starting, Deep Research asked the user which aspects of retail should be focused on, such as consumer behavior, supply chains, or new technologies.
Once Deep Research knows which areas are important, it searches the internet for relevant data. For instance, it might find a source stating that e-commerce accounted for 19.4% of global trade in 2023, with analysts predicting an increase to 22.6% by 2027. The AI then draws conclusions and seeks additional data to support these findings. Ultimately, it delivers a complete report with all essential topics and explanations for its assumptions.
Deep Research is particularly useful for expert-level tasks. For example, in linguistics, it can help users develop a new language intended for use five centuries in the future. This language would be based on English but incorporate characteristics from other languages to reflect historical invasions by various nations. According to OpenAI, Deep Research can complete such a task in just a few minutes, whereas human experts might take around five hours.
Deep Research can also be used in other areas. For instance, users can search for a scene from a specific series. The AI can scan product texts and series descriptions to provide accurate answers, including a detailed scene description and character names. In contrast, GPT-4o might provide incorrect information for the same query.
Despite its capabilities, Deep Research still has some limitations. As an early version of the AI agent, it may occasionally experience hallucinations, struggling to distinguish facts from rumors. Formatting issues can also occur. However, OpenAI expects to resolve these issues in the coming weeks and months based on user feedback.
Initially, Deep Research is available only to Pro users in the USA, who can make up to 100 requests per month. Access will later be extended to Plus and Team users, and finally to Enterprise users. It is not yet known when the feature will be available to users in Europe. OpenAI plans to develop a more cost-efficient version of Deep Research that provides equivalent responses, intending to lift current limitations and allow for more requests per month.
