ChatGPT’s New Deep Research Feature: A Game-Changer for AI Research?

The generative AI landscape is developing by leaps and bounds, and amid growing rivalry, OpenAI presents its deep research agent. What is it, and what does this advancement mean for the industry? We spoke with Volodymyr Getmanskyi, Head of the Data Science Office at ELEKS, to bring you the latest expert insights.

What is OpenAI’s deep research?

Deep research is a new agentic capability in ChatGPT that can conduct research by browsing the internet. Its concept is similar to the Deep Research feature presented by Google late last year. Think of it as a personal assistant that can make your life easier and quickly become an expert on any topic.

OpenAI's deep research runs on the o3 version–their newest and top-ranked model according to LiveBench, a benchmark for LLMs)–designed to browse the web and crunch data. In practice, this means you can ask the AI agent to research something, and it'll go out and read tons of websites, put all that data together, and give you back a report with proper sources, kind of like what you'd get from a professional researcher.

What are the current usage parameters of the deep research feature?

Currently, the deep research feature is available only to Pro plan users, who are allowed up to 100 research queries per month. OpenAI indicates that completing a research task can take anywhere from 5 to 30 minutes, but as some users indicate it can take even longer. In addition, the deep research feature is quite compute-intensive; therefore, the more time it takes to research a query, the greater the computing resources required.

At this point, the deep research can scan publicly available information and analyse any files you upload. However, the company is exploring options to connect with more specialised data sources and to expand access to subscription-based or internal resources.

And what are the limitations of this research agent?

As the OpenAI states, it can still hallucinate some facts in responses or make incorrect inferences. It may also struggle to distinguish between factual information and internet rumours. When first using the research agent, users might notice some small formatting issues in the reports and citations, which might take a minute to get going.

In my opinion, the key limitation so far is that it can only work with what's out there on the open web and can't access information behind paywalls, textbooks, government databases, or old documents that haven't been digitised yet.

Additionally, data quality can be a real challenge. This research tool will need to make sense of all information available, but they might end up using some not-so-great sources or even AI-generated stuff that's full of hallucinations or mistakes.

What is the difference between OpenAI's and Google's deep research agents; are there any other competitors?

If you're looking for a general overview, Google's research version is excellent. However, if you want a researcher to explore multiple sources and delve into the details with a strong opinion, then OpenAI's agent is a better choice. It's important to note that neither platform has access to paywalled research and publications, which currently limits the capabilities of both agents.

There is also news circulating that a group of developers at the AI development platform called Hugging Face has created an "open" version of OpenAI's deep research tool. And what’s interesting is that they have reached quite promising results with converting textual instructions to code. This can be seen in the prompting chain examples on GitHub.

It's important to note that several reproductions of OpenAI's deep research can be found online, some of which use open models and tools. However, these reproductions are missing a crucial component–o3, the model that underlies deep research and its speed.