Click on any major artificial intelligence tool — including ChatGPT, Google Gemini, or Perplexity — and you might notice a new option known as Deep Research.
Companies like OpenAI, Google, Perplexity, xAI, and Microsoft have all rolled out similar Deep Research AI tools in recent months, part of an industry-wide push for autonomous models focused on fact-finding.
Deep Research-style models are more accurate and comprehensive than other forms of AI. However, they remain reliant on human fact-checking and guidance, even as their use in business and education has already begun to surge.
What is Deep Research?
Deep Research tools can investigate a variety of sources to generate a research report. When prompted by a user, these models search the internet for information and produce an analysis of their findings.
Offerings from Google, OpenAI, Microsoft, and Perplexity can also recalibrate their research steps, indicate their progress, and accept follow-up questions.
“Every output is fully documented, with clear citations and a summary of its thinking, making it easy to reference and verify the information,” OpenAI has said. “It is particularly effective at finding niche, non-intuitive information that would require browsing numerous websites.”
These tools are part of AI companies’ move towards agentic AI. Unlike generative AI, which creates new content like text or audio based on user input and training data, agentic AI can autonomously take action on behalf of its users.
Google has been “investing in developing more agentic models”, the company said in December, while OpenAI announced a new slate of tools for building AI agents in March.
Is it actually useful?
While Deep Research tools typically score higher on accuracy than other AI models, they can still provide incorrect information.
OpenAI’s Deep Research answered the BrowseComp question set — designed to test AI agents’ abilities to find information — with 51.5% accuracy, better than any other OpenAI model. But while the humans OpenAI tested could only answer 29.2% of the questions, they did so with 86.4% accuracy.
Meanwhile, Perplexity’s Deep Research scored 93.9% accuracy on the SimpleQA test, which measures AI’s skill at answering purely factual questions.
“One of the things we've got to recognise is we've still got to be discipline or subject matter experts to a degree, and then use AI to augment and work with us,” Lynn Gribble, a University of New South Wales Associate Professor and AI education practices expert, tells Azzet.
“Because of how AI is programmed, it will always sound plausible. It's designed to give you an answer that's plausible, that's creative, and that will meet what you asked it, whether or not that's truthful and trustworthy.” The human user is ultimately responsible for evaluating how valuable or accurate the AI model’s answer is, Gribble says.
Deep Research tools also require greater amounts of time and compute to compile their reports, compared to other AI models. In April, OpenAI began rolling out a “lightweight” version of Deep Research, which offers shorter answers from a different AI model, and OpenAI, Google, and Perplexity all limit queries to free users.
What impact could Deep Research have?
The rapid spread of AI research tools is largely due to “the efficiency argument”, according to Gribble, as companies emphasise productivity.
Deep Research models are frequently targeted at businesses. “Deep Research is the perfect tool if you’re an entrepreneur launching a small business and want to quickly gather a competitor analysis and recommendations for suitable locations, or if you’re a marketer researching recent AI-powered marketing campaigns to benchmark for 2025 planning,” Google has said, for example.
Around 69% of Australia and New Zealand C-suite executives who aim to prioritise AI will focus on agentic AI in 2025, according to a February report by Salesforce and YouGov. An April survey by Cloudera also found that 96% of IT leaders respondents plan to increase their use of AI agents in the next 12 months.
AI research tools have also begun to affect academia, with students and researchers turning to technology to write or research their work. “We’re already seeing articles being retracted,” Gribble says, as AI-induced inaccuracies have been found in published papers.
A 2024 report by Atomi found that 60% of Australian teachers surveyed were concerned about the ethical implications of AI in a classroom setting. The use of AI to write assessments could lead to gaps in students’ learning, respondents said.
“There is a real need to reimagine education and assessment so that we're ensuring that students who come through the disciplines actually have that disciplinary knowledge,” says Gribble. “It's not about just being able to produce a piece of written work, it’s about being able to engage in the conversation.”
“We need people to have strong moral compasses, and understand the ethical implications” as they collaborate with AI, says Gribble. “But the future is bright, because if I can have a human and a robot operate on me, it will always be better than one or the other.”
