As LLM-powered agents become more widespread, academic research is undergoing a fundamental transformation in how it is conducted. Simply put, large language models and multimodal generative AI applications can be used to assist in the entire research process including:

• Brainstorm research topics
• Identify research gaps and emerging trends
• Formulate hypotheses and design experiments
• Extract and synthesize data
• Conduct literature reviews, systematic and scoping reviews, and meta-analysis
• Evaluate and mitigate biases, inaccuracies, and inconsistencies in literature review and analysis
• Code, analyze, and visualize qualitative and quantitative data
• Develop disciplinary and interdisciplinary methodological and theoretical frameworks
• Recommend sources and citations
• Summarize text and write abstracts
• Assist in research dissemination, presentation, knowledge translation, and mobilization

Evidence of this paradigm shift in research can be found in a recent study at Stanford titled  Mapping the Increasing Use of LLMs in Scientific Papers. In an analysis of 950,965 papers published between Jan 2020 – Feb 2024 on the arXiv, bioRxiv, and Nature portfolio journals, the study found that 17.5% of computer science papers and 16.9% of peer review text had at least some content drafted by AI.

Deep Research: A New Paradigm in Research

While keeping pace with the new developments in generative AI and LLM reasoning models is in itself a challenge, new “deep research” agents are beginning to surface at a remarkably rapid rate. Briefly defined, deep research in the context of generative AI is the use of LLMs, Retrieval Augmented Generation (RAG), and reasoning models to conduct in-depth, multi-step, comprehensive, and detailed extraction, analysis, and synthesis of scholarly information, empirical data, and perspectives from a broad range of sources to generate a final report with citations. Similar to the rise of numerous search engines in the 1990s, we are witnessing a growing competition in the “deep research” environment with various LLMs striving to gain popularity and acceptance by researchers, scholars, students, policy analysts, and the general public. In just a short span of four months (December to February), Google Gemini, Perplexity.ai, OpenAI, and Grok introduced their “deep research” platforms as agentic research options resembling an advanced option for users to experiment with. Similar to chatbots, deep research tools are offered as free (limited) and subscription-based models. OpenAI, for instance, offers its deep research option on a subscription basis, whereas Perplexity.ai provides limited free access to its deep research tool. The introduction of DeepSeek, the Chinese open-source large language model was a major disruptor, reminding Silicon Valley AI developers what could be imagined if they gradually made their models open source to stimulate innovation. An example of the impact of this disruption can be found in Google Gemini’s announcement of a freely accessible Deep Research tool on March 13, 2025.

Considering that the “deep research” tools are built on different LLMs, they share some similarities and also offer model-specific functionalities. For instance, the four deep research tools named above have the capability to break down a submitted research topic/question prompt into specific facets and questions and allow a user to narrow down and focus on a particular question or research direction. They all list the scholarly sources and publicly available web and social media information they use to conduct deep research. As one of the early users of Perplexity.ai, I should particularly give them credit for committing to transparency, allowing users to view the information sources and citations that they use to respond to prompts. Some of the named deep research platforms provide an interesting functionality to modify or exclude less credible, non-academic information from the list of sources with the opportunity to focus only on reliable and selected sources. This approach is very much based on decades of information search behaviour studies and interactive information retrieval research where a user is provided with opportunities to provide a search statement, reformulate, expand, modify, or change their original query terms to achieve search task goals. In almost all of the four deep research tools, the user is presented with information on the different steps of the research process, namely searching, reading, reasoning, analyzing, synthesizing, wrapping up, and producing a final report. In a couple of the deep research tools, the user is shown the time it takes to go through each stage, the waiting time, and a progress bar.

They also offer different and unique user experiences and functions in terms of the number of sources they use for research in free and paid versions, whether or not the user has control over the different research steps to modify the initial prompt, and whether or not they can add or remove sources from the provided list, as well as how many research topics and prompts they can submit to the system. One of the major drawbacks of deep research tools—commonly criticized in large language models—is the lack of consistency in their outputs when addressing the same research question over time.

Research Integrity and Transparency

While deep research platforms provide more transparency and clarity in the use of information sources and in demonstrating the research process, the ways in which they conceptualize research, or deep research for that matter, are different. These differences impact how a research topic is understood, explored, analyzed, investigated, and presented. One may argue that the rationale for the emergence of “deep research” tools is largely due to the common and widely cited criticisms of LLMs as platforms that hallucinate and fabricate data and citations and produce inaccurate, biased, discriminatory, and often inconsistent and incomplete outputs. It is evident that these criticisms along with the increasing emphasis on key human-centred AI principles such as transparency, explainability, reliability, trustworthiness, and safety are gradually giving rise to more accountable and transparent LLM-based research agents.

Although “deep research” platforms offer an opportunity to re-imagine and re-envision how research could be carried out, the same significant concerns and challenges associated with the ethical and responsible use of LLMs hold true for deep research tools. Examples of these issues may include misrepresentation of data and authorship, difficulty in replication of research results, data and algorithmic biases and inaccuracies, user and data privacy and confidentiality, quality of outputs, data and citation fabrication, and copyright and intellectual property infringement. Dan Russell points out three big issues with deep research platforms, namely relevance of the produced report as sometimes the texts in the reports are tangential or off-topic, getting a silly answer to a silly research question, and not being careful to hand your cognitive work off to an AI bot. I would add the fundamental issues of repeatability, reproducibility, and output consistency in conducting research using deep research platforms.