Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

Deep Research

nounEmerging AI Researchalso AI deep research agents, agentic research, AI research assistants

In one line

Deep research is an AI capability in which an agent plans a research task, searches and reads many sources over several minutes, reasons about what it finds and produces a structured, cited report, automating work that a human analyst might spend hours on.

Deep research is a feature of modern AI agents that carries out multi-step investigations on the web and in documents. Rather than answering from memory in one pass, the system breaks a question into sub-questions, searches, reads pages and files, compares what it finds, notices gaps and writes a report with citations. OpenAI introduced a deep research capability in ChatGPT in February 2025, describing an agent that finds, analyzes and synthesizes hundreds of online sources and takes from about five to thirty minutes on a task [1]. Google had introduced a similar mode in Gemini in December 2024, and other developers have followed [2]. The approach relies on reasoning models, test-time compute, retrieval-augmented generation and context engineering.

How deep research works

### Planning

The agent interprets the request, may ask clarifying questions, and drafts a plan of sub-topics and search strategies. Some systems show the plan so the user can edit it.

### Searching and browsing

The agent issues many queries, opens pages, reads documents such as PDFs and sometimes uses tools to analyze data or run code. It follows leads and revises its queries as it learns more, which separates it from a single retrieval step.

### Reasoning and verification

A reasoning model weighs evidence, notes contradictions and decides when it has enough. Better systems compare sources, prefer primary documents and flag uncertainty. Extended thinking at this stage is where test-time compute is spent.

### Managing context and memory

Because research generates far more text than a model can hold at once, the system must summarize, store notes and retrieve them later; see memory in AI agents and context engineering. Some designs use a lead agent that delegates sub-tasks to several parallel sub-agents and merges their findings.

### Writing the report

The agent drafts a structured report with headings, summaries, tables and citations, often in minutes. Users can ask follow-up questions or request revisions.

### Training

Developers train these systems with reinforcement learning on browsing and research tasks, rewarding accurate, well-supported answers; see reinforcement learning.

Uses

  • Market, competitor and policy research that would take an analyst hours.
  • Literature reviews in science and medicine; see AI in scientific research.
  • Due diligence and legal research, with careful human verification; see AI in law.
  • Journalism background research; see AI in journalism.
  • Education, where students learn to read and evaluate sources; see AI in education.
  • Consumer decisions such as comparing products or planning travel.
  • Enterprise knowledge work, searching internal documents and the web together.

Evaluating deep research

Evaluating long reports is hard, because quality involves accuracy, coverage, source quality, reasoning and clarity. Researchers use expert-level question sets such as Humanity’s Last Exam, introduced in 2025 with questions written by thousands of experts [3], and browsing benchmarks that require locating hard-to-find facts, as well as human review of reports. Reliable evaluation checks whether each claim is supported by the cited source, whether important perspectives are missing and whether the system handles conflicting information; see AI evaluation.

Limitations and risks

  • Hallucination and unsupported claims. Reports can contain wrong statements or citations that do not support the text; see hallucination.
  • Source quality. Agents may rely on low-quality, outdated or biased web pages, or on search-optimized content.
  • Missing paywalled and private material. Important sources may be inaccessible.
  • Confident tone. Polished writing can hide uncertainty and weak evidence.
  • Prompt injection. Web pages can contain hidden instructions that try to manipulate the agent; see AI cybersecurity.
  • Privacy and confidentiality. Queries and documents sent to services may be sensitive; see AI privacy.
  • Copyright and publisher impact. Reports may summarize paid content and reduce traffic to publishers; see AI copyright.
  • Skill erosion. Over-reliance can weaken the habit of reading sources and thinking critically.
  • Cost and energy. Long agent runs use significant computing resources.

Good practice

Users get better results by stating the goal, audience, scope, date range and preferred source types, asking for primary sources, requesting a list of open questions and checking key claims against the cited pages. For important decisions, a human expert should review the findings; see human oversight and AI literacy.

A worked example

A city official asks a research agent for a briefing on how other cities have regulated electric scooters. The agent proposes a plan covering safety, parking, insurance and enforcement, and the official adds a request to focus on cities of similar size. Over twelve minutes, the agent searches government websites, news reports and academic studies, compares rules, and notes that figures on injuries differ because cities define incidents differently. It produces a ten-page report with a summary table and links. The official checks five key claims against the cited sources, finds one outdated ordinance and corrects it, and then uses the report as a starting point for discussions with the transportation team.

Research agents compared with search engines

A search engine returns a ranked list of pages and leaves reading and synthesis to the person. A deep research agent does the reading and synthesis, which saves time but moves judgment into the system. That shift makes transparency important. Good tools show which sources were opened, which were ignored and where evidence conflicts, and let users inspect the trail of searches. Users who treat the report as a map rather than the territory, and who follow citations into the original documents, gain speed without giving up verification; see explainable AI.

Impact on research skills and publishing

Libraries, schools and newsrooms are adapting. Instructors teach students to formulate questions, evaluate sources and cross-check AI output, while publishers and standards bodies discuss attribution, licensing and the effect on web traffic. Researchers propose practices such as recording prompts and dates, keeping the original source list and disclosing AI assistance in published work, so that others can reproduce and audit findings; see AI governance.

Further reading and next steps

Readers can continue with AI scientific discovery, AI harness and human-AI collaboration to see how agents are built and used in practice.

FAQs

What is deep research in AI? It is an agent mode that plans, searches many sources, reasons about them and writes a cited report on a topic.

How is it different from a normal chatbot answer? It works over several minutes using many searches and steps, instead of replying from memory in one pass.

Can I trust a deep research report? It is a strong starting point, but claims and citations should be checked, especially for important decisions.

Does deep research replace human researchers? It speeds up gathering and drafting, but judgment, verification and original analysis still need people.

What are the main risks? Unsupported claims, weak sources, prompt injection, privacy concerns and over-reliance.

Key facts

  • OpenAI introduced a deep research capability in ChatGPT in February 2025, describing an agent that finds, analyzes and synthesizes hundreds of online sources to produce a research report, with tasks taking from about five to thirty minutes1.
  • Google introduced Deep Research in Gemini in December 2024, and other providers including Anthropic, Perplexity and xAI have since offered similar agentic research modes2.
  • Benchmarks such as Humanity's Last Exam, introduced in 2025 with questions from thousands of experts, are used to test whether AI systems can answer difficult expert-level questions3.

Related concepts

Quick checkWhat is deep research in AI?Show answer

Deep research is an AI capability in which an agent plans a research task, searches and reads many sources over several minutes, reasons about what it finds and produces a structured, cited report, automating work that a human analyst might spend hours on.

Sources

  1. OpenAI, Introducing deep research, https://openai.com/index/introducing-deep-research/
  2. Google, Gemini Deep Research, https://gemini.google/overview/deep-research/
  3. Center for AI Safety and Scale AI, Humanity's Last Exam, https://lastexam.ai/

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue