Aiodic

Back to the table

53 Intermediate Re Research Applications

Research

Gathering and synthesizing information.

What it is

Research, as an AI application, means asking an AI tool to do the legwork of finding things out: work out what to look for, search for it, read what comes back, compare the sources, and write up what it found with links you can follow. You give it a question such as "What did Victorian households actually eat for breakfast?" or "Which print-on-demand services sell well to schools?", and a few minutes later you get a report.

Every chatbot with a web search button can do a little of this. Ask a question, it runs a handful of searches, and it answers with links. What changed from late 2024 onwards is a bigger version, usually called "deep research". Google, OpenAI, Perplexity, Anthropic and others now offer a mode that takes a single question and works on it for several minutes, sometimes much longer, running dozens of searches and reading many pages before writing a long, structured report with citations. Names vary: Deep Research, Research, Deep Search. The idea is the same.

A deep research tool is an agent, meaning an AI that works through a task in steps, deciding what to do next from what it has just learned, instead of answering in one go. A typical run goes through five stages:

Under the hood, some of these tools split the work between several AI workers. In June 2025 Anthropic described the system behind Claude's research feature: a lead agent writes the plan and sends subagents off to search different parts of the question in parallel, and a separate step matches the finished report's claims to places in the sources. The Context Engineering article covers why that split helps. For you as a user, the important part is simpler: a research tool reads far more than you could in the same time, and then tells you about it in its own words.

That last phrase is the catch. A research report is a summary of sources the tool chose itself, written by a model that can misread, blend, or invent. The report looks finished. It isn't. It's a well-organised first draft of your research, with a reading list attached.

There are also specialised research assistants for academic work. Tools such as Elicit and Consensus search databases of published scientific papers rather than the open web, and help you find, summarise and compare studies. They are useful for anyone writing non-fiction that leans on science, but the same rule applies: the summary of a paper is not the paper.

Why it matters

It collapses the slowest part of research. For an author or creator, the expensive part of research isn't thinking. It's finding: twenty searches, forty open tabs, and an afternoon deciding which pages are worth reading. A research tool does that first sweep in minutes. For a novelist checking what a 1920s telephone exchange looked like, or a course creator mapping what competitors teach, that's a real gift.

It gives you a map before you start. A good research report shows you the shape of a topic: the main schools of thought, the key names, the arguments, the obvious gaps. Even if you then read every source yourself, you read them in a better order, knowing what to look for.

The output looks more trustworthy than it is. A long report with headings, footnotes and forty links feels like expert work. That appearance is exactly the danger. The tool can cite a weak source as if it were strong, attach a real link to a claim the page doesn't make, or invent a source outright. The examples below include two official reports, one from a government and one from a major consultancy, that went out with citations to things that don't exist.

It favours what's easy to reach. Research tools can mostly read only what's freely available on the web. Books, paywalled journals, newspaper archives, and specialist databases are often invisible to them. Anthropic reported that its own early research agents kept choosing search-optimised content farms over more authoritative sources such as academic PDFs, and it had to teach the system to judge sources better. If the best evidence on your topic sits in a library, the report may quietly leave it out and still sound complete.

Gaps come out sounding like answers. When a tool can't find something, it doesn't always say so. It may fill the gap with a plausible guess, which is a hallucination in research clothing. Worse, a confident sentence about "what studies show" can rest on one blog post.

Your name goes on the result. If you publish a book, a course, or a client report that repeats an error from an AI research tool, the error is yours. Readers and clients don't care which tool you used. That's why research with AI is a two-part job: the tool gathers, and you verify.

Real-world examples

Google introduces Deep Research in Gemini (11 December 2024)

On 11 December 2024, Google announced Deep Research for Gemini Advanced, its paid subscription. Google's product post described the pattern that most tools have since followed. You ask a question; Gemini writes a multi-step research plan, which you can revise or approve; it then browses the web, searching, reading, and starting new searches based on what it learns; and finally it produces a report of key findings with links to the original sources, which you can export to a Google Doc. At launch it was available on the web in English, with a wider rollout promised for 2025.

What made this launch matter was the plan step. Showing the plan before the research starts gives you a chance to steer: add a sub-question, cut one that doesn't matter, or insist on a type of source. It's the single most useful control these tools offer, and it's the one people skip most often.

OpenAI and Perplexity follow within weeks (February 2025)

On 2 February 2025, OpenAI launched its own deep research feature in ChatGPT, first for subscribers to its most expensive Pro plan. OpenAI said it used a version of its o3 reasoning model, and that a single task could take from 5 to 30 minutes. It reported that deep research scored 26.6 percent on Humanity's Last Exam, a benchmark of very hard expert-level questions across many subjects, far ahead of the models it compared against. That figure is OpenAI's own. OpenAI was also candid about limits: it said the tool could hallucinate facts or make incorrect inferences, might struggle to tell authoritative information from rumour, and often failed to convey uncertainty accurately.

In mid-February 2025, Perplexity released its own Deep Research mode, free to try with a daily limit and with more use for paying subscribers. Perplexity said most tasks finished in under three minutes and reported a score of 21.1 percent on the same benchmark.

Within about two months, three companies had shipped the same idea, and others followed. Access, limits, and the models behind these tools have changed several times since, so check each tool's current terms. The lasting lesson is in OpenAI's own caveats. The company that built the tool warned that it might confuse rumour with fact and sound surer than it should. Those are exactly the failures a researcher has to check for.

The government report that cited studies that didn't exist (May 2025)

On 22 May 2025, the Make America Healthy Again Commission, chaired by US Health Secretary Robert F. Kennedy Jr., released a report on children's health that cited more than 500 studies and other sources. On 29 May, the nonprofit news outlet NOTUS reported that at least seven of the cited sources did not appear to exist. One citation listed the Columbia University epidemiologist Katherine Keyes as first author of a paper on adolescent anxiety. She told NOTUS it was not a real paper she or her colleagues had been involved with. NOTUS also found broken links and real studies whose conclusions had been misstated.

The Washington Post later reported that some references contained "oaicite", a marker that experts linked to citations generated by OpenAI's tools. The White House called the problems formatting issues, and an updated version replaced the missing citations.

For authors, the pattern is the thing to remember. The fake studies had plausible titles, real journals, and real researchers' names, the classic fingerprint of an invented citation. A reviewer who had opened each reference would have caught them in minutes.

Deloitte refunds part of a government contract (July to October 2025)

In July 2025, Australia's Department of Employment and Workplace Relations published an independent review by Deloitte Australia of the IT system it uses to apply penalties to welfare recipients. The contract was worth about 440,000 Australian dollars. Christopher Rudge, a University of Sydney academic, spotted problems: references to reports that didn't exist, attributed to real academics, and a made-up quote attributed to a Federal Court judgment.

A corrected version was published in late September 2025, removing the fabricated references and quote and adding a disclosure that a generative AI tool chain based on Azure OpenAI's GPT-4o had been used. Deloitte agreed to repay the final instalment of the contract, which was later reported as more than 97,000 Australian dollars. The department said the substance of the review and its recommendations were unchanged.

The Deloitte case is the business version of the MAHA report. A professional firm, paid for expert work, used a generative AI tool and then let references through that nobody had properly checked. The cost was money and reputation, and it landed on the firm, not the tool.

An everyday scenario: the historical novelist

Picture Priya, who is writing a novel set in a Scottish fishing town in 1912. She asks a deep research tool for a report on herring fishing, women's work gutting fish, and daily life in the town that year. Ten minutes later she has 3,000 words, twenty-six citations, and a vivid detail about the wages paid to herring girls.

She runs a source audit before she writes a word. Most links are fine: a museum site, a local history society, a digitised newspaper. Two are travel blogs repeating each other. The wage figure, the detail she liked best, is attached to a page that mentions herring girls but gives no number at all. She asks the tool where the number came from, and it can't find it. The best source on her topic, a book of oral histories, never appeared in the report, because the tool couldn't read it. She orders the book. The report saved her a week of searching, but it didn't do her research. It told her where to start.

Try it yourself

You'll need a chatbot with either a deep research mode or ordinary web search. At the time of writing, ChatGPT, Gemini, Claude, and Perplexity all offer some form of research mode, several with a small free allowance, but names, limits, and plans change often, so check each tool's current help pages. Allow about forty minutes, including the check at the end, which is the most important part.

  1. Choose a real question. Pick something you actually need for a project, where you know enough to spot nonsense. A good research question is specific: not "the history of coffee" but "how coffee houses were used for business in 1700s London".
  2. Write a research brief. Research tools do much better with a clear brief than a bare question. Copy this template, fill in the brackets, and keep it for reuse:
Research question: [your question]

Why I need it: [e.g. background for a chapter of a non-fiction book aimed at general readers]

What I want back:
- A report of about [1,500] words with headings.
- Every factual claim followed by a link to the source it came from.
- A short list of the 5 most important sources, with who published each one and when.
- A section called "Gaps and disagreements" listing what you could not find, where sources disagree, and anything that rests on a single source.

Source rules:
- Prefer primary and authoritative sources: original documents, academic papers, museums, archives, official sites and established publishers.
- Avoid content farms, AI-written listicles and uncredited blogs, or flag them clearly if you use them.
- Tell me about important sources you could not read, such as paywalled articles or books.
- If you are unsure of something, say so. Do not fill gaps with guesses.
  1. Run it and read the plan. In a deep research mode, the tool may show you a plan or ask clarifying questions first. Read the plan properly. Add anything missing and cut anything off-topic before you approve it. If you only have ordinary chat with search, switch search on, paste the brief, and add: "Search as many times as you need before answering."
  2. Read the report once, quickly. Don't take notes yet. Look at the "Gaps and disagreements" section first, because that's where an honest tool tells you what it couldn't do. If it's empty, be suspicious.
  3. Audit five citations. Pick the five claims you'd most like to use: the surprising numbers, the quotes, the specific dates. For each one, open the link and fill in this checklist:
Claim: [copy the sentence from the report]
Link opens? yes / no
Who published it? [organisation or author]
Kind of source: primary / reputable secondary / weak / unknown
Does the page actually say this? yes / partly / no
Exact supporting sentence: [paste it, or write "not found"]

Use your browser's find function (Ctrl+F, or Cmd+F on a Mac) to search each page for a key number or name. The question that matters most is "Does the page actually say this?" A working link attached to a claim the page doesn't make is the commonest error, and the easiest to miss.

  1. Send back what failed. For any claim that didn't check out, paste this into the same chat:
I checked these citations and they do not support the claims:
[list the claims and links]

For each one, either quote the exact sentence from a source that supports it, with the link, or tell me plainly that you cannot find support and remove the claim from the report.
  1. Look for what's missing. Ask: "What kinds of sources would an expert use for this question that you could not access? Name specific books, archives, or databases if you know of them, and tell me which of those names I should double-check." Then check that any books it names actually exist before you order them.
  2. Score the run. Out of your five citations, how many fully supported their claims? Write the number down. Do this every few projects and you'll learn how far to trust each tool on each kind of question, which is worth more than any single report.

Common mistakes and misunderstandings

Treating the report as the research. A research report is a draft and a reading list. It tells you where to look and what might be true. Anything you plan to publish, teach, or send to a client needs to come from a source you've read yourself.

Trusting a citation because it's there. A footnote looks like proof. The MAHA and Deloitte reports both had tidy footnotes. Open the link, find the sentence, and only then count it as supported. The Grounding article explains why a real link attached to an unsupported claim is so common.

Assuming more sources means better sources. Forty citations can all trace back to the same two articles, copied and rewritten across the web. Check who published each source, and look for the original, not the copy.

Believing the web holds everything. Much of the best evidence sits in books, paywalled journals, newspaper archives, and specialist collections. A research tool that can't read them may produce a confident report that simply leaves them out. For serious work, pair the tool with a library catalogue and a librarian.

Skipping the plan. The plan step is where you steer. Ignore it and the tool may spend its effort on the wrong sub-questions or the wrong kinds of source.

Asking vague questions. "Tell me about the Victorian era" produces a vague report. A specific question with a stated purpose, audience, and source rules produces a useful one. The research brief above is a prompt template you can reuse for every project.

Taking quotes on trust. Quotations are the riskiest thing in any AI report, because a model can produce a sentence that sounds exactly like something a person would say. Never publish a quote from a research report without finding it in the original source. The Deloitte report's fabricated court quote is the warning.

Letting the tool's confidence set yours. Research tools rarely say "I'm not sure". As OpenAI itself said of its own tool, they often fail to convey uncertainty accurately. Ask for gaps and disagreements explicitly, and treat anything resting on a single source as provisional.

Confusing research with analysis. A research tool gathers and summarises what others have said. Working out what your own numbers mean is a different job; the Data Analysis article covers that.

Leaving no human check. For anything that goes into print, into a client's hands, or into a decision about money or health, keep a human in the loop who checks sources, not just reads the summary.

Related elements