Skip to content

Research properly with AI

Use AI to speed up research without letting it become your source. A five-step loop, worked through by tracing a real headline statistic back to its original study.

Level
Beginner
Reading time
13 minutes
Sources checked
11 Nov 2026
Status
Draft, awaiting editorial review

Who it’s forAnyone who uses AI to find things out for work, study or a decision, and wants answers they can defend.

Before you startNone. Check an AI answer before you rely on it covers the checking side in more depth.

By the end you'll be able to

  • Know which research jobs AI is good at, and which it can't do for you
  • Use a five-step loop that keeps every claim tied to a source you've read
  • Trace a statistic back to where it started, and notice when it has changed on the way

AI is a research assistant, not a source

AI tools are fast at the parts of research that used to be slow: getting oriented in a new topic, suggesting search terms, summarising a long document you've given them. They're unreliable at the part that matters most, which is telling you exactly where a fact came from.

Columbia Journalism Review's Tow Center tested this directly in 2025. "We ran sixteen hundred queries (twenty publishers times ten articles times eight chatbots) in total," the researchers wrote, asking each AI search tool to identify the original news article an excerpt came from. "Collectively, they provided incorrect answers to more than 60 percent of queries."

So the rule for this lesson is simple. AI can help you find and read sources faster. It can't be the source. Anything you'd put your name to needs to trace back to something you've opened and read yourself.

Sort it

Which jobs suit AI?

Sort each research job into where it belongs.

  1. 1.Suggest search terms and related ideas for a topic you're new to
  2. 2.Summarise a report you've pasted in
  3. 3.Give you a statistic with a citation for a slide
  4. 4.Decide whether a source is trustworthy enough to rely on
  5. 5.List counter-arguments to a position you hold
Choose a category for every item to check.

The five-step research loop

  1. 11. Frame the questionWrite down what you need to know, why, and what would change your mind. A clear question stops you collecting interesting facts that don't help.
  2. 22. Orient with AIAsk for key terms, the main positions, and the kinds of sources experts would use. Treat every name, number and title in the answer as a lead to check, not a fact.
  3. 33. Find sources yourselfSearch for the actual sources: official pages, the original study, the primary data. For an academic citation, Crossref lets you "find a particular work by searching for a full citation" or by its DOI.
  4. 44. Read and traceOpen each source and find the claim in it. If the source is reporting someone else's work, follow it back to the original. This is where most errors show up.
  5. 55. Record what you foundFor each claim, log the source, the exact wording, and how confident you are. A source log turns research into something you can defend and someone else can check.

Step 4 borrows from Mike Caulfield's SIFT method for checking information online: "Stop, Investigate the source, find better coverage, trace the original context." It works just as well on an AI answer as on a social media post.

Knowledge check

Quick check: tracing

An AI tool tells you a statistic and links a news article. You open the article and the statistic is there. What's the next step?
Need a hint?

Ask yourself: is this article the place the number was first published?

Worked example: tracing a headline statistic

Here's a real trace, using the study from the start of this lesson. Suppose an AI tool tells you: "AI search engines fail to produce accurate citations in over 60% of tests." That wording is real; it's the headline Nieman Lab used on 10 March 2025. It's a good source, but it's reporting someone else's work, so it isn't the end of the trail.

What the trace found (checked 11 November 2026)
QuestionNieman Lab headlineThe original CJR report
What was measured?"accurate citations"Whether each tool correctly identified the original news article for an excerpt
How much?"over 60% of tests""incorrect answers to more than 60 percent of queries", collectively
How big was the test?Not in the headline1,600 queries: 20 publishers, 10 articles each, 8 chatbots
Who did it?Tow CenterKlaudia Jaźwińska and Aisvarya Chandrasekar, Tow Center, 6 March 2025

The headline is fair. But the original tells you something the headline can't: the test was about finding the source of news excerpts. It doesn't tell you how often AI tools get other kinds of question wrong, and it's a combined figure across eight tools, not a score for any one of them. If you were writing "AI search is wrong 60% of the time" in a report, the trace would stop you.

Scenario

The statistic for the slide

You're preparing a slide for a team meeting tomorrow. An AI tool gives you a strong statistic about customer behaviour, with a link to a marketing blog that mentions it. The blog doesn't link to any study.

What do you do?

Prompts that keep AI in its lane

Orienting prompt: get leads, not answers
I'm researching [topic] so I can [decision or output].

Give me:
1. The key terms and names I should search for.
2. The main positions or schools of thought, in one line each.
3. The kinds of primary source an expert would check (e.g. official statistics, the original study, regulator guidance).

Don't give me statistics or quotes. I'll find those in the sources myself.
Reading prompt: work only from a source you've given it
Below is [document]. Using only this document:
1. List each claim it makes about [question], with the exact sentence it comes from.
2. Say where the document is reporting someone else's work, and name that original source if it does.
3. If the document doesn't answer something, say so. Don't fill gaps from your own knowledge.

[paste document]

Try it

Guided practice

Trace one claim to its original

About 20 minutes0 of 5 steps ticked

You’re done when

  • Every claim you'd use has a source you've opened and read
  • At least one claim has been traced past a summary to its original
  • Anything you couldn't trace is marked as such or dropped

What you’ll get from it: A short, defensible answer to a real question, and a source log someone else could check.

Free worksheet

The source log

Keep every research claim tied to a source you've opened: the claim, where it came from, the exact wording and how sure you are.

The online version saves your answers only in this browser. The PDF is A4 and prints on US Letter too.

What to remember

  • AI is good at widening your view and reading what you give it. It's unreliable at saying where a fact came from.
  • A real source that reports someone else's work isn't the end of the trail. Trace to the original.
  • The CJR study tested news citation. It's strong evidence for checking citations, not a general accuracy score for AI.
  • Don't paste confidential material into a research tool your organisation hasn't approved.

Tools for this guide

  • Gemini Notebook (NotebookLM): A research notebook that answers questions using the documents you add, with citations back to the passages it used. (Researched from docs)

Your progress

You’ve finished 0 of 4 activities in this lesson.

When you've tried the practice task on your own work, mark the lesson complete. It's your call: finishing the activities shows you've practised, not that you've mastered the skill.

Progress is saved only in this browser, on this device. There’s no account, and nothing is sent to us or anyone else. Clearing your browser data, or using another browser or device, starts you afresh. We save which activities you finished, never what you type.

Suggested next lesson

Make AI useful with documents and dataResearch often ends in a pile of PDFs and spreadsheets. Next, get reliable answers out of your own documents.Intermediate, 14 minutes

Next steps

Sources and evidence

The research loop is our editorial advice, built on the sources below. The worked example is real: we traced the statistic on the date shown, using the two published articles linked. We haven't measured how much time the loop saves.

  1. Columbia Journalism Review (Tow Center): AI Search Has a Citation Problem, 6 Mar 2025 (opens in a new tab)Checked 11 Nov 2026
  2. Nieman Lab: AI search engines fail to produce accurate citations in over 60% of tests, 10 Mar 2025 (opens in a new tab)Checked 11 Nov 2026
  3. Mike Caulfield: SIFT (The Four Moves), 19 Jun 2019 (opens in a new tab)Checked 11 Nov 2026
  4. Crossref Metadata Search: help (opens in a new tab)Checked 11 Nov 2026

This guide is a draft and hasn’t been through editorial review yet. How we review guides

One useful email a week.

Important model releases, a tool worth trying, one practical workflow and a few useful links, each with its source. Free, and easy to leave.

Signups aren’t open yet.

We won’t collect any email addresses until the newsletter launches, and there’s no form here to fill in. Until then, new guides and videos go out on our channels:

Questions in the meantime? understandingaihq@gmail.com