AI for Research and Learning

AI can find, summarise and explain faster than any library search — and invent a convincing reference just as fast. Learn to get grounded answers, check them properly, and use AI as a tutor that makes you learn rather than a crutch that lets you forget.

In 2023, researchers asked ChatGPT to write short literature reviews on 42 topics and then checked every one of the 636 references it produced. With the older GPT-3.5 model, more than half of the references did not exist; with GPT-4, about one in five were invented. Each looked perfectly plausible — real-sounding authors, a sensible journal, a year, page numbers. Tools have improved a great deal since then, especially those that search the web as they answer. But the lesson has not expired: a citation is a claim, and claims need checking.

From answers to grounded answers

When an assistant answers from memory alone, it is reconstructing what its training text tended to say — which is why it can be out of date or simply wrong, a hallucination. A grounded answer is different: the tool first retrieves sources (a web search, your uploaded documents, a database) and then answers from them, usually with links. Grounding does not make an answer true, but it makes it checkable — and that is the whole game in research.

Grounded answers come from AI search engines, from general assistants with web search switched on, and from tools that answer only from documents you supply (see Choosing the Right Tool). Whatever you use, ask for sources explicitly and say what kind you want.

Ungrounded
What are the health effects of intermittent fasting?
Grounded
Search for evidence on the health effects of intermittent fasting in adults. Prefer systematic reviews, meta-analyses and guidance from national health bodies published in the last five years. For each main finding, give the source with a link, the type of study, and how strong the evidence is. Clearly separate what is well established from what is uncertain or disputed. If you cannot find a source for a point, say so instead of answering from memory.

Checking citations: do they exist, and do they say it?

There are two separate failures to look for. The first — a reference that does not exist — is now rarer in tools that search, but still happens. The second is subtler and more common: a real source that does not say what the answer claims. The link is genuine, the page is reputable, and the claim is not on it, or is stated with conditions the summary dropped. Work through these checks for anything that matters.

  1. Does it exist?Open the link, or search the exact title in Google Scholar, a library catalogue or the publisher’s site. No trace means do not use it.
  2. Is it what it says it is?Check the authors, year and publication match. A real paper with the wrong year or journal is a warning sign that the rest may be shaky too.
  3. Does it say what is claimed?Find the passage. Use your browser’s find function on a key term. Does the source support the claim fully, partly, or not at all?
  4. Is the context intact?Look for the conditions: “in mice”, “in a small pilot study”, “among patients who…”. Summaries love to drop these.
  5. Is it a good source?A press release, a blog or a single study is weaker than a systematic review or an official guideline. Check the date too.
Make checking easier
For each claim in your answer, add the exact sentence from the source that supports it, in quotation marks, with the link. If no single sentence supports a claim, mark the claim [UNSUPPORTED].

Then actually open two or three of those links and confirm the quotes are there. Asking the model to quote does not guarantee the quotes are accurate — but a missing quote is quick to spot.

Deep research features

Since late 2024, several assistants — including ChatGPT, Gemini, Claude and Perplexity — have added deep research modes As of 2026. Instead of one quick search, the tool plans a series of searches, reads dozens of pages over several minutes, and writes a long, cited report. Names, limits and availability on free plans vary and change often.

These are genuinely useful for getting the lie of the land on an unfamiliar topic: what the main positions are, who the key sources are, what has been said recently. They are an agent-style feature — the model takes many steps on its own — which brings agent-style risks:

Scoping a deep research request
Research the current evidence on four-day working weeks for office-based organisations. Scope: trials and studies from 2019 onwards, any country, with particular attention to South-East Asia if available. Prioritise peer-reviewed studies, government reports and large structured trials over news articles and opinion. Structure: (1) what was tested and where, (2) outcomes measured, (3) results, (4) limitations and criticisms, (5) open questions. Flag any claim that rests on a single source.

Treat the report as a reading list with commentary. Pick the three or four sources it relies on most and read those yourself.

Read laterally

Professional fact-checkers have a habit that students and even academics often lack. In a well-known study, Stanford researchers watched people evaluate unfamiliar websites. Students and historians tended to read down the page, judging it by its look and its own claims. Fact-checkers read across: within seconds they opened new tabs to see what others said about the source. They reached better judgements, faster. This is lateral reading.

Mike Caulfield’s SIFT method packages it into four moves: Stop; Investigate the source; Find better coverage; Trace claims, quotes and media to the original context. It works just as well on AI answers as on web pages. Checking AI’s Work walks through each move. When an AI tells you something surprising, do not ask the same AI whether it is sure; open a new tab and see whether independent, reliable sources say the same.

AI can help with the lateral moves too:

Investigate a source
I found a claim on a website called [NAME]. Search for independent information about who runs this site, how it is funded, and what reputable organisations or journalists have said about its reliability. Give me links I can check, and do not rely on the site’s own “About” page.

Using AI to learn, not just to look up

Consensus Decades of learning research agree on a few things: we remember more when we try to recall information than when we re-read it (the testing effect, or retrieval practice), and practice spread over days beats cramming (the spacing effect). An AI assistant that simply hands you answers skips the effort that makes learning stick. Set it up to make you work instead.

You explain the idea in your own words; the AI finds the gaps. Explaining exposes what you only thought you understood.

Explain-it-back prompt
I am learning about compound interest. I will explain it in my own words. Do not correct me straight away: first tell me which parts of my explanation are right, then ask me one question that targets the weakest part. Repeat until I can explain it fully.

My explanation: [write yours]

One caution: an AI tutor can explain a concept wrongly with total confidence. Keep your textbook or lecture notes as the reference, and when an explanation conflicts with them, check before you believe the AI.

Limits: very recent and niche topics

Two kinds of question remain hard. Very recent events: a model’s built-in knowledge stops at its knowledge cutoff, and even with search, the first reports of breaking news are often incomplete or wrong. Niche topics: a small local regulation, a specialist technique, a little-studied species, a Malaysian state-level policy. Where little has been written, the model has little to go on and is more likely to fill the gap with plausible invention, sometimes borrowing details from a similar but different case — another country’s law, another company’s policy.

For both, go to primary sources: official government portals and gazettes, the organisation’s own announcements, specialist databases, librarians, and people who actually know. Use AI to help you find and read those sources, not to replace them.

Quick check

An AI answer cites a real, reputable website. What else must you check?

Real sources are often misquoted or stripped of important conditions.

What is lateral reading?

Fact-checkers judge a source by what others say about it, not by its own claims.

Which study habit is best supported by learning research?

Retrieval practice plus spacing beats re-reading and cramming.

When is an AI answer most likely to be invented, even with web search?

Where sources are scarce, models fill gaps with plausible-sounding details.

// Pause and reflect

  1. Think of the last fact you took from an AI answer and used. Did you check it? How would you have checked it in under five minutes?
  2. What are you trying to learn at the moment? Which of the four study prompts would suit it best?
  3. Which topics in your work are “niche” enough that you should go straight to primary sources?

Your notes are saved privately in this browser.

Try this

Ask an AI search tool a factual question in your field that you can answer well yourself. Take the three citations it relies on most and run the five-step check on each: exists, correct details, says it, context intact, good source. Score each 0–5. Then take a topic you are learning and do one round of the quiz-me prompt today and another in three days.

Key terms on this page

Sources & further reading