How to research with AI without losing rigour: sources, searches and checking

In March 2025, the Tow Center for Digital Journalism at Columbia University published a simple test: they took excerpts from real news articles, asked eight AI search engines to identify the original article, its publisher, its date and its link, and checked the answers. Overall, more than 60% were wrong. Many were wrong with complete confidence, without a hint of doubt.

That does not mean AI is useless for research. It means it is useful for something different from what it seems. It is excellent at orienting you in a new topic, suggesting where to look and summarising what it finds. It is unreliable as the last word on exactly what a source says.

What AI does well when you research

• Mapping a topic. In minutes it gives you the vocabulary, key concepts, main debates and names worth knowing.
• Writing better searches. It suggests technical terms, synonyms and words in other languages you would not have thought of.
• Finding candidate sources. With web search on, it gathers relevant pages much faster than you can.
• Summarising and comparing. Once you have the sources, it helps you extract the essentials and see where they agree or contradict each other.

What it does not do reliably is guarantee that each claim matches exactly what the linked page says. That is why the method below separates two phases: exploring with AI and checking yourself.

Search and deep research modes

The main assistants now have two ways of working with current information:

• Web search. The assistant runs a few searches, reads some results and answers with links. It is fast and useful for specific questions.
• Deep research. The assistant plans, launches dozens of searches, reads many pages and writes a long report with citations. It takes several minutes and is useful for broad questions.

Without search turned on, the model answers from what it learned in training, which has a cutoff date and includes no verifiable links. For any current or checkable fact, turn search on (we explain why in Why AI makes things up).

The five-step method

1. Start with the question, not the tool

Write in one or two sentences what you need to know and why. "I want to know what Colombian rules say about electronic invoicing for individuals who sell online, to decide whether I have to invoice this year" is much better than "electronic invoicing Colombia".

2. Ask for a map before answers

Before looking for facts, ask the assistant to orient you: subtopics, key terms, types of source that would hold the answer (a law, a public body, an academic study, a report by an international organisation) and questions you should ask yourself. That way you will know what to look for and where.

3. Ask for sources and separate what is checked from what is inferred

With search on, ask for:

• One link per important claim, not a general list at the end.
• The type of source: primary (the study, the law, the original data) or secondary (a news story or blog that discusses it).
• The distinction between what the sources say and what the model deduces or summarises on its own.
• What it did not find, explicitly.

4. Open the links and go up to the primary source

This is the step nobody wants to do and the one that matters most. For each fact you are going to use:

• Open the link and search for the phrase or figure on the page (Ctrl+F or Cmd+F).
• Check it says the same thing, in the same context. It is common for the figure to exist but refer to another year, country or population.
• If it is a secondary source, follow its link to the primary one. A news story about a study is not the study.
• Note the full reference once you confirm it: author, title, date and link.

5. Cross-check with lateral reading

Professional fact-checkers do not assess a source by reading it top to bottom, but by leaving it: they open other tabs to see what others say about that source and whether other reliable outlets report the same thing. Researcher Mike Caulfield summed this up in the SIFT method, which applies just as well to an AI answer:

• Stop. Before you accept or share, pause for a moment.
• Investigate the source. Who is behind the page? Does it have authority on the subject?
• Find better coverage. Are there more reliable sources saying the same thing?
• Trace the claim. Follow it back to its original context.

A prompt for researching with sources

I want to research this topic: [TOPIC]
What I need it for: [GOAL]
Country or scope, if relevant: [SCOPE]

1. First, give me a brief map: subtopics, key terms and what kinds of
   sources would hold the answer (laws, public bodies, studies...).
2. Then search and answer with these rules:
   - One specific link for each important claim.
   - Say whether each source is primary or secondary.
   - Mark anything that is your own conclusion as [INFERENCE].
   - Prioritise official, academic and recognised institutional sources.
   - If you cannot find something, say so; do not fill it in.
3. End with a list of what I should check first.

There is a fuller version in the Deep research with sources template, and to review a text you already have, use Fact-check claims.

Warning signs in an answer

• Links to the home page of an outlet or institution instead of the specific document.
• Very round or very precise figures with no specific source next to them.
• Direct quotes attributed to a person with no link to where they said it.
• Academic references with authors, journal and year but no link, or a link that does not open: check that they exist.
• Suspicious unanimity. If the topic is debated and the answer mentions no disagreement, it is probably missing part of the picture.

Keep reading

• How to check an AI answer before you use it: the method for reviewing a text you already have.
• Why AI makes things up: where false citations and figures come from.
• How to summarise long documents with AI: what to do once you have the sources.
• RAG without the hype: when you need vector search and when you don't: how companies connect AI to their own documents.
• In the glossary: grounding, hallucination and RAG.

Sources

• Jaźwńska, K. and Chandrasekar, A. (2025). AI Search Has A Citation Problem. Columbia Journalism Review, Tow Center for Digital Journalism.
• Caulfield, M. (2019). SIFT (The Four Moves). Hapgood.
• Anthropic. Reduce hallucinations. Claude documentation.