Research assistance and fact-checking AI-generated claims

What AI is good for in research

Used well, an assistant can speed up research without being the source of truth:

• Planning — "What questions should I answer to decide whether to install a heat pump in a 1960s house? Group them by topic."
• Vocabulary — "What terms do specialists use for this, so I can search better?"
• Explaining sources you provide — "Explain the methods section of this paper in plain language."
• Structuring notes — "Turn my notes into a table of claims, sources and open questions."
• Finding gaps — "Which of these questions do my sources not answer?"

What to treat with caution

• Facts, figures and dates produced from the model's memory.
• Citations and links — models can invent references that look real, or attach real references to claims they do not support.
• Recent events beyond the knowledge cutoff.
• Specialist advice (health, law, finance, safety) — use qualified sources and people.

Tools that search the web or read your files reduce some of these risks, but retrieved pages can be wrong, outdated or misread. Always open the source.

A fact-checking routine

For every claim you intend to rely on:

1. Isolate the claim. Ask the model to list its factual claims as separate bullet points, or do it yourself.
2. Find an independent source. Search for the claim yourself rather than asking the model whether it is right. Prefer primary sources (the official statistics office, the original study, the law itself, the manufacturer's documentation).
3. Read laterally. Check what other reputable sources say about the source and the claim, rather than judging a single page by its appearance.
4. Match the details. Numbers, dates, units, scope ("in the UK" vs "worldwide") and conditions must match.
5. Record it. Note the source and date checked next to the claim.
6. Decide. Confirmed → use with citation. Contradicted → remove or correct. Unverifiable → remove, or present clearly as unconfirmed.

Prompts that make checking easier

• "List every factual claim in your answer as a numbered list, marking any you are unsure of."
• "Answer only from the documents below. If the documents do not contain the answer, say 'not in the provided sources'."
• "For each point, quote the sentence from the source that supports it." (Then check that the quotes really exist.)

These prompts help, but do not replace checking: a model can still mark a false claim as certain or produce a quote that is not in the source.

Asking the model to "double-check" is not verification

"Are you sure?" often makes a model change a correct answer or defend an incorrect one. Verification means comparing with an independent, reliable source.