Module 3 of 6 · 30 minutes

Fact-checking AI output

Professional fact-checkers have fast, proven techniques. This module adapts them for AI output.

Learning objectives

  • Explain why AI shouldn't be used to verify itself
  • Apply the SIFT method
  • Use lateral reading to judge sources
  • Check citations in four steps
  • Prompt for verifiable output and triangulate important claims

3.1 Verify at the source

The single most important rule: don't verify an AI claim by asking the AI — or a second AI — whether it's true. Models can confidently "confirm" their own errors. Go to an independent, authoritative source.

3.2 The SIFT method

SIFT, developed by digital literacy expert Mike Caulfield and widely taught by universities and libraries, is a fast, practical approach that works well for AI output:

StepWhat to do with AI output
StopPause before using or sharing. Do you know where this claim comes from? How much does it matter?
Investigate the sourceIf the AI cites a source, who is it? Is it authoritative on this topic? Does it exist?
Find better coverageSearch independently for the claim. What do trusted sources — regulators, official statistics, peer-reviewed research, reputable outlets — say?
Trace to the originalFollow claims, quotes and statistics back to the original context. Is the figure current, correctly quoted and not cherry-picked?
SStopDo I know wherethis comes from?IInvestigatethe sourceReal? Authoritativeon this topic?FFind bettercoverageWhat do trustedsources say?TTrace to theoriginalCurrent, accurate,in context?
SIFT, a fast fact-checking routine used by fact-checkers and university libraries — adapted here for AI output.

3.3 Lateral reading

Professional fact-checkers don't read a source top to bottom to judge it. They read laterally: open new tabs and see what others say about the source and the claim. It's faster and more reliable than judging by how professional something looks — which, with AI, tells you nothing.

Source the AIcitedRegulator /official statsReputable newscoverageWikipediabackgroundWho fundsthis source?Open new tabs: what do others say about the source and the claim?
Lateral reading: instead of judging a source by how professional it looks — meaningless with AI — check what independent sources say about it.

3.4 Checking citations

  1. Does it exist? Search the exact title. For academic papers use a library database or Google Scholar; for UK law use official sources such as legislation.gov.uk or the National Archives' Find Case Law service.
  2. Are the details right? Authors, year, publication and link.
  3. Does it say what's claimed? Find the relevant passage. AI often cites real sources for things they don't say.
  4. Is it current and authoritative? A 2019 blog post isn't evidence about 2026 regulation.
1. Does it exist?2. Are the details right?3. Does it say what's claimed?4. Is it current and authoritative?Most fake or misattributed citations fail at step 1 or 3
The citation check funnel. A citation is only usable once it passes all four.

3.5 Prompt for verifiability

You can make checking easier from the start:

  • Ask for sources with links for every key claim, and for confidence levels.
  • Supply the sources yourself and say "answer only from these".
  • Ask for direct quotes from documents, with page or section numbers.
  • Ask "What are you least sure about in this answer?" — useful for prioritising, though not a substitute for checking.

3.6 Triangulation

For important claims, look for two or more independent, reliable sources that agree. If sources disagree, that's valuable information — report the disagreement rather than picking the answer you prefer.

Try it: SIFT a claim

Ask an AI tool: "What percentage of UK adults used generative AI in the last year? Give a source." Apply SIFT: stop, investigate the cited source, find better coverage (e.g. official statistics or Ofcom), and trace the figure to its original report. Was the AI's figure accurate, current and correctly described?

Key takeaways

  • Verify at independent, authoritative sources — not by asking the AI.
  • SIFT: Stop, Investigate the source, Find better coverage, Trace to the original.
  • Read laterally: what do others say about this source?
  • Citations: exists? details right? says what's claimed? current?
  • Triangulate important claims; report disagreements.