The Newspaper Test
Imagine you are a newspaper editor in 1985. A junior reporter hands you a story full of fascinating facts. Your job before you print it: call the sources, check the numbers, verify the quotes. You do not distrust the reporter — you just know that even good reporters make mistakes, and your name goes on the front page.
That is exactly how to treat AI answers. Not with suspicion, but with the same professional habit of verification you would apply to any source. AI is a junior researcher who works very fast, almost always points you in the right direction, but occasionally gets confident about something it should not. A 60-second check is your editorial layer.
The 5-Step Fact-Check Process
Step 1 — Identify the checkable claims
Read the AI response and highlight anything that is a specific fact: a number, a date, a named person, a quoted source, a statistic. General explanations ("photosynthesis converts sunlight into energy") rarely need verification. Specific claims ("a 2019 Harvard study found that 68% of adults…") always do.
Step 2 — Search for the claim independently
Open a new tab and search for the claim on Google or another search engine — not more AI. You are looking for a primary source: the original study, the official government page, the established news article that reported the fact first.
Step 3 — Check the source quality
Not all sources are equal. Government websites (.gov), peer-reviewed journals, well-known universities, and major established news organizations are high-quality sources. Random blogs, AI-generated content farms, and sites you have never heard of are not. Look at who published the information and when.
Step 4 — Look for the original study or data
If AI cites "a study," find that actual study. Search its title in Google Scholar or PubMed. If the study does not exist, or the AI described its findings inaccurately, you have caught a hallucination. This step is especially important for medical, scientific, or financial claims.
Step 5 — Ask AI to reconsider
If something does not check out, paste the correction back: "I searched for this and the actual figure appears to be X, not Y. Can you update your response?" Good AI models will acknowledge the correction gracefully. This also helps calibrate your trust in AI on that topic going forward.
What to Verify vs. What to Trust
- Usually safe to trust: General explanations of how things work, brainstorming ideas, writing suggestions, broad historical summaries, coding help, creative writing.
- Always verify: Specific statistics and percentages, citations and paper references, medical dosages or treatment recommendations, legal requirements, recent news (after the AI's training cutoff), names of real people and their affiliations.
Try This Today
- Ask AI a factual question about something you know well. Evaluate how accurate it is.
- Ask AI for a statistic on any topic. Spend 60 seconds trying to find the original source of that statistic. Notice whether the source exists and whether the AI quoted it accurately.
- Try: "Give me three surprising facts about [topic you like] — and for each one, tell me where I could verify it." This trains both you and the AI to be more accountable.
Common Questions
Not every single response. Casual tasks like drafting emails or brainstorming need little fact-checking. Reserve careful verification for medical, legal, financial, or professional claims you plan to act on or share publicly.
For general facts: Wikipedia (as a starting point, then check its citations). For science: PubMed, ScienceDirect. For news: established news outlets. For statistics: government sites like CDC.gov or census.gov, or the original study.
You can, but it is not reliable. Two different AI models may both produce the same wrong fact if their training data had the same errors. Always verify against primary human-curated sources.
For most claims, 30–60 seconds of searching is sufficient. A quick search confirming the main claim is often all you need for everyday use.