Key points
- Match the effort to the stakes: casual ideas need no checking; anything you’ll publish, submit, pay for or act on does.
- Pull out the checkable claims — names, numbers, dates, quotes, citations — and verify each at a primary source.
- Use prompts that make checking easy: ask for sources, quotes and confidence levels.
- If you can’t verify it, don’t use it as fact.
Why you need a routine
AI chatbots are right most of the time on common topics, which is exactly what makes their mistakes dangerous: after twenty good answers, it’s natural to trust the twenty-first. But language models generate plausible text rather than retrieve verified facts, and research suggests training methods have tended to reward confident guesses over admitting uncertainty (we explain this in Why AI hallucinates). A short, consistent checking routine lets you keep the speed of AI without inheriting its errors.
Step 1: Decide the stakes (10 seconds)
Not everything needs checking. Use a simple scale:
- Low stakes — brainstorming, first drafts, recipe ideas, rewriting your own text. No check needed beyond common sense.
- Medium stakes — a work email with facts, a school assignment, a social media post, a purchase decision. Check key facts.
- High stakes — health, money, legal matters, safety, anything published under your name, anything affecting others. Check every claim, and consult a qualified professional where appropriate.
Step 2: Extract the checkable claims (1 minute)
Read the answer and list every specific claim: names, numbers, dates, prices, quotes, statistics, citations, product features, legal or medical statements. Opinions and general explanations are less risky; specific facts are where errors hide. You can ask the chatbot to do this extraction for you:
“List every factual claim in your previous answer as a numbered list. For each, say how confident you are (high, medium, low) and what kind of source would confirm it.”
This won’t verify anything, but it gives you a checklist and often flags the weakest points.
Step 3: Verify at primary sources (2–3 minutes)
For each claim, go to the original source rather than another summary:
| Claim type | Best place to check |
|---|---|
| Product price or feature | The company’s official pricing or plan page |
| Statistic | The organisation that published it (statistics office, research body) |
| Research finding | The paper itself (read the abstract and the relevant section) |
| Law or regulation | The official government or legislative website |
| Health information | National health services, medical bodies, your doctor or pharmacist |
| Quote | The original speech, interview or document; search the exact words in quotation marks |
| News event | Two or more established news organisations, ideally with original reporting |
When checking a citation, confirm four things: the source exists, the author and date match, it says what the AI claims, and it’s still current. Fake or mismatched citations are one of the most common AI errors.
Step 4: Check the date (30 seconds)
A fact can be true and outdated. AI models have knowledge cutoffs, and prices, product features, laws and statistics change. Look for the “last updated” date on the source. For fast-moving topics like AI products themselves, information older than a few months may already be wrong — that’s why every comparison on this site shows a “last checked” date.
Step 5: Fix, label or cut (30 seconds)
Correct what you can, label what’s uncertain (“reportedly”, “according to…”), and remove anything you couldn’t verify. If you’re sharing the information, cite the primary source, not the chatbot.
Prompts that make checking easier
Prevention beats cure. Add these to prompts where accuracy matters:
- “Cite a specific source for each factual claim, with a link if you have web access.”
- “If you’re not sure, say so — don’t guess.”
- “Use only the document I’ve provided. Quote the relevant sentence for each point.”
- “Separate what you’re confident about from what you’re unsure about.”
Anthropic’s guidance on reducing hallucinations recommends several of these techniques, including allowing the model to say it doesn’t know and asking for direct quotes from source documents.
Worked example
You ask a chatbot: “How much does Google AI Pro cost and what storage does it include?” It answers with a price and a storage amount. Stakes: medium (you’re about to subscribe). Claims: the price, the storage, and implicitly the country. You open Google’s official subscriptions page for your country. When we checked the US page on November 19, 2026, it listed Google AI Pro at $19.99 per month with 5 TB of cloud storage. If the chatbot’s figures differ, the official page wins — and prices may differ in your country.
Mistakes people make when checking
- Checking against another AI summary instead of the original source.
- Stopping at “the link works.” A real page can still say something different from the AI’s claim.
- Ignoring the date on fast-moving topics.
- Checking only the surprising claims. Familiar-sounding errors are the ones that slip through.
Over time you’ll learn which kinds of answers your favourite chatbot gets right and which it tends to get wrong. Keep notes; that knowledge makes your checking faster and more focused.
Remove this box by adding your content here, or hide all notes with
"showEditorNotes": false in site.config.json.Tools that help
- Chatbots with web search and citations make step 3 faster, because the links are right there. Compare them in our guide to the best AI for research.
- Search engines with exact-phrase search (put the phrase in quotation marks) for quotes and titles.
- Scholarly search engines and library databases for academic citations.
- Reverse image search for images that might be AI-generated or out of context — see how to spot AI scams and deepfakes.
For students and teachers
UNESCO’s guidance on generative AI in education emphasises human agency and critical evaluation of AI output. In practice, that means treating AI answers as a starting point for your own research, not a source to cite. Most schools require you to cite the original sources you verified, not the chatbot — check your institution’s rules. Our guide to using AI for studying covers this in detail.
Make it a habit
The routine takes five minutes for a typical answer and becomes automatic with practice. The payoff is large: you get the speed and creativity of AI with the reliability of your own judgement.
Frequently asked questions
Can I ask the AI to fact-check itself?
It can help — asking “which claims here are you least sure about?” often surfaces weak spots — but a model checking its own work can repeat the same mistake. Always verify important claims at an independent source.
Is Perplexity more accurate because it shows sources?
Showing sources makes checking much easier, but the summary can still misread or misquote a source. Click through and confirm the claim is really there.
What counts as a primary source?
The original origin of the information: the company’s own pricing page, the government agency’s page, the research paper itself, the official statistics release, the court’s published decision.
How do I check if a citation is real?
Search the exact title in quotes in a search engine or a scholarly database, then open the paper and confirm the author, year and the specific claim.
What if I can’t verify a claim?
Leave it out, or clearly label it as unverified. Never pass on an unverified AI claim as fact in work, school or public posts.
How long does fact-checking an AI answer take?
For a typical answer with a few specific facts, about five minutes. High-stakes answers with many figures or citations take longer — and deserve it.
Sources
- Anthropic — Reduce hallucinations (opens in a new tab)
- Kalai et al. (2025), Why Language Models Hallucinate (opens in a new tab)
- UNESCO — Guidance for generative AI in education and research (opens in a new tab)
- Google Search Central — Creating helpful, reliable, people-first content (opens in a new tab)
- NIST — AI Risk Management Framework (opens in a new tab)
- Google AI plans (US) — used in the worked example (opens in a new tab)
Facts, prices and features were checked against these sources on November 19, 2026. AI products change quickly; if you spot something out of date, please tell us.