The most dangerous word in AI is confidence. A model that is wrong is still sure, and it explains its confidence beautifully. The fix is not more caution — it is a mechanical check on the claim.
The one-question test
Ask the chatbot: “What is your primary source for that, and give me its URL?” Models that answer with a phrase like “according to RBI” will usually also produce, upon request, the specific page. Your job is to open it.
Why it matters
- Fluency and accuracy are unrelated skills in a model.
- If no URL leads back to an official source, the claim is a rumour with syntax.
- The stakes decide the check: money/identity first, trivia never.
If it cannot say (or the page is dead)
Key takeaways
- Confidence is not evidence.
- Ask for the source URL and visit it.
- Money and identity claims always go to an official portal first.
Treat it as unverified. For laws, rules and portals of India, the correct answer trace is always: official portal + gazette/service page — checks that a chatbot can shortcut but never replace.
The doubt ladder
- Anything about your money, identity or legal status: verify the official site directly.
- Statistics and dates: check the quoted source page, not a summary.
- General knowledge: a wrong city or date is a comfort issue, not a risk issue.
What you should do
- Ask for sources and open them — 5 minutes of curiosity defeats most hallucinations.
- Verify anything legal/policy-related strictly on official government portals.
- When a claim affects a decision, route it once through an official channel before acting.