The dividing line between people who master AI and people who get caught up in trouble is right here: whether you think of AI's answer as a "finished product" or as a "draft."
AI's Answer Is a "Confident Draft"
AI carries itself with poise even when it's wrong. So "it's stated definitively, therefore it's correct" doesn't hold. It's safer to assume the confidence of the tone and the correctness of the content are unrelated.
What to Verify and What You Can Skip
Verifying everything is exhausting. Prioritize by risk.
- Always confirm: numbers, dates, names of people and companies, quotes, law/medicine/taxes, URLs, whether code works
- Lightly confirm: gaps in summaries, nuance, the use of technical terms
- Mostly unnecessary: rough drafts of writing, brainstorming, paraphrasing, idea generation
Concrete Ways to Verify
- Confirm against primary sources β Check official sites, the original text, reliable statistics
- Re-run it through another AI or search β Ask the same question of a different tool and look for discrepancies
- Have the AI re-check itself β Ask it to "point out errors and weak spots in this answer"
- Ask the opposite β "What are the counterarguments or counterevidence to this claim?" to expose bias
Have the AI Critique Itself
Surprisingly effective is having AI hunt for flaws in its own answer.
For the answer you just gave, honestly list any spots that might be wrong or where the basis is weak.
This surfaces the dubious parts lurking in the first answer.
Summary
Verification looks like a hassle, but once you're used to it, it takes tens of seconds. And once you accept that "AI is a tool that produces drafts," you become able to use it boldly with peace of mind. Doubting is the prerequisite for trusting AI.