Discernment: trust, but with judgement
Learn to tell when an answer is good, when it needs checking and when it belongs in the bin.
AI writes well. Too well: even when it’s wrong, the text is tidy, confident, convincing. Discernment is the ability not to be charmed by the form and to look at the substance.
Three things to evaluate
The 4D framework suggests looking at an answer from three angles:
- The product — is it correct? Complete? Right for my purpose and audience?
- The process — do the steps make sense? Any logical leaps or rushed conclusions?
- The performance — did it follow my instructions? Is it just agreeing with me?
Watch out for the AI that “agrees with you”
Assistants tend to be accommodating. If you ask “my idea is brilliant, right?”, they’ll probably say yes. For an honest opinion, ask explicitly: “Tell me the 3 weak points of this idea.”
The warning signs
These are the signals that should make your antennae go up:
- Very precise numbers, dates and percentages without a source.
- Quotes attributed to famous people.
- Titles of books, studies, laws or specific court rulings.
- Links and web addresses (often made up or broken).
- Absolute promises: “always”, “never”, “guaranteed”, “risk-free”.
- Recent news, if the assistant didn’t search the web.
The Description ↔ Discernment loop
Describing and evaluating aren’t two separate moments: they form a loop. Ask → evaluate → adjust the request → evaluate again. That’s how you get to a good result.
How to check (without going crazy)
You don’t need to check everything. Use common sense and the risk level from the Delegation lesson. When it matters, here’s how:
The second-opinion trick
Open a new chat (or another assistant) and ask it to review the answer: “Find any mistakes in this text”. It’s no guarantee, but it often brings weak spots to light.
Watch out for bias too
AI learned from texts written by people, and inherited their BiasA skewed tendency in the answers, inherited from the texts the AI learned from. For example assuming a doctor is a man. too. For example it might assume “the nurse” is a woman and “the engineer” a man, or always give examples from just one culture. If you notice it, point it out and ask for a more balanced version.