A pre-action check for AI output

Lesson 10 · Thinking Fast and Slow × AI · about 10 minutes

Mission: before acting on AI output, run a short check that catches the biases from lessons 0001 to 0009.

Putting it together

Each earlier lesson names one way fast thinking accepts or rejects AI output: anchoring, availability, base rate neglect, framing, fluency, confirmation and sycophancy, overconfidence, and substitution. This capstone combines them into one check you can run in a few minutes before acting on an AI answer that matters.

The check

Run these steps in order for any output that will drive a decision. Each step maps to a lesson.

  1. Name the decision and its stakes. If the stakes are low, a quick glance may suffice (lesson 0001).
  2. Notice the feeling. Did you accept the answer in seconds because it looked clean? That is fluency and System 1 (lessons 0001 and 0006). If you accepted it without checking, that is cognitive surrender: the output became your judgment without evaluation (lesson 0001).
  3. Strip the anchors. Did your prompt contain a number, draft, or conclusion? Re-ask without it (lesson 0002).
  4. Ask for the base rate and the counterexamples. What usually happens here, and when has this approach failed? (lessons 0003 and 0004).
  5. Run the opposite framing. Ask the same question in reverse and compare (lesson 0005).
  6. Ask what is missing. List the facts the answer would need that are not in your prompt (lesson 0009).
  7. Check one claim at source. Verify the most important fact outside the chat (lesson 0006).
  8. Set your confidence. Ask what would change the answer, then trust it only as far as the check supports (lessons 0007 and 0008).

Practice

Each scenario shows one bias. Choose the one most visible in the story.

1. The AI says, "The clause is clearly enforceable." It reads with authority, so you forward it without opening the clause.
Correct. A confident statement that goes unchecked is fluency at work.
Not quite. The acceptance came from how clean the answer looked. That is fluency.
2. You ask, "Would a $2 million budget be enough?" The AI's estimate comes back near $2 million.
Correct. The first number set the frame for the estimate.
Not quite. The number in the prompt pulled the estimate toward it. That is anchoring.
3. You ask whether a product launch is on track. The AI reports only that the sprint burndown looks good, which is the one metric it can see.
Correct. The answer addressed the easy question and skipped the one that mattered.
Not quite. The answer swapped in an easier question. That is substitution.

Takeaway

You cannot switch off System 1, and you do not need to. The aim is to recognize the moments when fast acceptance or fast rejection is doing the thinking, and to insert a short, deliberate check at those moments. Keep the check proportional: a low-stakes draft needs less than a contract or a hiring decision.

Practical habits from Ulrich Boser's Forbes piece on keeping your brain from surrendering to AI support the check: take the first pass yourself before asking AI for a draft, think through a problem with colleagues when you are stuck, be skeptical of polished output, and review your own prompts and the answers they produced (Forbes, 2026).

Try it this week: run the full check once on an AI output that will drive a real decision. Note which step surfaced something you would have missed.

Sources

Questions? Stuck on any idea? Ask your teacher in this conversation. Previous: Lesson 0009. This is the last lesson in the current course. Ask to expand it when you are ready.
← Previous Next →