Availability: what comes to mind first

Lesson 3 · Thinking Fast and Slow × AI · about 10 minutes

Mission: stop judging AI's reliability by the last vivid error I saw.

The core idea

The availability heuristic judges how common or likely something is by how easily examples come to mind. Recent, vivid, or emotional cases feel more frequent than they are. Kahneman's book covers this in its Part II heuristics discussion (see Sources).

With AI this shows up in two places. Your memory of one dramatic failure can make you distrust a tool that is right most of the time. And the AI's own answers can be shaped by what is most common or most cited in its training, so the "typical" answer it gives may be the most available one, not the most accurate.

Why it matters for AI use

Automation bias research finds that trust in automated systems is shaped by experience, and that both over-reliance and under-reliance are errors (Goddard et al. systematic review). A vivid failure can swing you into under-reliance, and a string of smooth successes can swing you into over-reliance. Either way, your sense of the tool's reliability is built from a few memorable cases.

Practice

Choose the bias the scenario shows.

1. You read three news stories about AI inventing court cases. You now decide AI is unreliable for all legal research, though you have not measured it.
Correct. One vivid case is shaping a judgment about the whole category.
Not quite. The judgment is driven by how easily the one vivid case comes to mind.
2. You saw one AI error last week. This week you double-check every output, and you no longer check at all after a few good answers.
Correct. A recent error is driving extra checking, and that will fade.
Not quite. Your checking habit changed because of one memorable case, which is availability.
3. You want to test whether your view of AI reliability is skewed. What should you ask the AI?
Correct. Asking for counterexamples surfaces the cases your memory is missing.
Not quite. The fix is to surface cases you have not seen, such as counterexamples.

Takeaway

A memorable example is not a frequency. Before deciding how far to trust a tool, ask how often it fails on a sample you chose deliberately, not on the cases that happened to stick.

Try it this week: pick one task you use AI for, and write down ten recent outputs. Mark each as right or wrong. Compare the count to your gut feeling.

Sources

Questions? Stuck on any idea? Ask your teacher in this conversation. Previous: Lesson 0002. Next: Lesson 0004, base rates.
← Previous Next →