AI principles · Chapter 3 of 8
Strong here, weak there
AI does some hard tasks well and fails at easy-looking ones, can state false things with confidence, and is more useful when it says it does not know.
Goals
- Answer six questions or say “I don't know”, then score them two ways.To do
- Sort six everyday tasks onto a dated map of where AI does well and where it fails.To do
- Pass the check.To do
Step 1 of 4: The read
Uneven, and not by difficulty
In chapter 2 you saw a model write by picking likely next pieces of text, learned from the examples it was trained on. What it does well depends on those examples, not on what people find hard. It can translate a sign into three languages in seconds, then miscount the letters in a word, because it reads pieces of words rather than single letters.
Researchers call this a jagged frontier: the edge between what AI does well and what it does badly zigzags. The phrase comes from a 2023 study in which consultants used AI on realistic tasks (Dell'Acqua and colleagues). On tasks inside the frontier their work got better. On a task just outside it, those using AI were right less often than those working without it. The frontier also moves as models change, so any map of it needs a date.
Confident and wrong
A model always produces an answer, and its words sound just as sure when it is wrong. Writing something false in a confident tone, such as a quote from a book that does not exist, is called confabulation; many people call it hallucination. Being calibrated means your confidence matches how often you are right: when you say you are sure, you nearly always are.
Why guessing pays on most tests
Abstaining means saying “I don't know” instead of answering. Whether it pays depends on the scoring rule, the way a test turns answers into points. If only right answers count, a guess can only help, so someone who always answers beats someone who admits doubt. A 2025 paper by researchers at OpenAI argues that most tests used to rank AI models score this way, so a model that guesses ranks above one that admits it is unsure. The fix it proposes: make a wrong answer cost more than “I don't know”.
In the visual
You answer six questions; nobody could know some of the answers from this page. Your answers are scored both ways, next to two simulated players. Then you sort six everyday tasks onto a dated map: usually fine, check carefully, or don't rely on it.
Step 2 of 4: Try it: Answer or abstain, then map the tasks
Answer or abstain, then map the tasks
Answer six questions, see them scored two ways, then sort six everyday tasks onto a dated map.
Mode:RealSimulatedReal: your scores are computed in this page from your answers. Simulated: the café, the two other players and the task map are scripted for this chapter (the map is a judgement as of 2026-11-03, not a measurement). No AI model runs here.
Step 3 of 4: Check
Check
Next: Question 2 of 3
Step 4 of 4: What you can do now
What you can do now
- Explain why AI can do a hard task well and fail at an easy-looking one.
- Say why a confident answer can still be wrong, and why scoring only right answers rewards guessing.
- For an everyday task, decide whether AI output is usually fine, needs checking, or should not be relied on.