Predict where it breaks

The AI that told you when Finland joined the EU and the AI that invented your town's swimming hall hours used the same calm, certain voice, because the voice is a learned style, produced whether the content is right or wrong. This lesson gives you the mechanism behind that, one sentence long, and the map that falls out of it: the places where these models predictably break, knowable before you ask. This week that means deciding in advance which answers you can use as they stand and which to treat as drafts, instead of finding out after something goes wrong. You don't have this map yet because the delivery gives no signal, so daily use teaches you nothing about where the edges are, however much you use it. The habit you walk out with is predict-then-check, and you'll run it tonight on material you know better than the machine does.

The AI that told you when Finland joined the EU and the AI that invented your town's swimming hall hours used the same calm, certain voice, because the voice is a learned style, produced whether the content is right or wrong.

What is a language model doing when it answers a question?

Predicting plausible next words from patterns in its training text, one word at a time. The answer is a continuation of your text, rebuilt on the spot.

Why does an invented answer sound as confident as a correct one?

Confidence is part of the writing style the model learned. The style gets applied whether the content is right or wrong.

Does the model know when it doesn't know?

It has no reliable internal signal for it. Remembering and inventing are the same operation, prediction, and reported certainty is generated text too.

Why is "who won last night's game" a danger-zone question?

The game happened after the training text ended, so the true result has no route into the prediction. What comes back is a plausible result.

Which survives better, a rough magnitude or an exact figure?

The rough magnitude. Training text agrees on the size of things while holding many nearby exact figures, and prediction picks a plausible one.

Why are word-for-word quotes risky even from famous books?

The demand is one exact string. Prediction produces likely wording, and likely wording is close and wrong.

A village of 500 people with a nationally famous festival: which question about it is safe?

The famous thing. The zone follows the amount of text written about the subject, and fame produces text. The size of the place decides nothing by itself.

When does the plausible wrong answer beat the true one?

When the plausible version appears more often in training text than the true one, like a tidy retold story versus the historian's correction. Prediction sides with the more common text.

What does writing the prediction before asking add over just checking?

The miss. Checking fixes one answer, while a written prediction that proves wrong fixes your map of where the machine breaks.

1. When a language model answers your question, what is it doing?: Predicting, word by word, a plausible continuation of your text, based on patterns in its training text
2. The same model gives you one correct answer and one invented answer, in the same certain tone. Why does the tone stay the same?: The certain tone is a writing style learned from the training text, and the style gets applied whether the content is right or wrong
3. Web search is off. Which of these four questions should you trust the model's answer to most?: Who won the marathon at the 1952 Helsinki Olympics?
4. A village of about 500 people hosts a festival known across the country. Which question about the village is the model most likely to answer correctly?: What the village is known for
5. Which change most reliably moves a question from the safe zone into a danger zone?: Asking for the exact figure instead of the rough size
6. You add to your prompt: "Only answer if you are certain. Otherwise say I don't know." How much protection did you buy?: Some, at the margins, but the certainty it reports is generated text like the rest, produced without a reliable signal of what it knows
7. In spring you ask an AI for this year's limit of the commuter deduction in Finnish taxation, and it gives you a precise euro amount. Why does this question sit in a danger zone twice over?: The limit is a precise number that changes from year to year, so the training text is full of earlier years' limits, and each of them is a plausible answer