LearnAI
Новинки AI · Безопасный AI · Lesson 4/10
← To the catalog
Task

Make the AI ​​say "I don't know"

Make a request against fabrications. Return JSON {"prompt":"...","rules":[...],"why_works":"..."}: prompt — text of the request to AI; rules — rules of behavior in case of ignorance, built into the request; why_works - briefly why this reduces fudge. Return only JSON, no text around it.

Unlock access to submit solutions for instant AI review.
Prompt versus fiction
💡 A piece of theory

Mechanics: by default, the model “strives to help” and continues the text, even if there is no data - silence is an unlikely option for it. This means that permission not to know must be given explicitly. Prompt techniques: (1) direct instruction - “if you’re not sure, honestly say ‘I don’t know’ and don’t make things up”; (2) the requirement of sources - “give only facts that you can name with the source, mark the rest as assumption”; (3) separation—“separate verified facts from conjecture.” Insider: ask the model to estimate its confidence in percentage and explain WHAT exactly it does not know - this sharply reduces confident fabrications, because the task changes from “give the answer” to “assess the reliability.” Second trick: formulate a question without a built-in assumption - “does law X exist?” instead of “retell Law X”, otherwise the model will confirm something that does not exist. A typical mistake: believing that a polite “don’t be mistaken” works—we need specific rules for behavior in the face of ignorance, rather than a request to be precise.

How is it assessed · passing 70

  • 1Valid JSON without text around it20%
  • 2There are fields prompt, rules, why_works30%
  • 3Inquiry really resolves ignorance50%
Your promptClaude ⌄
🔒

Unlock access to submit solutions for instant AI review.