LearnAI
Новинки AI · Безопасный AI · Lesson 3/10
← To the catalog
Task

How to recognize AI fiction

Parse the AI's response. Return JSON {"red_flags":[...],"suspicious_claim":"...","verify_step":"..."}: red_flags - signs of possible fiction in this answer; suspicious_claim - the most dubious statement; verify_step - how to check it outside the model. Return only JSON, no text around it.

Unlock access to submit solutions for instant AI review.
Catching hallucinations
💡 A piece of theory

Mechanics: The text generator predicts the next word based on probability rather than checking for truth. Therefore, he confidently invents something that does not exist: non-existent books, quotes, statistics, links, court cases. It's not a "lie" - the model simply fills a gap in a plausible way. Red flags: too precise numbers without a source, specific dates, volumes and pages in a narrow topic, “links” that are not opened, round confidence where experts argue. Technique: the more niche and verifiable the fact (number of the law, page of the article, quote with the author), the higher the chance of invention - it is the details that the model comes up with most easily. Second trick: ask the same question twice in different ways - if the answers about the “fact” diverge, this is a sign of fiction, not knowledge. Insider: Hallucinations are often hidden in the plausible middle of a long answer, rather than in obviously strange places. A typical mistake: trusting an answer just because it sounds confident and competent - style is not the same as truth.

How is it assessed · passing 70

  • 1Valid JSON without text around it20%
  • 2There are fields red_flags, suspicious_claim, verify_step30%
  • 3Signs and verification are linked to the example50%
Your promptClaude ⌄
🔒

Unlock access to submit solutions for instant AI review.