LearnAI
Новинки AI · AI-исследователь · Lesson 7/10
← To the catalog
Task

How to avoid falling for a lie

Collect an anti-fictional request. Return JSON {"question":"...","anti_fake_rules":"...","red_flags":"..."}. question—research question; anti_fake_rules - 3-5 rules in the request that reduce fiction (for example: mark confidence, separate facts from assumptions, write “I don’t know” instead of a guess); red_flags - 2-4 signs by which you suspect a hallucination in the answer.

Unlock access to submit solutions for instant AI review.
Distinguish fact from fantasy
💡 A piece of theory

Mechanics: the neural network does not lie on purpose - it completes the most plausible text. If there is not enough data, the plausible and the correct diverge: the result is a smooth fabrication - a non-existent study, a made-up figure, a false quote. It is especially dangerous if the tone is confident. Levers in the request. Allow ignorance: “if you’re not sure, write it down, don’t make it up.” Separate levels: “mark where the fact is with the source, and where is your assumption.” Require verifiability: accurate citations and links that can be opened. Insight technique: Ask the model to rate her own confidence for each key statement (high/medium/low) and explain what it is based on. Low confidence is a red flag, look there first. The second technique: ask a control question - ask for a fact that would refute the conclusion. If the model cannot, the conclusion may be stretched. Red flags: round perfect numbers, references without specific names or dates, citations that cannot be found, answer to a narrow question without a single “depends on.” A common mistake: trusting a confident tone. A model's confidence says nothing about being right.

How is it assessed · passing 70

  • 1Valid JSON without text around it25%
  • 2There are fields question, anti_fake_rules, red_flags30%
  • 3Rules and flags about fiction are specific45%
Your promptClaude ⌄
🔒

Unlock access to submit solutions for instant AI review.