LearnAI
Новинки AI · AI, который видит · Lesson 10/10
← To the catalog
Task

AI Vision Limits and Verification

Make a verification request to the sighted AI to catch a possible fabrication. Return JSON {"prompt":"...","verification_step":"...","risky_assumptions":"..."}: prompt - task, verification_step - a technique that will separate honest reading from speculation, risky_assumptions - what exactly AI can invent in this frame and how to catch it.

Unlock access to submit solutions for instant AI review.
Don't trust blindly
💡 A piece of theory

Sighted AI does not “photograph” the truth - it assumes the most probable. In a clear photograph this is almost always true, but in a blurred, glare or cropped area, the model completes what is plausible: it confuses 3 and 8, reads text that is not there, and confidently names the wrong brand. And the tone is as confident as with the correct answer - this is the main trap. Test technique: separate reading and interpretation. Ask the AI ​​to first rewrite verbatim what it sees (for example, each digit of the counter in order), and SEPARATELY - the interpretation. A discrepancy between “read” and “concluded” is a red flag. The second technique: he was directly instructed to mark the unreadable as “unknown”, rather than guess, and to indicate how confident he was in each fragment. Hidden lever: re-photograph the disputed area from a different angle or at a larger angle and compare the two answers - a stable match is more reliable than one confident one. A typical rookie mistake: taking a smooth, confident answer as a fact and letting the AI ​​fill in the obscured numbers with fiction.

How is it assessed · passing 70

  • 1Valid JSON without text around it25%
  • 2There are fields prompt, verification_step, risky_assumptions25%
  • 3Anti-fiction checking mechanism50%
Your promptClaude ⌄
🔒

Unlock access to submit solutions for instant AI review.