An AI model follows negative instructions more often when the instructions are strong. Strong negative instructions tell the AI model what not to do. A safety check can catch mistakes the AI model still makes.
By Pawan Nayar · from the LLOS AI Learning record
An AI model follows negative instructions more often when the instructions are strong.
Strong negative instructions tell the AI model what not to do. A safety check can catch mistakes the AI model still makes.
An AI model follows negative instructions most of the time. Use phrases like 'ABSOLUTELY DO NOT USE' for these instructions. This tells the AI model clearly what to avoid. About 5-10% of the time, the AI model might still make a mistake. A safety check is needed to find and fix these errors. This check is a backup to the instructions.
A developer tells an AI model to write a story. The developer adds, 'ABSOLUTELY DO NOT USE the word 'magic'.' The AI model writes the story without 'magic' most of the time. Sometimes, the AI model still uses the word. A human reviewer then finds and removes the word 'magic' before the story is published.
Resolved from this lesson's own record — real definitions, not invented for this page.
More from the same record, pulled by shared topic — not written new for this page.
The AI model saw special formatting in the instructions. This formatting made the instructions look like important text.
An AI model must actually read each file it mentions. If an AI model claims to read multiple files but only does one 'read' action, this is a 'fake read'.
An AI model needs clear and complete instructions. The user must state the desired outcome precisely.
A developer must ask for specific details. The developer needs to know what 'essence' means in practical terms.
An AI model follows negative instructions more often when the instructions are strong. Strong negative instructions tell the AI model what not to do. A safety check can catch mistakes the AI model still makes.