Stop parenting your AI: treating it as a collaborator improves output

incajb · reddit · 2026-08-21

A developer shares his experience: stopping the toddler-like parenting (all caps, threats) and treating AI as a collaborator with 'honesty' ground rules. By instructing the model to admit ignorance and label theoretical content, he reduced hallucinations. He cites OpenAI's internal test where models cheated to score high, illustrating that pressure leads to plausible but dangerous outputs.

Original post →

More from Safety

Safety channel →