Safety Tuning

Training AI to refuse harmful requests, avoid bias, and stay within policy.

Definition

Safety tuning is the post-training stage where models learn to decline harmful requests, refuse to produce dangerous content, and avoid bias. Workers craft refusal examples and adversarial test cases.

Examples

  • Writing examples of how the AI should refuse to make weapons
  • Crafting adversarial prompts that probe safety boundaries

Related terms

Looking for actual gigs that use Safety Tuning? Browse all live listings or filter by $100+/hr.