Bias (in AI)
When AI outputs systematically favor or disfavor groups — can be racial, gender, geographic, ideological, or other.
Definition
Bias in AI comes from biased training data (the internet is biased) and biased preference labels. Detecting and correcting bias is its own contractor track — pays well, requires careful judgment.
Examples
- AI consistently rating "doctor" as male and "nurse" as female
- AI being more polite to certain dialects
Related terms
Evaluation (Eval)
Testing AI outputs against quality criteria — the largest single category of AI training work.
Red Teaming
Adversarially probing AI for safety failures — getting it to produce harmful, biased, or dangerous outputs.
Safety Tuning
Training AI to refuse harmful requests, avoid bias, and stay within policy.
Looking for actual gigs that use Bias (in AI)? Browse all live listings or filter by $100+/hr.