AI中的偏见是指模型中的系统性不公平——产生对某些群体不公平的结果,通常反映了训练数据中的偏见。这是一个严重的伦理和实践问题,因为有偏见的AI可能造成真实伤害并加剧歧视。
什么是AI偏见
AI BIAS → systematic, unfair skew in a model's outputs:
→ the model treats certain groups unfairly (e.g. by race, gender, age) or makes skewed
decisions
→ usually stems from BIASED TRAINING DATA (the model learns the biases in the data)
→ 'bias in, bias out' → models reflect and can AMPLIFY societal biases in their data
→ AI can perpetuate or worsen unfairness/discrimination
偏见来自何处
✓ BIASED DATA → training data reflects historical/societal biases or isn't representative
→ e.g. hiring data favoring one group → the model learns to favor that group
✓ UNREPRESENTATIVE data → underrepresented groups → poor performance for them
✓ Biased labels, flawed problem framing, biased features → encode unfairness
→ bias mostly originates in the DATA (and how the problem is set up)
