偏见Bias
Bias
模型在不同人群、语言、文化上的表现不均：源头在训练数据。Uneven model performance across groups, languages, and cultures — rooted in training data.
国际 AI 安全报告指出：模型在训练数据里少见的语言与文化语境中表现下降。
表现的不均直接归因于训练数据的分布。
语料本身携带的刻板印象会被模型学到并复述。
人对 AI 建议不加检查就接受（自动化偏见）是另一层问题，别与模型偏见混为一谈。
The International AI Safety Report notes performance declines in languages and cultural contexts less common in training data.
The unevenness is attributed directly to the distribution of training data.
Stereotypes carried by the corpus are learned and reproduced by the model.
Humans accepting AI suggestions without checking (automation bias) is a separate layer — don't conflate it with model bias.
同一个模型对不同的人不一定一样好。评估 AI 产品时，「对你所在的语境表现如何」是必问题。The same model is not equally good for everyone. When evaluating an AI product, 'how does it perform in your context' is a required question.
回到名词地图看第2站的其他 35 个词 →Back to the map: the other 35 terms in station 2 →