bias (noun)
A systematic skew in an AI model's behaviour, traceable to an imbalance introduced at any point in the chain that produces and delivers it. Training data is one common source, but the same skew can enter through choices made in building or tuning the model, or through how the AI model is used.
*Example: "The team kept regenerating job-description drafts until one sounded like their existing hires — the bias wasn't in the model's training data, it was in which draft the team chose to keep."*
Related: responsible AI, guardrails, model