If AI bias mitigation methods continues to gain relevance, what signals indicate whether it changes user behavior in unexpected ways — and what second-order effects could follow?