Bias, representativeness and context
Data can reflect inequalities, omissions or a reality unlike the one in which a system will be used. A good average result can hide failures in one language, group or type of case. Look for who is poorly represented.
Define what would count as a significant error and compare results across relevant scenarios while respecting privacy. Include people who know the context in the evaluation. No single metric proves fairness, suitability or the absence of discrimination.
A request classifier may work well on formal messages and fail on informal Portuguese. Evaluate these cases separately before making it available to the team.
Always ask: for whom does it work, in what situation and with what consequences?
Try it yourself
Choose an answer containing three factual claims. Mark each as confirmed, incorrect or needing verification. Record a primary source and one limitation. Explain the impact of publishing the answer without review.