Whose Ground Truth? Embracing Ambiguity in Human-Centered AI

Current AI systems oversimplify human judgment by treating it as noise: Conventional machine learning assumes one correct answer exists and averages out disagreements between human annotators. This approach discards valuable information about legitimate differences in how people interpret the same situation, which is particularly problematic for decisions affecting real people.

Many real-world decisions don't have a single "right" answer: For human-centered AI tasks, multiple reasonable interpretations often exist simultaneously. The paper argues that reducing this meaningful ambiguity to one ground truth "risks overlooking meaningful information about the diversity of human perception, judgment, and experience"—meaning AI systems trained this way may fail to account for valid alternative perspectives.

AI development needs to shift toward modeling multiple valid interpretations: Rather than forcing agreement, the paper calls for AI systems to be designed, trained, and evaluated based on the full range of plausible human judgments. This requires distinguishing between genuine ambiguity (where multiple answers are legitimately reasonable) and actual annotation errors.

How we build AI affects how it gets deployed and governed: The paper emphasizes that this perspective should influence not just how AI learns and is evaluated, but also how it's represented to users and regulated. This suggests that acknowledging ambiguity in AI systems may require changes to how organizations communicate about AI capabilities and limitations.