Systemizing Multiplicity: The Curious Case of Arbitrariness in Machine Learning

Multiple "good" models can make very different predictions: Even when machine learning models perform equally well on test data, they may reach opposite conclusions for the same person or situation. This arbitrariness matters because the specific model chosen for deployment could significantly affect real-world outcomes, yet this choice often appears arbitrary rather than principled.

Design choices throughout model development introduce hidden arbitrariness: Beyond just the final model selection, arbitrary decisions made during data collection, feature engineering, and hyperparameter tuning all contribute to which predictions get made. The paper formalizes how these upstream choices compound to create multiplicity, suggesting that "good performance" alone doesn't guarantee consistent or fair outcomes.

Multiplicity is distinct from—but related to—other sources of AI unreliability: The paper clarifies that multiplicity differs from uncertainty (not knowing what will happen) and variance (inconsistency across data samples). Understanding this distinction is important because it means some AI systems could appear confident and stable while still being fundamentally arbitrary in their specific decisions.

Multiplicity presents both opportunities and risks for responsible AI: While multiple valid models could enable more democratic or context-sensitive AI systems, they also risk obscuring accountability—organizations might use model multiplicity to justify unfavorable decisions as inevitable rather than chosen. The paper emphasizes that multiplicity itself is neutral; how we respond to it determines whether it helps or harms responsible AI.

This is an emerging research area with critical open questions: The paper identifies that we still lack clear guidance on when multiplicity matters most, how to communicate it to stakeholders, and whether some forms of arbitrariness are more acceptable than others in different contexts.