Who Gets Credit or Blame? Attributing Accountability in Modern AI Systems

Researchers propose a framework to attribute AI model behavior to specific development stages (pretraining, fine-tuning, alignment), enabling accountability tracking without model retraining. The method quantifies how each stage contributes to model outputs and can identify spurious correlations, advancing transparency in AI development.

The accountability attribution problem addresses a fundamental challenge in modern AI development: determining which stage of model development is responsible when deployed systems succeed or fail. Current AI systems undergo multiple transformation stages, each modifying model weights and behavior in distinct ways. This complexity obscures responsibility, making it difficult for developers, regulators, and users to understand what causes specific model outputs. The research introduces a counterfactual framework that efficiently estimates stage effects by analyzing how model behavior would differ if particular development stages were excluded, without requiring expensive model retraining. This efficiency gain is crucial because retraining modern large language models costs millions of dollars.

The technical contribution accounts for optimization dynamics like learning rate schedules, momentum, and weight decay—factors often overlooked in simpler attribution methods. The work demonstrates practical applications by identifying and removing spurious correlations in image classification and toxicity detection tasks. This matters because spurious correlations often lead to discriminatory or unreliable model behavior in production. For the AI industry, better accountability mechanisms address growing regulatory scrutiny around model transparency and responsibility. Developers gain actionable insights for debugging model failures and improving training pipelines. The framework represents progress toward more interpretable AI systems, though it currently applies primarily to understanding existing models rather than preventing problematic behaviors during development. Future work likely involves scaling these methods to larger models and integrating accountability attribution into standard development workflows to catch issues earlier.

• →Framework enables tracing model behavior to specific development stages without retraining, reducing analysis costs significantly.
• →Method accounts for optimization dynamics including learning rates, momentum, and weight decay overlooked in simpler attribution approaches.
• →Practical applications include identifying and removing spurious correlations in image and text classification tasks.
• →Addresses regulatory demands for AI transparency by clarifying responsibility across multi-stage model development.
• →Represents incremental progress toward interpretable AI but operates on post-deployment analysis rather than preventive development controls.