New Framework Uses Shapley Values to Fairly Credit Sources in AI Summaries

When large language models summarize information from multiple sources to answer user questions, the original creators of that content often go unrecognized and uncompensated. Researchers have now developed a framework to solve this attribution problem using game theory.

The Core Problem#

Today's AI-powered search engines and assistants pull information from many documents and serve synthesized answers directly to users, obscuring which sources actually contributed to the response. This lack of transparency threatens creator revenue and the sustainability of content ecosystems. The challenge lies in fairly measuring each document's contribution to the final summary—a task that's mathematically complex and computationally demanding.

Cluster Shapley Approach#

To address this, the research team proposes a method grounded in Shapley values, a concept from game theory that assigns each participant a fair share of total value based on their marginal contributions. Because computing exact Shapley values becomes prohibitively expensive at scale, they developed Cluster Shapley, which groups semantically similar documents together using LLM embeddings and computes Shapley values at the cluster level rather than for each individual document. This approach comes with formal guarantees bounding both the approximation error and the resulting attribution mistakes.

Performance Validation#

Testing on Amazon product review data revealed that existing shortcuts—such as Monte Carlo sampling or Kernel SHAP—underperform in LLM contexts. By contrast, Cluster Shapley achieves substantially better results on the efficiency-accuracy tradeoff. Simple allocation heuristics like equal distribution or relevance-based weighting are computationally cheap but produce highly unfair attributions. The framework is designed to work with any LLM, summarization method, or evaluation approach, making it broadly adaptable across different systems.

Why It Matters#

Fair source valuation is essential for content creators who depend on audience reach for revenue. By transparently accounting for each document's contribution to summaries, this work could support compensation systems that reward original reporting and writing.