Measuring Human Contribution to Human–AI Collaborative Content Generation in the Wild

With the widespread adoption of generative artificial intelligence, content creation is shifting from independent human authorship towards human-guided collaborative generation with AI. The extent of human contribution therefore varies substantially across works, challenging conventional notions of originality. Existing approaches to measuring human contribution on a continuous scale rely primarily on internal model information, such as token probabilities. Meanwhile, content generation increasingly relies on black-box services provided by frontier models such as GPT, Claude, and Gemini. This mismatch between the white-box access required for evaluation and practical access constraints limits the applicability of existing methods in the wild. To address this limitation, we propose the Human Contribution Evaluation Framework (HCEF) for measuring human informational contribution using only observable human inputs and AI outputs. Drawing on the information-theoretic notion of side information, HCEF measures human contribution as the usable information gain that human inputs provide about the realized output. It normalizes this gain by the output’s baseline information content to obtain a human contribution score. This enables continuous, recomputable instance-level assessment without access to the generating model’s internal states or probability distributions. Extensive experiments demonstrate that HCEF effectively captures varying degrees of human contribution under strictly black-box conditions, with further validation in the wild.