Negative Transfer in AI: Four Root-Cause Mechanisms Defined in a Chinese Governance Series

This second installment of the Carbon-Silicon Dao Code series defines four types of negative transfer in cross-domain AI: NT1 mechanism mismatch, NT2 semantic drift, NT3 unknown completion, and NT4 power leakage.

It argues that current evaluation based on fit accuracy and test-set pass rates cannot detect whether the underlying mechanisms match.

The article is a Chinese-language theoretical and governance piece, and the summary covers only the framework it presents, not empirical results.