A Tally of Framings Is Not a Census of Causes: What the AI Risk Repository's Human-Versus-AI Split Measures, Why Its Gap Closed From Seventeen Points to Four as the Corpus Grew, and the Attribution Rule a Causal Claim Would Need

Abstract

Life of Research

The AI Risk Repository extracts every risk named in published AI risk frameworks and codes each one by the entity presented as causing it. Its journal version reports that "human decisions cause nearly as many AI risks (38%) as the AI systems themselves (42%)", which it calls contrary to common assumptions. Its first version, coding a smaller corpus with the same Entity categories (version 3 later reworded the definition of Other), reported 51 percent AI and 34 percent human and concluded that risks tend to be attributed to AI. In this analysis the author reviews and synthesises published results to show that the split is a tally of how framework authors phrase risks, not a census of what causes them, and that it moves with three things a causal claim would have to hold fixed. The first is the corpus: on the repository's own printed counts, and if no existing risk was recoded, the 722 coded risks added between its first and second versions leaned human by about twelve points, enough by themselves to close most of the gap. The second is the coding rule: each risk receives exactly one entity, as presented in the source sentence, with joint causation placed in a residual bin, and the first version itself warned that bias risks its sources meant to lay at developers' feet were worded as the model's doing. The third is domain composition: 15 of 24 subdomains are coded at least 60 percent one way, so the aggregate is mostly a mixing weight. Reweighting the subdomains moves each share by under three points, which is a robustness the author credits, although it roughly halves the gap; removing a single domain moves each share by up to ten points, two of the seven single-domain removals reverse which one leads, and recoding the bias and discrimination subdomains as the first version said their sources intended reverses it narrowly. Incident databases, misuse studies and expert surveys give different splits, each for a structural reason. The paper offers a protocol for an attribution census grounded in actual causation and states what the existing tally can and cannot support.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 99 papers by Pranay Mahendrakar, or his ORCID record.