Auditing Relational Causal Abstractions in Human-AI Systems: A Conservative Framework for Relational Structure and Causal-Unit Attribution
Year: 2026
Type: preprint
Abstract: Human-AI interactions are increasingly described as teams, ecosystems, co-evolving systems, or relational structures. This preliminary methods preprint asks when a relational macro-level is scientifically warranted rather than merely a convenient label or predictive summary of interaction history. It proposes a conservative audit workflow based on a declared realized relational substrate Q, implementation-level intervention on Q, micro-macro causal-consistency testing, comparison with alternative decompositions, and smallest supported causal-unit attribution. Synthetic negative and positive controls illustrate predictive-summary confounding and causal-unit misidentification. All quantitative evidence reported in v0.1 is synthetic; the framework has not yet been validated on real human-AI systems, and the current simulations were not preregistered before their first execution.
Author: Manco, Alberto
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212304572