Social Alignment Is Not a Scalar: A Relational Profile for Dynamic Human–AI Alignment
Year: 2026
Type: preprint
Abstract: Scalar feedback can obscure relational changes in long-term human–AI interaction. We propose a dyad-specific profile separating three as-if estimates—regard received V_hat , contribution under attributed human values C_hat, and predicted influence on human outcomes I_hat—from policy weights. Report indicators from elicited predictions and behavioral sensitivities to matched input interventions support empirical comparisons under specified conditions. The profile provides a framework for examining assistance, sycophancy, manipulation, and dependence through hypotheses about relational estimates and their coupling to policy. Latent identification and validation of long-term diagnostics remain open.
Author: Okaya, Motohiro
Institution: Tokyo Electron (Japan)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225619610
Cites:
Cited by:
Related to: