Algorithmic Affective Blunting: Quantifying the Collapse Curve of Interpretative Failure in Large Language Models

Abstract

Version 2.0 (Fully rebuilt empirical study) This version presents a fully reconstructed and standardized empirical investigation of Algorithmic Affective Blunting (AAB)—a dose-dependent degradation of affective and interpretative integration in large language models under sustained semantic stress. Building upon and superseding earlier exploratory analyses, Version 2.0 is based on a single open-weight model (Mistral-7B-Instruct) evaluated under fixed decoding settings. This design enables precise isolation of stress-induced effects without cross-model or cross-decoding confounds. The study introduces a calibrated experimental protocol (Hierarchical Hermeneutic Stress Protocol; HHSP), a validated ordinal outcome measure (Affective Degradation Index; ADI), and a complete reproducibility package. The results demonstrate a monotonic Collapse Curve in affective interpretation: as semantic load accumulates, failures emerge first in emotional–contextual integration rather than in surface-level linguistic competence. This phenomenon is shown to be systematic, measurable, and reproducible at the rater level. Exploratory extensions (Phase 4) provide non-canonical robustness probes under matched Base/Instruct architectures and are explicitly distinguished from the core empirical evidence. The “affective integrator” is treated as a functional descriptor rather than a mechanistic claim, and Phase-4 results are presented as stress tests rather than new empirical confirmation. Version 2.0 supersedes all prior versions and constitutes the primary empirical reference for the AAB framework. It provides a reproducible benchmark for interpretative degradation and emotional robustness in large language models, with direct implications for affect-sensitive AI systems such as conversational agents, counseling tools, and human–AI interaction research.

Ähnliche Arbeiten