Apology Without Complete Repair: Divergent Restoration of the Gricean Maxim of Quality in Large Language Models
Polyglot: Journal of Linguistics, Literature, and Language Education · 12.08.2026 · 2 (1) · 39-53
DOI: 10.64850/polyglot.v2i1.414 ↗Kurz gefasst
Fachartikel in der Zeitschrift Polyglot: Journal of Linguistics, Literature, and Language Education, online erschienen am 12.08.2026. Verfasst von Li Weili und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
In human–AI interaction, users generally expect large language models (LLMs) to provide information that is accurate and reliable, an expectation that functions as a cooperative one. When a model produces an error, this expectation is broken, and the model typically responds by apologizing, explaining, or revising its answer in an attempt to renegotiate cooperation. Whether such a post-error response genuinely restores cooperation, however, remains unclear. Drawing on a corpus of 360 Chinese responses elicited from 30 standardized error scenarios, this study examines six consumer-facing LLM systems and reports three main findings. First, a post-error response is not a simple apology but a combination of relational, evidential, corrective and deferral resources. Second, in a 90-response core directly concerned with the maxim of quality, 41 responses achieved complete repair, 48 achieved partial repair and one failed. Complete-repair rates differed significantly across the six consumer-facing systems under the observed interface conditions, although only one pairwise contrast remained significant after Holm correction. Third, at the pooled descriptive level, error acknowledgment, relational softening and evidential boundaries were associated with complete repair, but no single surface form was sufficient; same-task contrasts show that repair depends on whether the old answer is functionally invalidated and the available evidence is used, rather than on relational softening, abstention or a promise of later verification alone. These findings suggest that cooperative stance and cooperative substance may diverge in human–AI interaction: relational softening, including apologies and expressions of thanks, cannot by itself establish that cooperation has been restored.
Zitieren
APA
Weili, L., & Shuo, Z. (2026). Apology Without Complete Repair: Divergent Restoration of the Gricean Maxim of Quality in Large Language Models. Polyglot: Journal of Linguistics, Literature, and Language Education, 2(1), 39-53. https://doi.org/10.64850/polyglot.v2i1.414
Vancouver
Weili L, Shuo Z. Apology Without Complete Repair: Divergent Restoration of the Gricean Maxim of Quality in Large Language Models. Polyglot: Journal of Linguistics, Literature, and Language Education. 2026;2(1):39-53. doi:10.64850/polyglot.v2i1.414
BibTeX
@article{Weili2026,
author = {Weili, Li and Shuo, Zhao},
title = {Apology Without Complete Repair: Divergent Restoration of the Gricean Maxim of Quality in Large Language Models},
journal = {Polyglot: Journal of Linguistics, Literature, and Language Education},
year = {2026},
volume = {2},
number = {1},
pages = {39--53},
doi = {10.64850/polyglot.v2i1.414}
}
EUVIMED