EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Plausible but Unfaithful? Auditing LLM-Generated Explanations in Cultural Recommender Systems

Markos Konstantakis, George Caridakis

Machine Learning and Knowledge Extraction · 06.11.2026 · 8 (11) · 367

DOI: 10.3390/make8110367 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Machine Learning and Knowledge Extraction, online erschienen am 06.11.2026. Verfasst von Markos Konstantakis und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Large language models (LLMs) increasingly recommend cultural venues and tell visitors why, but such explanations can be plausible without being faithful. In a preregistered counterfactual audit, four commercial LLMs recommended venues in three Greek destinations to visitor profiles based on the ACUX typology, from fixed 20-venue catalogues. One profile attribute was changed at a time, and every change was measured against the variation between identical requests (4032 planned calls, 4030 valid). Pooled across attributes, cited attributes were associated with larger changes in recommendations (counterfactual faithfulness score 0.26–0.42), but within attributes the association was near zero (between −0.02 and 0.04): explanations name the kinds of attribute that matter, not what changed a given recommendation. Demographic attributes were almost never cited and had small effects. Explanations often went beyond the venue’s catalogue entry (29–83%, mostly atmosphere and crowding), whereas fabricated visitor preferences were rare (4–14%); a grounding instruction removed most such additions in three models. When a poorly matching venue was imposed under a neutral request, two models seldom disclosed the mismatch (3% and 13%); when the request allowed disagreement, all four disclosed it in 83–100% of cases. Model comparisons based on LLM-judge codes with limited human agreement are exploratory. LLM explanations should be treated as persuasive text and grounded in the information the system holds.

Zitieren

APA

Konstantakis, M., & Caridakis, G. (2026). Plausible but Unfaithful? Auditing LLM-Generated Explanations in Cultural Recommender Systems. Machine Learning and Knowledge Extraction, 8(11), 367. https://doi.org/10.3390/make8110367

Vancouver

Konstantakis M, Caridakis G. Plausible but Unfaithful? Auditing LLM-Generated Explanations in Cultural Recommender Systems. Machine Learning and Knowledge Extraction. 2026;8(11):367. doi:10.3390/make8110367

BibTeX
@article{Konstantakis2026,
  author = {Konstantakis, Markos and Caridakis, George},
  title = {Plausible but Unfaithful? Auditing LLM-Generated Explanations in Cultural Recommender Systems},
  journal = {Machine Learning and Knowledge Extraction},
  year = {2026},
  volume = {8},
  number = {11},
  pages = {367},
  doi = {10.3390/make8110367}
}
Kostenlos starten →