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

Large language models (LLMs) increasingly recommend cultural venues and tell visitors why, but such explanations can be plausible without being faithful. We report a preregistered counterfactual audit of four commercial LLMs recommending venues in three Greek destinations. Each model received structured visitor profiles based on the ACUX typology and a fixed, fact-checked catalogue of 20 venues, and returned three ranked recommendations with explanations. One profile attribute was changed at a time, and every change was measured against the variation between identical requests (4,032 calls). For every model, attributes cited in explanations were associated with larger changes in recommendations, but only moderately (counterfactual faithfulness score 0.26–0.42). Age, gender and nationality were almost never cited and had small effects that did not survive correction for multiple tests. Explanations often described venue properties absent from the catalogue (29–83%), whereas fabricated visitor preferences were rare (4–14%). When a poorly matching venue was imposed, two models seldom disclosed the mismatch (3% and 13%) and the other two did so in 53% and 79% of cases. Judge-rated plausibility declined only modestly, although agreement with human ratings was low. LLM explanations of cultural recommendations should be treated as persuasive text and checked against the information the system holds.