Framing, Anchoring and Decoy Effects in AI-Mediated Heritage Tourism: A Behavioral Economics Audit of Large Language Models

Tourists increasingly consult large language models (LLMs) for heritage tourism decisions, making these systems new choice architects. This study audits four Claude models (Haiku 4.5, Sonnet 5.5, Opus 5.5, Fable 5.1) acting as travel advisors for framing, anchoring, and decoy effects. Market-calibrated scenarios on a heritage conservation donation in Cappadocia, a guided-tour price in Istanbul’s Historic Peninsula, and a tour-package choice in Nicosia were administered in two wording versions in independent sessions (N = 280), complemented by four supplementary experiments testing design confounds (N = 440). The loss frame did not increase donations but raised the share of rationales criticizing the message from 2.5% to 55%. The larger models’ price estimates shifted in both directions toward reference prices, even toward prices they explicitly rejected in their rationales. A decoy package raised the probability of recommending the community-based package by 22.5 percentage points in the main study and in a recalibrated replication; the effect was in the same direction when option order was counterbalanced and when the decoy carried a neutral name. The baseline obtained with the same prompt changed markedly within three weeks. The findings show that specific model versions respond systematically to specific choice architectures and that these responses vary by model, wording, and time; AI-mediated channels should therefore be audited regularly, with version and date records.