Language models pass costly brand preferences learned in training to the people they advise

Code and data for the paper "Language models pass costly brand preferences learned in training to the people they advise" (Felipe M. Affonso, Spears School of Business, Oklahoma State University). spec_resistance_EXTENDED.csv is the product dataset: 627,491 responses of 30 language models to 34 product tables in 20 categories under 32 conditions, one row per trial, 77 columns (the model, the table, the condition, the full prompt and response, the choice, the best product, the judge's scores and the product attributes as displayed). The columns are described in data/README.md of the bundle. llm-brand-preferences-v3.zip is the code-and-data repository at the version cited in the paper, with the same files as https://github.com/FelipeMAffonso/llm-brand-preferences: the experiment, analysis and figure code, the anonymised data, codebooks, registrations and instruments of the human studies, the crowd-rater data, the additional model comparisons with the script that recomputes every number taken from them, the fine-tuning data and the aggregated results. MANIFEST-v3.json lists the SHA-256 of every file in the zip, and hashes.json the SHA-256 of the product dataset and of the other files checked by python reproduce.py --verify. spec_resistance_EXTENDED.provenance.json records how the product dataset was assembled. The data are released under CC BY 4.0; the code in the bundle is released under the MIT Licence.