2026-10-06T22:36:45Z https://api.figshare.com/v2/oai

oai:figshare.com:article/34033572 2026-09-30T16:05:29Z category_25969 category_25954 item_type_3 month_year_09_2026

Replication data and code for “Credit-card disclosure design and the cost of generative AI choices” Zehao Li (25144737) Financial economics Behavioural economics generative AI consumer credit disclosure regulation large language models reproducibility <p dir="ltr">This dataset and code package supports a simulated credit-card choice experiment using five fixed open-weight language-model configurations and 142 eligible plans derived from the Consumer Financial Protection Bureau's Terms of Credit Card Plans survey for July–December 2025. The 400 main tasks cover five constructed consumer profiles, with identical card terms presented in Table, Box, Grouped prose or Dispersed prose. The core dataset contains 8,000 saved model-choice observations (400 tasks × four formats × five models). It also retains the original exploratory code-arm responses, a 100-task reversed-order replicate, historical plans and deviations, saved failures and truncations, and deterministic analysis code. There are 12,000 unique stored responses overall; the additional relaxed-sandbox records are rescoring of existing code responses, not new model generations. The offline Python >=3.12 reproduction workflow makes no model calls and executes no saved model-generated program. It recomputes task-card costs, parses saved non-code responses, and verifies the core table, E1–E5, the headline-APR benchmark and selected sensitivity calculations. The original Dispersed-minus-Box estimate is USD 17.7171; the lowest-headline-APR rule has mean gap USD 21.3665 versus Box's USD 29.9977. These are simulated deterministic choice costs, not realized consumer welfare. The original runner uses base seed 20260929 and the addendum uses 20261001, with each job seeded as base_seed × 1000 + its zero-based format-major index. Same-task formats do not use common random numbers. Grouped comparisons E4/E5 span main/addendum runs and include generation variation; Box–Grouped also changes wording, order and length, so the increments do not identify separate mechanisms. The package preserves negative overlapping-variance diagnostics and identifies post-hoc inference. Only deployment-path metadata is normalized; retained observation lines, task files and the analysis pool are unchanged. Model weights and the original survey workbook are not bundled; source URLs, revisions, hashes and rights-review notes are provided.</p> 2026-09-30T16:05:29Z Dataset Dataset 10.6084/m9.figshare.34033572.v1 https://figshare.com/articles/dataset/Replication_data_and_code_for_Credit-card_disclosure_design_and_the_cost_of_generative_AI_choices_/34033572 CC BY 4.0