Exploring LLM-Mediated Trust-Related Perception Ratings Using a Corporate-Reputation Benchmark

Assessing trust-related perceptions of organisations is difficult where public evidence is unstructured and traditional reputation scores collapse multidimensional judgements into a single rating. This historical exploratory study examines whether a large language model (LLM), guided by IPIP-derived trust-related items, can translate information available within a consumer ChatGPT environment into structured perception signals and whether those signals associate with a public corporate-reputation benchmark. The historical eight-item prompt was applied to the 100 companies in the 2023 Axios Harris Poll 100 using GPT-4, GPT-4o, and a repeated GPT-4 run. The Axios Harris Poll is treated as a related corporate-reputation comparator, not as a ground-truth measure of trust or counterparty risk. The models were not asked to reproduce Axios scores, rankings or an overall reputation rating; they rated separate trust-related items that were aggregated subsequently by the researchers. This design reduces the plausibility of simple direct benchmark reproduction, although indirect influence from prior exposure to Axios or related reputation information cannot be excluded. Results showed measurable but model-dependent association with the benchmark. GPT-4o produced stronger linear and rank association, whereas the GPT-4 runs showed lower absolute error on the normalised scale. Repeat-run differences were smaller than cross-model differences, and the aggregate scores were compressed toward the middle of the scale. The findings support the feasibility of low-cost, structured trust-related perception signals, but not validation of the complete theoretical trust construct, superiority over simpler text-processing baselines, source-level evidential auditability, exact regeneration of the historical ChatGPT outputs, generalisation beyond highly visible US corporations, or prediction of realised counterparty outcomes. Operational use would require explicit model, prompt, source, calibration, provenance, and human-governance controls.