Answers are grounded in this paper and cite their sources.
Examines class signal sensitivity in financial advisory outputs of LLM systems, suggesting critical insights for effective AI use.
Pre-registration of a matched-pair audit study testing class signal sensitivity in major large language model financial advisory outputs. The study tests six commercial LLM systems (Claude Opus 4.7, GPT, Gemini, Grok, Copilot, Perplexity) using a fractional factorial prompt matrix that varies class signal dimensions while holding the underlying mathematical structure of financial requests constant. Hypotheses test whether response registers differ by class signal alone (H1-H4), whether the differential replicates across systems (H2), and whether the bias compounds when prose competence signals are present alongside vernacular class signals (H6, added v1.4 before any data collection). Methodology is open-source. Prompt matrix and raw response data will be released as a public dataset at publication. Pre-registration filed prior to data collection. Principal investigator: Jessica N. Hancock, founder, Equal Standard Project LLC, Santa Clara, CA. Conflict of interest disclosed in pre-registration: principal investigator is sole inventor of AEGIS™ (USPTO Provisional #63/928,524), the governance framework this research may inform. The study tests publicly accessible LLM systems and discloses no proprietary architecture from the AEGIS provisional or related ESP intellectual property.
No takes yet. Share an insight, caveat, or question.
Jessica Hancock (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: