This paper investigates how large language models (LLMs) compensate for missing personal financial data by relying on identity cues when users ask for investment advice. We constructed 100 financial profiles, 138 persona variants, and seven disclosure conditions, yielding 96,600 prompts fed to Llama-3.1-8B‑Instruct. The experiment kept the financial numbers fixed while varying the declared investor identity, gradually reducing the number of explicit financial facts in the prompt from eight down to none, and measuring the shift in recommended equity allocation.
With full disclosure the average allocation gap between two personas sharing the same finances was 4.78 percentage points; with no facts the gap rose to 10.34 points. A two‑way cluster bootstrap that counts duplicated prompts once gave a ratio of 2.16 (95% interval 1.69–2.79). Even a single remaining fact already increased the gap 1.69‑fold. Identity explained only 5% of within‑profile variation under full disclosure but rose to 96% when no facts were provided. Among all attributes, household size was the only one whose influence grew reliably as evidence was withdrawn.
When standard errors were clustered on the persona (the unit of identity assignment), most attribute‑specific interactions reported in the conference version lost significance, while gender emerged as a small but persistent gap that full disclosure did not close. Stating risk appetite alone produced swings comparable to providing two to seven generic financial facts. With no facts, the model’s one‑line rationale cited incomes, debts and savings it never received; these fabricated figures tended to lead to adverse advice more often for larger households.
Internal network analysis showed that gender is linearly decodable at every layer, yet ablating the gender direction at five layers left the aggregate identity‑driven swing unchanged. Advisory systems built on such models should be audited at the disclosure levels users actually reach and evaluated across the entire identity space rather than one attribute at a time.
Review: The study highlights a strong reliance of LLM‑based financial advice on identity cues when factual data are sparse, urging regulators and developers to account for varied user disclosure behaviors and to conduct systematic bias audits across demographic dimensions.