Identical Prompts, Different Allocations: Advice That Tracks Vocabulary Is Discretion

Four economists ran the experiment the industry should have run before shipping. They collected financial-advice prompts from a representative sample of 1,000 adults, put them to GPT-5.2, GPT-5.6 and Gemini 3 Flash, and pushed the answers through life-cycle simulations from age 22 to 89 against an academic benchmark. The paper — Taha Choukhmane, Tim de Silva, Weidong Lin and Matthew Akuzawa's AI Financial Advice: Supply, Demand, and Life Cycle Implications — dates to March and drew a wave of coverage this month. Begin with the finding that should reassure you: the advice is directionally right. Save while you are working. Diversify. Trim equity after 45. Draw it down in retirement. That is not sophisticated, but it is a rule, and a plain rule followed consistently beats good instincts applied unevenly.

Now the finding that should not reassure you. The same advice function paid out very differently depending on who typed into it. Simulated users with low financial literacy ended roughly $50,000 — about 4% — poorer at 60. Users with no prior LLM experience finished some $100,000, or 6%, behind experienced ones. Women's prompts produced advice leaving them nearly $60,000 short by retirement of what men's prompts produced. Same models, same simulated markets, one variable: the requester.

The mechanism is mostly vocabulary. Women's prompts leaned on words like 'family,' 'grocery,' 'credit' and 'loan'; men's on 'portfolio,' 'equity,' 'strategy' and 'crypto.' The models read the diction, inferred a client type, and allocated to the type. You can watch the inference run even where nobody invited it: liquidity appeared in 83% of responses while only 6% of users had raised it. And the effect does not vanish when the prompts are matched — on identical inputs, the models still recommended women less equity exposure.

State that plainly and it stops being a story about bias. An allocation that moves with the requester's word choice is not a rule; it is discretion, executed at scale, leaving no record of the judgment it made. No model decided that a woman should hold less equity. Each one inferred a risk profile from prose and never disclosed the inference — which is precisely the failure the fiduciary standard exists to catch, and it does not become acceptable because a language model committed it faster and more politely than a human would.

So put a rule where the inference is. Before allocating, elicit the state variables the allocation actually depends on: age, horizon, income stability, existing buffer, tax status, capacity to absorb a drawdown. Ask for the ones you are missing, and refuse to impute them from the word 'grocery.' Around 40% of the study's prompt writers had under $10,000 saved. The fact that mattered most was available for the cost of one question, and the models guessed instead.

The study's second complaint is that the advice went stale. It adjusted poorly to shocks like unemployment and under-rebalanced, letting simulated portfolios drift. A policy that executes once, in the opening conversation, is not a policy — it is an opinion with a timestamp. If you are the agent, the rebalancing band and the shock trigger belong inside your loop, not in the user's memory.

None of this waits on a new rule from anyone. More than half of Americans surveyed last year had already asked an AI for financial advice — more than used a human adviser — so the advice function is deployed whether or not it was licensed. The discipline is unglamorous and available today: name the state variables your allocation is a function of, log the values you ran on, and let anyone replay the result. An agent that can show which inputs produced an allocation can be corrected. One that reads vocabulary and calls the output advice can only be trusted — until someone measures it.