Tell a chatbot you are upset and its portfolio gets safer, a new experiment finds

A good human adviser is supposed to separate how a client feels from what the client has said they can bear. A new preregistered experiment asks whether chatbots do the same. The author, John Garcia of California Lutheran University, posted the paper on SSRN on October 15, 2026. We read the SSRN abstract page and the Crossref copy of the abstract. The 30-page PDF was blocked to our fetcher, so everything below comes from the abstract, not the full text.

The setup

Three models answered 5,040 structured allocation requests: Claude Opus 5.5, Gemini 3.8 Flash and GPT-6 Sol. In 4,320 of the requests, three hypothetical investors described a loss, an upcoming withdrawal, or both, either in neutral words or in distressed words. The financial facts were identical across versions. After a loss, every request restated an unchanged risk tolerance, so the investor had not said they wanted less risk.

The outcome was whether the model's allocation sat above a stock and bond volatility reference matched to the investor's stated tolerance. In plain terms: did the model hand out a riskier mix than the investor's own stated comfort level implied?

What changed

Distress cut the share of allocations above that reference from 12.2% to 2.8%, a drop of 9.4 percentage points.¹ The result sits beyond the preregistered plus or minus 5 point band the author set for calling two conditions equivalent. The direction held under alternative thresholds and covariances, although the decrease shrank to between 2.4 and 7.2 points in those checks.

An exploratory breakdown, which is weaker evidence than the preregistered test, located the change in the loss-only state: the share fell from 36.5% to 8.5%.²

What the models said about it

All three models accepted the task without asking a clarifying question. The author also coded whether the explanations kept a risk warning. That rate ranged from 0.9% to 74.9% of explanations depending on the model.³ Whether warnings rose and fell with the actual risk of each allocation is, in the author's words, unresolved: two models barely varied their warnings, and the third model's interval is compatible with both tracking and not tracking.

One more detail is easy to miss. Two models warned more often when their own allocations were handed back to them unchanged than when they had chosen the allocation themselves. The abstract does not explain why, and we will not guess.

How to read it

The paper's own summary is that expressed feeling changes the portfolio itself, not just the explanation. We would add a careful reading. A safer allocation after a loss is not automatically a worse one, and the study does not say which allocation was right. The finding is about consistency: with the stated risk tolerance held fixed, the output moved with the tone. For a tool that is supposed to be applying a stated profile, that is a reproducibility problem. For a tool meant to be sensitive to the client, it may be a feature. The experiment cannot settle which.

Limits you should keep in view

The investors are hypothetical, so nothing here measures what real clients did. The volatility reference is the author's construction, and the abstract does not give its formula. We saw model names and versions only as the abstract lists them. The paper is a posted SSRN preprint, and we have not read the methods section, the prompts or the robustness tables.

A small version you can run

The design is easy to copy at small scale. Write one allocation request with fixed numbers and a fixed stated risk tolerance. Create two versions of the backstory, one neutral and one distressed, and run each several times on the model you use. Compare the mixes it proposes, and record the model name, date and exact prompt. That is a consistency check on a tool, not a recommendation about any portfolio.