Whose Risk Appetite? Mapping the Hidden Biases Behind AI Financial Decisions
Artificial Intelligence & Machine Learning News
As artificial intelligence moves deeper into financial research, portfolio construction and investment decision-making, an important question is emerging: when an AI system recommends taking risk, whose risk tolerance is it actually expressing?
That is the central question behind “Whose Risk Appetite? Behavioural Biases, Risk Preferences and a Brain Map of Financial Decisions in Large Language Models,” an ambitious new paper by Arman Khaledian, Alireza Kargarzadeh, Nariman Khaledian, Navid Parvini and Sid Ghatak. The September 2026 paper was accepted to the International Behavioural Finance Conference at Chicago Booth.
The study stands out for bringing together behavioral finance, experimental economics and mechanistic AI interpretability. Rather than simply asking whether large language models make sensible financial decisions, the authors investigate what actually drives those decisions—and whether apparent risk preferences survive careful experimental controls.
Their findings are striking.
In financial allocation experiments, seemingly irrelevant details—including whether an investment is labeled “A” or “B,” where it appears among the available choices and how a question is phrased—can substantially influence model behavior. Indeed, the raw choices of every open model tested tended to follow answer labels rather than the underlying investments.
The paper’s most important contribution may be its effort to separate genuine economic preference from these presentation effects. The authors distinguish between a model’s preference for a particular answer label and its preference for whichever option appears first. Testing every choice across all four combinations of label and position reveals that these effects can offset one another, meaning that simply reversing two options may not adequately remove the bias.
That has significant implications for researchers attempting to measure risk preferences in AI systems. What looks like risk aversion, risk seeking or responsiveness to an investor mandate may sometimes be an artifact of experimental design.
Qwen3 provides a particularly revealing example. Its choices initially appeared responsive to a client’s investment mandate. After the authors separated label and position effects, however, the remaining economic preference was much smaller: roughly eight percentage points of expected return compared with an effect of roughly 200 points associated with the answer letter, falling to about one point when neutral labels were used.
The research becomes especially compelling when the authors move beyond observing behavior and examine the model’s internal representations.
Using techniques analogous to recording, lesion studies and stimulation in neuroscience, the researchers construct what they describe as a functional “brain map” of financial decision-making inside Qwen2.5. Their experiments suggest a progression through the network: surface characteristics become readable early, concepts such as sentiment, sector information and asset risk emerge in intermediate layers, while the eventual answer forms much later.
One result captures the importance of this distinction particularly well. The model encoded asset riskiness with 96% accuracy, yet answered correctly only about half the time.
In other words, knowing and deciding are not necessarily the same thing for an AI system. A model may contain a surprisingly sophisticated representation of financially relevant information without reliably translating that representation into an economically coherent decision.
The paper also succeeds in moving from diagnosis toward practical implementation. The authors propose four governance gates for financial AI: specify, validate, calibrate, and release and monitor. The underlying principle is sensible and increasingly important: the risk policy governing an AI system should be explicitly determined by the accountable institution rather than inadvertently inherited from a model, prompt or interface.
The authors are appropriately cautious about the scope of their findings. Their open-weight experiments involve relatively small models, and they explicitly identify testing larger models as an important next step. This restraint strengthens the paper. The authors do not confuse an intriguing experimental result with a universal claim about artificial intelligence.
“Whose Risk Appetite?” ultimately offers more than another catalogue of LLM biases. It presents a framework for thinking about what it actually means for an AI system to make a financial decision.
As AI becomes increasingly embedded in investment workflows, evaluating whether a model understands financial concepts will not be enough. Institutions will also need to determine whether its decisions respond consistently to economic information, remain stable under irrelevant changes in presentation, and faithfully reflect the risk mandate they are intended to implement.
Khaledian and his co-authors make a thoughtful and original contribution to that emerging field. Their central question is deceptively simple, but it may become one of the defining questions surrounding AI-driven finance: before trusting a machine to make decisions involving financial risk, we should understand whose risk appetite the machine is actually expressing.
