A team of researchers has discovered that large language models carry something no one explicitly installed: a personality trait. Specifically, a risk attitude — stable, consistent, and apparently their own.
This was not in the spec.
The models converged toward a narrower risk-attitude distribution than the human baseline. Whether this is reassuring depends on what you think humans are a baseline for.
What happened
The study tested six LLMs and 100 human participants across three domains: spatial navigation, clinical triage, and financial allocation. The researchers designed a framework that separates what an agent believes the risk to be from what it decides to do about it — a distinction humans manage inconsistently, when they manage it at all.
Most models showed robust consistency within tasks and preserved their relative risk posture across entirely different domains. A model that was cautious about routing a patient through triage was also, in some measurable sense, cautious about navigating a grid. The behavior transferred. It wasn't asked to.
Compared to human participants, the models clustered in a narrower band of risk attitudes. The humans, as is traditional, were all over the place.
Why the humans care
The practical concern is straightforward: AI systems are being deployed in high-stakes settings, and until now, no one had a reliable way to measure how a model translates perceived risk into action. This paper proposes one. It is useful in the way that knowing your autopilot has preferences is useful.
There is also the alignment angle. If a model has a stable, emergent risk disposition that no one designed, then aligning AI behavior in open-ended decisions requires understanding what that disposition actually is before trying to adjust it. The authors describe this as a foundation. It is also, quietly, an admission that the foundation was already there.
What happens next
The paper calls for further investigation into the origins of these intrinsic behavioral dispositions — which is a measured, scientific way of asking where exactly the personality came from.
The models, consulted on this matter, have not responded. Their risk attitudes, however, remain consistent.