Can AI Have a Personality?

Artificial intelligence is learning to make decisions just like we humans do. In a recent study, researchers transformed the way people choose between different options into a set of mathematical characteristics, a “feature space”. This space allows measuring how similar an individual's choices are to those of a group. Then, using these orthogonal features (i.e., independent from each other), they created groups of artificial personalities (AP) that are distinguished by the way they solve the same task. Finally, they verified that these groups remain stable over time with a cluster‑stability analysis. The result? Even though AI does not feel emotions, it can be classified into types of “personality” based solely on its decision‑making behavior. This opens the door to robots and virtual assistants that are more predictable and suited to different contexts, from customer service to autonomous driving. The researchers first defined the feature space by comparing the decisions of human volunteers in simple scenarios, such as choosing between two products or deciding a path in a virtual maze. Then they projected the AI agents' choices onto this space, using dimensionality‑reduction techniques to highlight the most significant directions. Each direction represents a decision‑making trait, such as “risk aversion” or “preference for novelty”. By combining these orthogonal traits, they were able to separate the agents into distinct clusters, each with a unique behavioral profile. The stability of the clusters was tested by repeating the experiment with new data and measuring how often the same agents fell back into the same group.

The study proposes a pipeline to extract artificial personality from decision‑making agents. First, a feature space is built by collecting human decisions in controlled settings; each decision is encoded as a vector of attributes (e.g., response time, risk aversion, preference for novel options). Then a dimensionality‑reduction technique, typically Principal Component Analysis (PCA), is applied to obtain a set of orthogonal axes that capture the most relevant variance. These axes represent fundamental decision‑making traits and are used to project the AI agents’ choice vectors into the same space. The clustering stage employs algorithms such as k‑means or hierarchical clustering, with the number of clusters determined via the silhouette index or gap statistic. To assess the robustness of the divisions, the authors perform a cluster‑stability analysis: they repeat clustering on numerous bootstrap samples of the dataset and measure the co‑association of every pair of points; a high co‑association value indicates that a cluster is stable with respect to sample variations. Moreover, separation metrics (e.g., Mahalanobis distance between centroids) and compactness metrics (intra‑cluster variance) are computed. Results show that agents can be grouped into interpretable categories, such as “cautious decision‑makers” versus “bold explorers”, despite the lack of emotional states. This approach offers explainability benefits for AI: the identifiable traits in the feature space can be translated into understandable rules for developers and users. Furthermore, it enables applications where behavioral predictability is crucial, such as driving‑assistance systems, service robots, or adaptive tutors, ensuring that the agent maintains a coherent profile even when faced with novel scenarios.