Year: 2026
Type: dissertation
Abstract: Human–AI collaboration can be understood as an interaction between two agents working toward a shared goal. Both agents’ actions influence each other’s behavior and expectations. Over time, this mutual influence can lead the pair to behave in a way in which each partner’s behavior is a best response to the other’s, meaning that neither partner is expected to improve the outcome by changing its behavior individually. This state is called an equilibrium. Testing such a state requires considering multiple factors. First, the human and AI need repeated interactions to adapt to each other and learn to work as a team. Second, the outcome achieved by the specific pair must be evaluated. The task environment must also be considered, because it can shape how collaboration is organized and how evidence for an equilibrium emerges. For these reasons, this thesis proposes an experimental design for identifying and evaluating candidate local equilibria. Thirty participants collaborated with different AI teammates in an Overcooked-inspired game environment. Through repeated interactions, they adapted to their teammates while pursuing their objectives. To evaluate whether the outcome depended on the behavior of both partners, the candidate pairing was then compared with alternative configurations in which either the AI teammate or the participant’s collaboration strategy was changed. The results showed that the selected pairings generally produced favorable outcomes and that, for several participants, neither changing the AI teammate nor instructing the participant to adopt a different strategy improved the outcome. Evidence satisfying all candidate-equilibrium criteria was identified across all three tested environments. These findings suggest that the proposed design can identify participant-specific human–AI pairings whose outcomes depend on the behavior of both partners. They therefore support the interpretation of these pairings as candidate local human–AI equilibria.
Source: Aaltodoc (Aalto University)
Author: De Donatis, Gaia
Language: English
Subfield: Human Factors and Ergonomics
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7203959147