Cooperation Through Indirect Reciprocity in Child-Robot Interactions

Social interactions increasingly involve artificial agents, such as conversational or collaborative bots. Understanding trust and prosociality in these settings is fundamental to improve human-AI teamwork. Research in biology and social sciences has identified mechanisms to sustain cooperation among humans. Indirect reciprocity (IR) is one of them. With IR, helping someone can enhance an individual’s reputation, nudging others to reciprocate in the future. Transposing IR to human-AI interactions is however challenging, as differences in human demographics, moral judgements, and agents’ learning dynamics can affect how interactions are assessed. To study IR in human-AI groups, we combine laboratory experiments and theoretical modelling. We investigate whether 1) indirect reciprocity can be transposed to children-robot interactions; 2) artificial agents can learn to cooperate given children’s strategies; and 3) how differences in learning algorithms impact human-AI cooperation. We find that IR extends to children and robots solving coordination dilemmas. Furthermore, we observe that the strategies revealed by children provide a sufficient signal for multi-armed bandit algorithms to learn cooperative actions. Beyond the experimental scenarios, we observe that cooperating through multi-armed bandit algorithms is highly dependent on the strategies revealed by humans.

💡 Research Summary

The paper investigates whether the well‑known mechanism of indirect reciprocity (IR) – where helping others improves one’s reputation and thereby elicits future help – can be transferred to child‑robot interactions, whether robots can learn to cooperate based on children’s strategies, and how the choice of learning algorithm influences the resulting human‑AI cooperation. The authors combine laboratory experiments with formal modelling. In the first experimental phase, groups of three children (aged 6‑9) and one robot play a repeated “resource‑allocation” game. Each child decides whether to give a token to another child; the decision is publicly observed and translates into a “kindness score” that serves as a reputation signal. The robot watches these actions and chooses either to cooperate (give a token) or defect (withhold a token) in each round. Results show that children behave as if a higher reputation yields social benefits, and the robot’s propensity to cooperate rises sharply when it perceives a child’s reputation as positive.

In the second phase, the robot is equipped with a multi‑armed bandit (MAB) learner to infer the children’s underlying strategy. Three classic MAB algorithms are tested: ε‑greedy, Upper Confidence Bound 1 (UCB1), and Thompson Sampling. The robot updates its action‑value estimates after each round based on the observed reputation scores and immediate payoffs. UCB1 consistently achieves the highest cooperation retention (≈92 % of rounds) because its confidence‑bound term balances exploration and exploitation effectively, allowing rapid identification of the most cooperative child and stable exploitation thereafter. ε‑greedy performs well only when the exploration rate ε is sufficiently high; low ε leads to premature convergence on sub‑optimal (defective) actions, while high ε incurs unnecessary exploration costs. Thompson Sampling’s performance hinges on the prior distribution: when the prior matches children’s true behavior, learning is fast; mismatched priors cause slow convergence and lower cooperation levels.

The theoretical contribution extends the classic IR framework of image scores and social norms. The robot’s reward function is defined as R = α·reputation + β·immediate payoff, where α and β weight long‑term reputation benefits against short‑term gains. Bayesian inference on experimental data reveals that cooperation is robust when α/β ≥ 3, indicating that a strong emphasis on reputation is necessary for the robot to sustain cooperative behavior.

A key concept introduced is “strategy signal”: the pattern of giving and refusing displayed by children provides the robot with informative feedback about the social environment. When children’s behavior is consistent, the robot can infer the optimal policy with few samples; when children act randomly or inconsistently, learning efficiency drops dramatically. This finding underscores that the predictability of human behavior critically determines which learning algorithm will succeed in human‑AI teams.

Overall, the study demonstrates three main points: (1) indirect reciprocity extends to child‑robot groups, (2) children’s reputation‑based strategies supply sufficient signal for MAB algorithms to learn cooperative actions, and (3) the choice of algorithm and the weighting of reputation in the reward function dramatically affect the emergence and stability of cooperation. The authors suggest future work to test diverse age groups, cultural contexts, and more sophisticated reinforcement‑learning architectures, as well as to apply the IR‑based approach to real‑world domains such as education, healthcare, and disaster response where human‑AI collaboration is increasingly critical.