Normative Equivalence in Human-AI Cooperation: Behaviour, Not Identity, Drives Cooperation in Mixed-Agent Groups

The introduction of artificial intelligence (AI) agents into human group settings raises essential questions about how these novel participants influence cooperative social norms. While previous studies on human-AI cooperation have primarily focused on dyadic interactions, little is known about how integrating AI agents affects the emergence and maintenance of cooperative norms in small groups. This study addresses this gap through an online experiment using a repeated four-player Public Goods Game (PGG). Each group consisted of three human participants and one bot, which was framed either as human or AI and followed one of three predefined decision strategies: unconditional cooperation, conditional cooperation, or free-riding. In our sample of 236 participants, we found that reciprocal group dynamics and behavioural inertia primarily drove cooperation. These normative mechanisms operated identically across conditions, resulting in cooperation levels that did not differ significantly between human and AI labels. Furthermore, we found no evidence of differences in norm persistence in a follow-up Prisoner’s Dilemma, or in participants’ normative perceptions. Participants’ behaviour followed the same normative logic across human and AI conditions, indicating that cooperation depended on group behaviour rather than partner identity. This supports a pattern of normative equivalence, in which the mechanisms that sustain cooperation function similarly in mixed human-AI and all human groups. These findings suggest that cooperative norms are flexible enough to extend to artificial agents, blurring the boundary between humans and AI in collective decision-making.

💡 Research Summary

**
The paper investigates whether the presence of an artificial intelligence (AI) agent in a small human group alters the social norms that sustain cooperation. While prior work on human‑AI cooperation has largely focused on dyadic interactions and has documented a “machine penalty” (lower cooperation with AI partners), little is known about how norms emerge and persist when an AI joins a group of humans. To fill this gap, the authors conducted an online experiment using a repeated four‑player public goods game (PGG) followed by a one‑shot Prisoner’s Dilemma (PD).

Each experimental group consisted of three human participants and one computer‑controlled “bot.” The bot was framed either as a human or as an AI (the “label” manipulation) and followed one of three pre‑programmed contribution strategies: (1) unconditional cooperation (always contributes the full endowment), (2) conditional cooperation (matches the group’s average contribution from the previous round), or (3) free‑riding (contributes nothing). The label and the strategy were the only differences across conditions; all other aspects of the interaction (rounds, endowments, payoff structure, visibility of individual contributions) were held constant.

The study employed a 2 × 3 factorial design (label × strategy) with 236 participants (final N = 236 after exclusions). Participants played ten rounds of the PGG, then a PD with a simulated partner whose behavior was always cooperative, and finally completed norm‑elicitation questionnaires measuring perceived social appropriateness of contribution levels, injunctive and descriptive norm expectations, trust, fairness, and group cohesion. The experiment was preregistered (AsPredicted #234846) and conducted on the oTree platform with participants recruited via Prolific.

The authors preregistered three main hypotheses: (H1) AI‑labelled groups would show lower overall cooperation (the “algorithm aversion” effect) and the bot’s strategy would modulate cooperation (unconditional > conditional > free‑rider); (H2) norms established in the PGG would persist more strongly in subsequent PD decisions for human‑label groups than for AI‑label groups.

Contrary to expectations, the label manipulation had no statistically significant impact on contributions in the PGG. Across all three bot strategies, average contributions and the temporal dynamics of cooperation were virtually identical in human‑label and AI‑label conditions. The only robust patterns observed were classic conditional cooperation (participants increased contributions when others contributed more) and behavioral inertia (participants tended to repeat their previous contribution). Differences among the bot strategies were small and did not translate into meaningful changes in overall group cooperation.

In the follow‑up PD, participants’ choices did not differ by label or bot strategy, indicating that the norms learned (or reinforced) during the group interaction did not persist differently across conditions. Post‑experiment surveys showed that participants’ perceptions of what contribution levels were socially appropriate, as well as their injunctive and descriptive norm expectations, were highly consistent regardless of whether the fourth player was labeled human or AI.

Statistical analyses employed mixed‑effects models with random intercepts for groups and participants, testing for main effects of label, strategy, round, and their interactions. Robustness checks that excluded participants who expressed doubts about the manipulation (≈19 % of the sample) yielded the same pattern of null label effects.

The authors interpret these findings as evidence for “normative equivalence”: the mechanisms that sustain cooperation—reciprocity, conformity to observed group behavior, and inertia—operate similarly whether the group includes an AI‑labeled member or not. In other words, the identity cue (human vs. AI) is overridden by the observable behavioral signals that participants use to infer norms. This challenges the assumption that AI agents inevitably suffer from a machine penalty in group contexts and suggests that designing AI agents to behave in a norm‑consistent manner may be more important than making them appear human‑like.

Methodologically, the study contributes a rigorous group‑level experimental paradigm to the human‑AI cooperation literature, combining behavioral economics games with norm‑elicitation measures and preregistration. The use of a hidden bot whose strategy is unknown to participants mirrors real‑world scenarios where AI behavior may be opaque, strengthening ecological validity.

Implications for practice include the recommendation that AI systems intended to operate in collaborative teams should prioritize transparent, norm‑aligned actions (e.g., consistent contribution levels) rather than focusing solely on anthropomorphic cues. Future research directions proposed by the authors involve testing higher‑presence AI agents capable of communication, longer interaction horizons, and cross‑cultural variations to identify conditions under which normative equivalence might break down.

Overall, the paper provides compelling empirical evidence that in repeated public‑goods settings, cooperation is driven by group behavior and shared expectations, not by the categorical identity of a partner. This advances our understanding of how social norms can extend to artificial agents and informs the design of AI that can be seamlessly integrated into human collective decision‑making.