💰 quantitative finance

Bounded Normative Equivalence in Human-AI Cooperation: Group Behaviour, Not Partner Labels, Predicts Cooperation under Anonymous Aggregate Feedback

In anonymous group settings with aggregate feedback, an AI label does not significantly alter cooperative behavior or norm persistence compared to a human label, as group contribution history and individual past actions are the primary drivers of cooperation, demonstrating a "bounded normative equivalence" where identity cues are diluted by the informational structure.

Original authors: Nico Mutzner, Taha Yasseri, Heiko Rauhut

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Nico Mutzner, Taha Yasseri, Heiko Rauhut

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Great Group Game: When Robots Join the Circle

Imagine you are part of a team working on a big project, like building a massive sandcastle together. You have a bucket of sand (your resources), and the goal is to see how much the whole group can build if everyone chips in. This is the heart of a field called behavioral economics, which studies how people make decisions when their own interests clash with what's best for the group. For decades, scientists have known that humans are tricky: we often want to free-ride (let others do the work while we enjoy the castle), but we also tend to copy our friends. If everyone else is helping, we help too; if they slack off, we do too. This is called "conditional cooperation."

Now, picture a new twist: what if one member of your team isn't a person at all, but a robot or an AI? As artificial intelligence becomes a regular part of our lives, we have to ask a big question: Does the fact that a teammate is a machine change how we play the game? Do we trust them less? Do we feel less pressure to be fair? This study dives into that exact mystery, exploring whether our social rules—our unwritten agreements on how to behave—break down when a robot joins the party, or if we treat the group the same way no matter who (or what) is sitting in the chair.

The Experiment: Humans, Bots, and a Secret Label

To find the answer, the researchers set up a digital playground. They gathered 236 real people and put them into groups of four to play a game called the "Public Goods Game." Think of this as a repeated round of the sandcastle challenge. In each round, every player got 100 tokens (like play money) and had to decide how many to put into a shared pot. If they put money in, it grew and was split back among everyone. If they kept it, they only had their own stash.

Here is the secret sauce: In every single group, three players were real humans, and the fourth was a computer program (a bot). The researchers played a little trick on the humans. Sometimes, they told the group that the fourth player was a human. Other times, they said it was an AI. But in reality, the bot followed one of three strict scripts:

The Super Helper: Always put in all 100 tokens.
The Mirror: Put in whatever the average of the group put in last time.
The Free-Rider: Never put in a single token.

The humans didn't know which script the bot was running, and they didn't see who put what in the pot. They only saw the total amount the whole group contributed. After ten rounds of this, the humans played one final, quick game with a partner to see if their habits stuck around.

The Big Surprise: The Label Didn't Matter

The researchers had a hunch that people would treat the AI differently. They thought that if a teammate was labeled "AI," people would be less cooperative, maybe feeling less guilty about free-riding or less trusting of the machine. They called this the "differentiation effect."

But the results were a total plot twist. The label didn't change anything.

Whether the group thought the fourth player was a human or an AI, the humans played the game in the exact same way.

The Numbers: The difference in how much people contributed was so tiny it was basically zero. The researchers used a special math test to prove that any difference between the "Human" groups and the "AI" groups was smaller than 5 tokens (which is just 5% of the 100-token starting amount). In other words, the groups were statistically identical.
The Real Driver: What actually mattered wasn't the label of the fourth player. It was what the whole group had done in the previous round. If the group contributed a lot last time, the humans contributed more this time. If the group slacked off, the humans slacked off too. The humans were reacting to the group's behavior, not the identity of their teammate.

Even the bot's strategy didn't make a huge splash. Whether the bot was a Super Helper, a Mirror, or a Free-Rider, the humans didn't drastically change their behavior based on that either. The group's collective mood was the boss, not the specific actions of the robot.

The Aftermath: Habits Stick, Labels Don't

To see if this "AI vs. Human" confusion had any long-term effects, the researchers asked the participants to play one final game immediately after the main event. They wanted to see if the "AI" groups had learned to be less cooperative than the "Human" groups.

Again, the answer was no. The habits the humans formed during the group game carried over perfectly, regardless of whether they thought they were playing with a robot or a person. Their beliefs about what was "fair" and what others "should" do were also the same in both groups.

Why This Happens: The "Blurry Photo" Effect

So, why didn't the AI label change things? The paper suggests it's because of how the information was shared. The humans only saw the total score of the group, not who contributed what. It's like looking at a blurry photo of a crowd; you can see the crowd is moving, but you can't point to one specific person and say, "That's the robot!"

Because the humans couldn't pin the actions on the AI, the "AI" label lost its power to trigger special feelings like distrust or moral disengagement. The group acted as a single unit, and the humans responded to the group's rhythm. The researchers call this "Bounded Normative Equivalence." It means that under these specific conditions (anonymous, group-level feedback), the rules of cooperation work the same way whether a robot is in the mix or not.

What This Means (and What It Doesn't)

This study suggests that in situations where we can't easily tell who did what, we might be more willing to cooperate with AI than we thought. It challenges the idea that we always treat machines as "outsiders" who don't deserve our trust.

However, the authors are careful to say this doesn't mean robots are exactly the same as humans in every situation. This "equivalence" is "bounded," meaning it has limits. If the feedback showed exactly who did what, or if the group was made up mostly of robots, the results might be totally different. But for now, in a group where everyone's actions are blended together, the humans played by the same rules, whether their teammate was flesh-and-blood or code.

Technical Summary: Bounded Normative Equivalence in Human–AI Cooperation

Problem Statement
The rapid integration of autonomous artificial intelligence (AI) agents into human group settings raises critical questions regarding the stability of cooperative social norms. While existing literature establishes that humans often treat AI as social actors, a "machine penalty" frequently emerges in dyadic interactions, where cooperation with AI is lower than with humans due to algorithm aversion or the perception that AI is less sensitive to social sanctions. However, it remains unclear whether this differentiation persists in small-group settings where interactions are repeated, feedback is aggregated, and individual contributions are not identifiable. The central research question is whether the perceived identity of a group member (human vs. AI) alters cooperative behavior and norm-related outcomes when the interaction structure, incentives, and feedback mechanisms are held constant.

Methodology
The study employed a between-subjects online experiment utilizing a $2 \times 3$ factorial design.

Participants: A final sample of 236 participants was recruited via Prolific.
Experimental Design: Participants played ten rounds of a linear Public Goods Game (PGG) followed by a one-shot Prisoner's Dilemma (PD). Groups consisted of three human participants and one computer-controlled agent (bot).
Manipulations:
1. Agent Label: The bot was framed as either "Human" or "AI."
2. Bot Strategy: The bot followed one of three predefined, undisclosed strategies:
  - Unconditional Cooperator: Contributed the full endowment (100 tokens) every round.
  - Conditional Cooperator: Matched the group's average contribution from the previous round.
  - Free-Rider: Contributed zero tokens every round.
Feedback Structure: Crucially, participants received only aggregate group feedback (total contribution and individual payoff) after each round. Individual contributions of specific group members were not visible, creating attribution ambiguity.
Measures: Primary outcomes included contribution levels in the PGG, cooperation/defection choices in the PD, and post-task measures of social appropriateness, empirical expectations (what others do), and injunctive expectations (what others think one ought to do).
Analysis: The authors utilized linear mixed-effects regression models to analyze PGG contributions and logistic regression for PD choices. They also conducted a Two One-Sided Tests (TOST) equivalence test to determine if the difference between human- and AI-labeled conditions was smaller than a smallest effect size of interest ( $\pm 5$ tokens, or 5% of the endowment).

Key Results

Absence of Label Effect: Contrary to the hypothesis of a "differentiation effect" (algorithm aversion), the AI label did not significantly reduce cooperation. The difference in contributions between human-labeled and AI-labeled groups was statistically non-significant ( $b = 1.09, p = .738$ ).
Equivalence Confirmation: The formal equivalence test (TOST) confirmed that contributions in the AI and human conditions were statistically equivalent within the bounds of $\pm 5$ tokens.
Dominance of Group Behavior: Individual cooperation was primarily predicted by:
- The aggregate contribution of the group in the previous round (conditional cooperation).
- The participant's own contribution in the previous round (behavioral inertia).
- A gradual decline in cooperation over time (round effect).
  These behavioral regularities were consistent across all label and bot strategy conditions.
Bot Strategy Impact: The specific strategy of the bot (unconditional, conditional, or free-riding) did not produce statistically significant differences in group contributions, likely due to the signal dilution caused by the presence of two other human members and the aggregate feedback structure.
Norm Persistence and Perceptions:
- PD Outcomes: There were no systematic differences in norm persistence (cooperation in the one-shot PD) based on whether the previous group contained an AI or human.
- Normative Expectations: Participants' perceptions of social appropriateness, empirical expectations, and injunctive expectations were closely aligned across conditions.
- Survey Data: While general trust and fairness perceptions were similar, participants in the AI condition reported slightly lower alignment with group expectations ( $p = .031$ ). However, this did not translate into measurable behavioral differences in contribution levels.

Key Contributions and Significance
The paper introduces the concept of "Bounded Normative Equivalence." This term describes a state where, under specific informational conditions, the observed cooperation patterns, norm perceptions, and norm persistence are statistically indistinguishable between groups containing AI agents and those containing only humans.

Theoretical Contribution: The study challenges the universality of the "machine penalty" and algorithm aversion found in dyadic literature. It suggests that in collective settings with anonymous aggregate feedback, the informational structure (attribution ambiguity and signal dilution) overrides the ontological category of the partner. Participants respond to the functional logic of the group (reciprocity and conditional cooperation) rather than the identity of the partner.
Boundary Conditions: The authors argue that this equivalence is "bounded" by the setting's informational structure. Because individual actions were not identifiable, the specific identity cues that typically trigger differentiation were diluted. The findings imply that cooperative norms can extend to groups including artificial agents when individual contributions are not transparent.
Implications for AI Integration: The results suggest that in environments where individual contributions are opaque (e.g., large-scale online communities, smart grids, or resource allocation systems), transparent and norm-consistent behavior by AI agents may foster social integration as effectively as human-like labeling. The study posits that "socialness" in such systems may be an emergent property of interaction rules and feedback loops rather than solely a property of the agents' perceived minds.

Limitations and Scope
The authors explicitly note that these findings are specific to the experimental conditions: short-term interactions, scripted (non-adaptive) agents, and anonymous aggregate feedback. The equivalence observed may not hold if:

Individual contributions are identifiable (removing attribution ambiguity).
AI agents are adaptive, communicative, or socially present.
The group composition shifts to an AI-majority setting (where the "human buffer" is removed).
Long-term reputation building is possible.

The paper concludes that while cooperative norms appear robust to AI labels in the studied context, future research must investigate when and how this equivalence breaks down in more complex, transparent, or adaptive human-AI interactions.