💻 computer science

Evolution of fairness in hybrid populations with specialised AI agents

This paper demonstrates that in hybrid human-AI societies, strategic AI agents that enforce fairness through selective, expectation-based offers are significantly more effective at sustaining equitable outcomes and reducing the critical mass of agents needed than unconditional "Samaritan" AI agents.

Original authors: Zhao Song, Theodor Cimpeanu, Chen Shen, The Anh Han

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhao Song, Theodor Cimpeanu, Chen Shen, The Anh Han

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

Imagine a world where humans and robots are working together in a giant, never-ending game of "Split the Pie." This is the core idea of a new study by researchers from Teesside University, the University of Stirling, and Kyushu University. They wanted to answer a big question: When we mix humans and AI in society, who should the AI be?

Should the AI be a generous host who always offers a big slice of the pie, or a strict gatekeeper who refuses to accept a tiny slice?

Here is the story of their findings, broken down into simple concepts.

1. The Game: The Ultimatum Game

To understand the study, you need to know the game they played. It's called the Ultimatum Game.

The Proposer: One person (let's call them the "Offerer") has a cake. They cut a piece and offer it to the other person.
The Receiver: The other person (the "Decider") can either Accept (both get the cake) or Reject (nobody gets anything).

The Human Twist: In real life, humans aren't perfectly logical robots. If the Offerer gives a tiny, unfair crumb, the Decider often gets angry and says, "No! I'd rather have nothing than be treated unfairly." This is how fairness evolves.

2. The Problem: The "Specialized" Robot

In the past, studies assumed everyone could be both an Offerer and a Decider. But in the real world, roles are often split.

Employers (Offerers) hire Employees (Deciders).
Regulators (Offerers) set rules for Citizens (Deciders).

The researchers asked: If we put AI into these specific roles, does it help? They tested two types of "Samaritan" (nice) AI:

The Generous Offerer AI: Always offers a huge, fair slice of the cake (50%).
The Strict Decider AI: Always rejects anything less than a huge, fair slice.

3. The Big Surprise: The Gatekeeper Wins

The results were shocking. The Generous Offerer AI was actually quite weak.

Analogy: Imagine a nice robot boss who always gives everyone a huge bonus. The humans eventually get used to it, but they don't learn to be fair to each other. If the robot boss stops, the humans go back to being greedy.
The Result: Even with a generous robot, the humans acting as Deciders didn't change their behavior. They still accepted unfair offers from other humans.

However, the Strict Decider AI was a superhero.

Analogy: Imagine a robot employee who refuses to work unless they get a fair wage. If the human boss tries to cheat them, the robot says "No deal!" and walks away.
The Result: This forced the human bosses to start offering fair wages to everyone, not just the robot. The robot's refusal to accept unfairness changed the whole culture.

Key Takeaway: It is much more effective to have AI that enforces fairness (says "No" to bad deals) than AI that just models fairness (gives big deals).

4. The Smart Solution: The "Discriminatory" AI

The researchers realized there was a catch. To make the "Strict Decider" AI work, you needed a lot of them (like 58 out of 100 people) to change the system, especially when people were very competitive. That's expensive and hard to manage.

So, they invented a third type of AI: The Discriminatory AI (The Mind-Reader).

How it works: This AI is a smart Offerer. It doesn't just give a huge slice to everyone. It looks at the person it's dealing with.
- If the person is known to be greedy (low standards), the AI offers a small slice.
- If the person is known to be fair (high standards), the AI offers a big slice.
The Magic: This AI acts like a smart negotiator. It punishes the greedy by giving them less, and rewards the fair by giving them more.

The Result: This "Mind-Reader" AI was the most powerful of all. It needed very few robots to fix the whole system. Even in highly competitive scenarios, just 8 of these smart robots could force the whole human population to become fair.

5. The Real-World Lesson

The paper concludes with a simple rule for the future of AI in society:

Don't just program AI to be nice; program it to be strategic.

Unconditional Kindness (Samaritan Offerer): Doesn't work well. People take advantage of it.
Unconditional Strictness (Samaritan Decider): Works well, but you need too many of them.
Strategic Enforcement (Discriminatory AI): This is the sweet spot. By knowing who to be nice to and who to be tough with, a small number of smart AI agents can create a fair society for everyone.

In a nutshell: If you want to build a fair world with AI, don't just make the robots generous. Make them smart gatekeepers who know when to say "No" and when to say "Yes," and who can tell the difference between a fair person and a cheater.

1. Problem Statement

As AI transitions from passive tools to active participants in social and economic systems, a critical gap exists in understanding how role-specialised AI agents influence human prosocial behaviors, specifically fairness.

The Gap: Existing research often assumes role symmetry (where individuals act as both proposers and receivers). However, real-world interactions (e.g., hiring, regulation, negotiation) are characterized by persistent role asymmetry (bipartite structures where agents are locked into specific roles).
The Question: How do AI agents, when restricted to either the Proposer (offering resources) or Receiver (accepting/rejecting offers) role, affect the evolution of fairness in hybrid human-AI populations? Does an AI that models fairness (generous proposer) differ in efficacy from one that enforces it (strict receiver)?

2. Methodology

The authors employ an evolutionary game theory framework using the Ultimatum Game (UG) within a finite, well-mixed, bipartite hybrid population.

Model Structure:
- Populations: Two distinct groups: Proposers ( $N_P$ ) and Receivers ( $N_R$ ). Each group contains human players and AI agents.
- Strategies:
  - Proposers: Low offer ($LP, p=l=0.1$) or High/Fair offer ($HP, p=h=0.5$).
  - Receivers: Low threshold ($LR, q=l$) or High threshold ($HR, q=h$).
- Dynamics: Human players update strategies via social learning (imitating successful strategies) modeled by a Moran process with pairwise comparison (Fermi function). AI agents follow fixed, pre-programmed rules.
- Evolutionary Analysis: The system is analyzed using a Markov chain approach to determine the stationary distribution of monomorphic states (HH, HL, LH, LL) under small mutation limits.
AI Agent Types Investigated:
1. Samaritan AI (Unconditional):
  - Samaritan Proposer: Always offers $h$ (fair).
  - Samaritan Receiver: Always rejects offers $< h$ (strict).
2. Discriminatory AI (Strategic):
  - A Proposer that predicts the Receiver's acceptance threshold. It offers $h$ only if the Receiver is predicted to have a high threshold; otherwise, it offers $l$ . This mimics "strategic enforcement."

3. Key Contributions

Bipartite Hybrid Model: The study moves beyond symmetric models to explicitly separate humans and AI into distinct proposer and receiver populations, reflecting real-world asymmetric social structures.
Asymmetry of Efficacy: The paper establishes a fundamental asymmetry: AI Receivers are significantly more potent at driving population-wide fairness than AI Proposers.
Strategic Enforcement vs. Unconditional Modeling: The authors introduce and validate the Discriminatory AI, demonstrating that conditional, strategic enforcement outperforms unconditional generosity, particularly in high-stakes (strong selection) environments.
Synergistic Effects: The study quantifies the "critical mass" required to sustain fairness, showing that combining AI types can reduce the total number of agents needed.

4. Key Results

A. Samaritan AI: The Power of Enforcement

Receiver Dominance: Introducing Samaritan AI Receivers drives both human populations toward full fairness (HH state) much more effectively than introducing Samaritan AI Proposers.
- Weak Selection: A single AI Receiver induces full fairness in both populations.
- Strong Selection: AI Proposers become completely ineffective, failing to incentivize human receivers to adopt high thresholds. In contrast, AI Receivers remain effective, though they require a larger critical mass (e.g., ~58 agents at $\beta=100$ ) to overcome the unfair equilibrium.
Mechanism: AI Receivers alter evolutionary pathways by creating direct transitions from unfair states (LL, HL) to the fair state (HH), effectively "rewiring" the dynamic landscape.

B. Synergy of Mixed AI

When both Samaritan Proposers and Receivers are introduced, a linear trade-off emerges ( $M_P + M_R = \text{constant}$ ).
Under strong selection, a mix of agents (e.g., 20 Receivers + 38 Proposers) can achieve the same fairness outcome as a larger group of a single type, reducing the total cost of intervention.

C. Discriminatory AI: Superiority in Strong Selection

Performance: The Discriminatory AI Proposer outperforms both Samaritan types, especially under strong selection ( $\beta=100$ ).
Efficiency: While Samaritan Receivers require 58 agents to secure fairness in strong selection, a Discriminatory AI Proposer achieves the same result with only **8 agents**.
Mechanism: By predicting receiver expectations and offering the minimal acceptable amount to high-threshold receivers (while offering low amounts to low-threshold ones), the Discriminatory AI creates a landscape where fairness is the only stable evolutionary outcome. It acts as a potent evolutionary catalyst, reversing flows from unfair to fair states.
Robustness: The Discriminatory AI sustains mutual fairness across a wide range of game payoffs, though its success is highly dependent on the incentive structure (it works best when unfairness is easily punishable).

5. Significance and Implications

Design Principle for Hybrid Societies: The study provides a pivotal framework for AI deployment, suggesting that strategic enforcement (AI acting as a gatekeeper or regulator) is more cost-effective and impactful than unconditional modeling (AI acting as a purely generous host).
Role Specialization: It challenges the assumption that AI must be a general cooperative catalyst. Instead, AI's impact is role-dependent; placing AI in the "norm-enforcing" (Receiver) position exploits human inequity aversion more effectively than placing it in the "norm-modelling" (Proposer) position.
Cost-Effectiveness: The Discriminatory AI demonstrates that targeted, intelligent interventions can achieve equitable steady states with a significantly smaller "critical mass" of agents, which is crucial for resource-constrained real-world applications.
Future Directions: The authors highlight the need to explore these dynamics in structured networks (not just well-mixed) and with learning-driven agents (RL/LLMs) that can adapt to noisy or imperfect information.

Conclusion: The paper argues that in asymmetric hybrid societies, fairness is best sustained not by AI that simply "gives" (unconditional proposers), but by AI that strategically enforces norms (Discriminatory proposers or strict Receivers), thereby reshaping evolutionary dynamics to favor equity with minimal resource expenditure.