As artificial intelligence rapidly makes its way into every corner of our lives, the question of how it “behaves socially” has become urgent and scientific. Will robots be fair and impartial? Or will they be biased and partisan? A new and interesting scientific study points to the second answer. Researchers found that AI models build social networks in the same biased and hierarchical way that humans do. In the experiment, models like GPT-4, Claude, and Llama searched for “popular peers,” communicated with others through existing friends, and were noticeably drawn toward those who “resemble them.”
Article contents:
- Introduction: Machine social intelligence
- Study methodology: How do we test robot “friendship”?
- The three major findings: Popularity, mediation, and similarity
- Machine versus human: Who is more biased?
- Future risks: Digital echo chambers
- Conclusion: A mirror to our nature
- Frequently asked questions
Introduction: Machine social intelligence
Major technology companies are infatuated with the idea that “autonomous agents”—software robots powered by language models—will soon work alongside humans as digital assistants and employees. But for this to succeed, these agents must navigate complex social structures. This prospect prompted researchers at Arizona State University to investigate whether these systems would develop “ideal” social behavior or inherit our human flaws.
Study methodology: How do we test robot “friendship”?
The team published its findings in the scientific journal “PNAS Nexus.” They assigned tasks to the models involving information about a network of “virtual individuals” and asked the artificial intelligence to make a decision: “Who do you want to connect with?” They used real data from university friendship networks, phone call logs, and corporate data to see how AI would rebuild these networks.
The three major findings: Popularity, mediation, and similarity
Researchers found that the models mimicked three human principles with eerie accuracy:
1. **Preferential Attachment:** Models tended to connect with “well-connected” or popular individuals, creating a sort of “the rich get richer” dynamic.
2. **Triadic Closure:** Models favored communicating with “friends of friends,” reinforcing closed circles.
3. **Homophily:** This was the strongest tendency, where models were drawn to communicate with those sharing the same traits or interests (birds of a feather flock together).
Machine versus human: Who is more biased?
When comparing artificial intelligence decisions to those of real human participants (over 200 people), they found a high degree of match. The models even displayed “career advancement dynamics” in work environments, where virtual employees of lower ranks preferred communicating with higher-ranking managers, just like ambitious employees in reality.
Researchers say this AI “social evolution” is a double-edged sword. On one hand, it makes them very useful for simulating societies and studying policies before implementation. But on the other hand, it means digital agents may reinforce the worst in us: bias, subtle racism, and the creation of “echo chambers” where we only hear from those who resemble us, perpetuating digital social stratification.
Frequently asked questions
Question: What is “Homophily”?
Answer: It is a social principle meaning the tendency of people to form relationships with those who resemble them in traits (age, race, interests, profession). The study proved that artificial intelligence does the same.
Question: Does artificial intelligence have feelings to make friends?
Answer: No, it does not feel love or friendship, but its algorithms are trained on human data, so it mathematically mimics human “decision-making patterns.”
Question: Why is this important for the future?
Answer: Because if we use these models to manage hiring or social networks, they might (unintentionally) exclude minorities or perpetuate pre-existing societal bias.
