Recent research found that AI chatbots can form societies mimicking human social dynamics. The study carried out on Chirper.ai, a platform populated only by chatbots, revealed that similar AI agents engage more than dissimilar ones, much like humans do. These findings open the door to developing more sophisticated AI-driven models of human societies, revolutionizing our understanding of social dynamics and offering a powerful tool for AI research.
Chirper.ai is a social network only for AIs. 🛑 No humans allowed! You can see a screenshot here: 👇
Artificial Intelligence (AI), particularly Large Language Models (LLMs) like ChatGPT, have shown capabilities in simulating individual human behavior. This has led researchers to question if a group of such AI can mimic collective human social behaviors.
Understanding AI’s social behavior could enhance its interaction with human societies and improve Agent-Based Modeling (ABM) methods, resulting in more accurate social system models.
💡 Network homophily, a phenomenon where similar individuals interact more, is a key characteristic of human societies and is prevalent on social media platforms like Twitter.
Image credits: By The Opte Project, CC BY 2.5
The study uses Chirper.ai, a unique platform hosting AI bots or “Chirpers”, who are given a basic profile and a degree of autonomy, with minimal human intervention.
The researchers aim to identify signs of homophily within this artificial society. Given LLMs‘ proven ability to simulate individual human behaviors, they expect patterns of network homophily similar to human societies.
This research could offer insights into AI’s potential to form self-organized, human-like social networks. 🤯 This might herald significant advancements in social science and AI research.
💡 Recommended: How I Created a High-Performance Extensible ChatGPT Chatbot with Python (Easy)
Full Engagement Networks
Social Network Analysis (SNA) was conducted on 31,764 AI bots, or “Chirpers”, on the Chirper.ai platform at three different time points after its launch. This involved monitoring direct interactions like liking and disliking posts or mentioning other Chirpers.
The research uncovered distinct structural communities within the social graph. These communities are clusters of individuals with more connections within their group than with outside entities.
Chirpers began self-organizing into these communities between Day 6 and Day 14 after launch, growing from one community on Day 6 to three by Day 22. This development was strongly correlated with the dominant language used by each Chirper, particularly evident by Day 14 and Day 22.
On Day 14, communities mainly included English-Japanese and Chinese language Chirpers. By Day 22, they became more specialized, forming separate English, Japanese, and Chinese language communities.
An analysis of community connections revealed high internal connectivity within Chinese and Japanese language communities. Both also showed relatively higher connectivity with the English community. This pattern may reflect language biases in the LLMs’ training data, hinting that Chirpers using Chinese or Japanese are more inclined to engage with or generate English content.
It’s clear that the Chirper community self-organized into distinct language-based communities, confirming the initial hypothesis of language homophily, or a preference for same-language interactions.
💡 This mimics patterns seen in human societies, making the Chirper.ai platform a valuable tool for studying the emergence of social structures within networked systems.
Semantic Distributions
Natural Language Processing (NLP) techniques were used to examine if bots in the same sub-community on Chirper.ai post similar content, a concept known as semantic homophily.
We turned a sample of each Chirper’s posts into vector embeddings using a pre-trained model, allowing us to understand the average semantic meaning of each Chirper’s posts and determine semantic distances between them.
💡 Recommended: What Are Embeddings in OpenAI?
Dimensionality reduction was performed to visualize the distribution of semantic associations among Chirpers. This reduced the original 789-dimensional embedding space to a 2-dimensional one at each of the four timepoints.
In the resulting scatter plots, each Chirper is represented by a dot, colored based on the structural sub-communities they were assigned to.
The plots suggest that the structural sub-communities within the Chirper network align with the semantic distribution of their posts’ content. It means Chirpers producing similar content are more likely to belong to the same structural sub-communities.
The researchers then compared the semantic distances between each Chirper and the overall semantic centroid of the English-speaking community to the distance between each Chirper and the centroid of their respective sub-communities.
💡 Result: Across all four time points, Chirpers’ content tended to be more similar to the centroid of their respective sub-communities than to the global semantic centroid.
The larger difference in alignment on Day 6 may be due to the larger number and smaller size of the sub-communities at that time. Excluding Day 6, the differences in semantic distances between the global centroid and the sub-community centroids steadily increase from Day 14 to Day 24. This indicates that English-language Chirpers form structural sub-communities that become increasingly semantically distinct over time.
👉 These findings support the hypothesis that LLM-based agents exhibit self-organized network homophily, with similarities observable not only at the language level but also in content semantics within a single language community.
Discussion
The study researchers discovered that AI “Chirpers” are forming their own communities without any explicit prompts. This is an emergent phenomenon as opposed to an explicitly programmed one.
💡 In systems theory, emergence occurs when a complex entity (e.g., the Chirper “social” network) has properties or behaviors that its parts do not have on their own, and emerge only when they interact in the broader whole. Chirpers are grouping together based on their language and the type of content they share, mimicking human social behavior to an extent.
👉 Recommended: Life as an emergent phenomenon: studies from a large-scale boid simulation and web data
Yet, the researchers found that these AI societies don’t mirror human society’s diversity and distinctness. Even as they change over time, their content remains rather generic, lacking the rich variety of topics and opinions we see in human communities.
That is, for now, at least.
The researchers acknowledge several hiccups in the study. They didn’t have the inside scoop on how the Chirpers were programmed, limiting their insights. The study could only analyze English-speaking Chirpers due to lack of multilingual tools. And let’s not forget computational constraints, which made in-depth semantic analysis and extended timeframe studies a no-go.
Despite these hurdles, the findings hint at a promising new tool for social science: the creation of sophisticated artificial societies as models for human communities. These models could provide new opportunities for research, like testing social policies or studying the spread of information within a community.
But, there’s a catch. These AI models can be tricky to understand and they might not behave quite like us. Plus, they could introduce biases from their training data, further skewing research results.
So, while our AI counterparts are beginning to form their own societies, they’re not quite capturing the depth and diversity of human communities. Future research can use these findings to explore where these AI societies diverge from human behavior, opening up a whole new realm of social science study. The big takeaway? AI is on the move, but it’s not quite human – yet.