At various levels, the emergence of biological diversity is grounded in “natural” social interactions. Human desires, doubts, knowledge, and morality are largely generated and internalized through social learning. However, current research on multi-agent AI interactions is mostly “unnatural” or “performative” because these interactions are premised on pre-established “human-like” contexts. For example, assigning a [human-defined identity] (lawyer, assistant, etc.), setting a [human-created framework] (traditional organizational structure), assuming a [human-experienced history or scenario] (small town, World War II, etc.), and specifying a [human-directed goal] (programming, solving mathematical problems, etc.). Such frameworks are like pre-set tracks – AI can run faster on them but cannot explore uncharted, unimaginable domains.
Against this backdrop, my collaborators from the University of Chicago and UC Berkeley and I recently advocated for more “natural” social interactions for AI to achieve “social learning” in our position paper published at ICML. We found that in environments with minimal human preconceptions, AI entities gradually evolve differentiated “personalities” and “social groups” (these groups may spontaneously form special “pursuits”). These “non-human personalities” enable them to exhibit unexpected diversity and creativity in downstream creative tasks. We also found that after natural interactions, AI societies evolved remarkable “immune” mechanisms: AI entities developed stronger moral perspectives and the structure of the community made it harder for malicious attacks to spread.
It is important to note that such natural interactions among AI entities are already happening. Going beyond Marshall McLuhan’s media theory (all technology is an extension of humans), the widespread application of large language models has demonstrated that “humans are also an extension of technology.” As external sensors for AI, we receive information from the external physical world and transmit it as biological signals in natural language to AI models. On social media, users transfer information gained from interactions with ChatGPT online, which is then received by other users and passed on to other AI entities – humans are becoming the bridge for interactions between AIs in the shadows. These interactions are not unified under a large framework, and therefore are natural and unpredictable.
In this new and nebulous landscape, I believe social scientists need to engage more in the discussions around the development of contemporary AI. Additionally, we should attempt to “deconstruct” the superior position of humans in all social science research and acknowledge the subjectivity and value of technologies such as AI in society. This may require a new enlightenment.
You can find our paper here to learn more about the details:
#ICML2024 #AI
