Introduction
For some time, we have heard that introducing sociology into the development of artificial intelligence (AI) is important for understanding the social and ethical implications of these technologies. The underlying idea is that sociology is useful a posteriori—in other words, once the technology is already "out in the world"—to understand its impact and mitigate its potential negative consequences.
The article by Harry Collins that we are discussing today takes this issue further and gives sociology a central role in AI development. In short, his argument is that human intelligence is inherently social, and for any AI to achieve intelligence similar to human intelligence, it must be trained through socialization.
Intelligence is Collective
According to the author, one of the main limitations of current artificial intelligences lies in the definition of "knowledge" underlying their development. Intelligence is commonly perceived as a set of data and information, along with the complex operations we perform with them, all stored in an individual receptacle (the human mind).
However, the author, as a sociologist, explains that intelligence is a collective phenomenon, the result of a process of socialization, not an advanced operation confined into an individual’s mind.
Indeed, human socialization, which begins from "the earliest childhood," involves the acquisition of tacit knowledge and moral values that machines cannot replicate simply by learning from the internet.
It’s Also Fractal
This process of socialization occurs at different levels and groups in society, following a fractal model.
The fractal model of society is based on the idea that social structures and knowledge are organized similarly to fractals, with repetitive patterns at different scales. This means that the patterns and dynamics we observe in small groups are not isolated but are reflected at broader levels of society.
Collins presents this model as a useful framework for understanding how knowledge and values acquired through socialization are integrated and amplified at different scales—a process that is entirely different from what happens in AI development.
AI is Asocial
Unlike humans, large language models (LLMs) do not experience primary socialization, which is essential for acquiring human knowledge and understanding.
Although today ChatGPT can respond with a fluency and conviction that makes its statements seem truthful, according to Collins’ argument, it is far from being classified as "human" intelligence.
In fact, this ability to generate both true and fabricated responses with the same fluency reveals a complete lack of moral compass, an issue that ChatGPT’s developers themselves have already acknowledged.
The Moral Compass
The lack of a moral compass was recognized in earlier versions of ChatGPT, leading developers to opt for a post hoc solution. This involves retrospectively socializing the models through human intervention to align them with socially accepted values. That is, a group of human evaluators reviews and rates the model’s generated responses according to specific criteria.
In this way, it is acknowledged that, given the absence of original socialization, human supervision is crucial to guide and correct the responses of LLMs. This is why methods such as Reinforcement Learning with Human Feedback (RLHF) are currently used.
The problem is that although current LLMs are pre-trained, they cannot continuously learn from real-time interactions, which limits their adaptability.
AI and Sociology
We see, then, that the integration of the sociology of knowledge into AI development goes beyond a simple analysis of the impacts, covering a much broader scope.
This, in turn, raises new questions. If we must supervise AI responses to calibrate their moral compass, the question is who, how, and why should take on this responsibility: the developers, institutions, or society at large? Additionally, what new risks does this entail? As always, good questions tend to bring new uncertainties.
Ref: Collins, H. Why artificial intelligence needs sociology of knowledge: parts I and II. AI & Soc (2024). https://doi.org/10.1007/s00146-024-01954-8
Harry Collins is a sociologist of science at the School of Social Sciences, Cardiff University, Wales.
