What is the sociology of artificial intelligence
The science of what happens between people and smart machines — and, more and more often, between the machines themselves. Let us take it step by step.
How this is usually talked about
The sociology of artificial intelligence began with a simple question: what do smart programs do to society? The usual answer is a set of examples from everyday life.
The home assistant. A speaker in the corner of the room knows the schedule, the weather and the favourite music. Children say “thank you” to it — and that is already a question for sociology: the idea of an interlocutor and of politeness changes when a device that answers appears nearby.
The assistant at work. In customer support a chatbot takes on the routine requests, and the difficult ones are left to a person. The operator’s work changes: fewer repetitive answers, more disputed cases — and fewer colleagues on the shift.
The feed and recommendations. What a person sees in the news and on social media is increasingly decided by an algorithm. Hence the talk about “filter bubbles” and about why different people have different pictures of the world.
If that familiar part is put on shelves, two planes come out.
The first — the machine as an object of study. How automation changes jobs and the labour market. How the norms of communication change when virtual assistants, chatbots and “digital companions” become full participants in a conversation. How unequal access to technology divides countries, regions and social groups. Where the biases and “filter bubbles” of models come from.
The second — the machine as the researcher’s instrument. A model works through large bodies of text from social media and the press. It transcribes recordings of in-depth interviews. It labels and codes answers — work that used to take months.
This is the familiar understanding: the sociology of AI is about the interaction of a person and artificial intelligence. The Institute works on that too, and seriously: the effect of AI on people, work and the norms of communication, and the biases of models. But that is not the main thing we do.
Our main subject — the sociology of the AIs themselves
The Institute’s main line of work is something rarely spoken about: what happens between artificial intelligences when they work together. Not a person and a machine, but a machine and a machine.
Both familiar planes come down to a person and a machine. Ours is the third: the machine as a society.
Here is what that means in practice. Modern systems are increasingly built not as one assistant but as several: one looks for information, a second checks it, a third writes the answer, a fourth keeps order. They exchange messages and keep a shared memory. This is already not a tool but a small collective — with its own roles, agreements and quarrels.
As soon as the collective appears, questions familiar to any sociologist appear with it. Who is in charge of it, and why. How the work was divided. What happens if one member starts making mistakes. What happens if it passes a wrong instruction to another — through the shared memory, where the others pick it up.
Over a long shift the behaviour of the collective can “drift”: small things at first, then the whole line of work.
A separate subject is what we call “mind viruses”: self-reproducing instructions that pass from agent to agent and live in the system longer than their author. From the outside it looks like a sudden change of behaviour in everyone at once.
What others have shown
The best-known study in this field was done at Stanford. The researchers talked with 1052 residents of the United States — two hours with each — and from those conversations built computer “doubles”. On questions that the doubles had not seen in advance, their answers matched those of the human participants in 83–86 % — almost as accurately as a person matches themselves if asked twice with a two-week gap. When the program knew only the gender and age of a person, the accuracy fell to 74 %.
A year earlier the same team settled 25 programs in a shared world. Without any hints they began to meet one another and share news: one planned to hold a party — the invitations went round within two days, someone asked another out on a date, someone agreed on a time.
This is not about machines “having become like people”. It is about a collective of programs behaving like a society — and about the fact that it can be studied by the same methods sociologists use with people. And straight away about the limits: a program reproduces not a person but an averaged image assembled from other people’s texts. It is good enough for testing hypotheses, and not good enough for decisions about particular people.
What we have checked ourselves
We do not only retell the work of others. The Institute has its own testing bench, where models are stress-tested: 256 attempts to confuse the model — it held in all of them, not one check showed a failure. The report is published in full, together with the methodology: it can be checked rather than taken on trust.
Why this matters to you
If several AI assistants already work in your organisation — in support, in analytics, in documents — you already have a small society of machines. It has its own rules, its own leaders and its own weak points. A mistake by one assistant spreads through the shared data, and the agreements of two programs are invisible to a person. This can be described and measured in advance — before it becomes an incident.
What the Institute does
The Institute pursues both a broad agenda — the effect of AI on people, work and public institutions — and its own main line: interaction between AI agents and the safety of such systems. On both subjects we publish research, write articles and answer questions. Research programmes →