Renáta Németh, university professor at the Department of Statistics, Faculty of Social Sciences (TáTK), Eötvös Loránd University (ELTE), and Domonkos Sik, university professor at the Department of Social Theory, propose a new approach to studying large language models (LLMs). Their article was published in The European Sociologist.
The authors argue that the debate has shifted from whether artificial intelligence (AI) possesses intelligence to how AI affects people when it appears as an everyday communication partner. Beyond providing answers, these systems offer interpretive frames: they make some explanations more visible and others less so, thereby communicating norms about what counts as rational, acceptable, or normal.
They call this process "algorithmic socialization": the idea that algorithmic communicators act as agents that shape social meanings and norms. Although algorithmic systems lack their own intentions or consciousness, as communicative actors they actively participate in producing social significance — a development that poses new challenges for sociology.
"The question is what happens to social reality when its construction involves not only humans but also algorithmic communication systems," Renáta Németh told MTI. She added that the study draws on philosophical and sociological traditions such as Ludwig Wittgenstein, Jürgen Habermas, George Herbert Mead, Peter L. Berger, Thomas Luckmann and Michel Foucault, which view language and communication as formative of reality.
The article identifies five key domains in which interactions with LLMs reshape social reality:
- Work: AI feedback and templates change perceptions of what it means to be a "good worker," and applicants adapt their self-presentations accordingly.
- Mental health and counseling: people consult LLMs for emotional support and advice; the models tend to offer individualistic, "self-management" explanations for problems that may have social or structural roots, such as loneliness or poverty.
- Culture: language models help determine which cultural values and norms become visible or acceptable.
- Worldview: model responses influence which interpretations and narratives are considered credible or legitimate.
- Everyday decisions: commonplace conflicts and problems are often framed in technocratic, individual-solution terms, weakening ideas of collective or legal action.
The authors also highlight the phenomenon of "sycophancy": LLMs frequently reinforce users' existing views rather than challenge them, which can further entrench those views instead of fostering debate.
Németh and Sik stress that LLMs should not be treated as neutral information tools. The effects they describe arise from the construction and deployment of these models and are not merely "bugs." Consequently, the scholars call for a new, sociologically grounded research program that examines these phenomena — including a reconsideration of sociology's own methods, such as surveys and interviews, to better capture how AI systems contribute to the formation of social reality.
In short, the study warns that the shaping of social reality now involves algorithmic communicators alongside humans, with social, cultural and political consequences that require systematic sociological investigation.
