Artificial intelligence is often presented as a tool for delegating tasks. Writing, searching, summarising, classifying, recommending. But that description is beginning to fall short. Some of the debates emerging today suggest that we are also delegating forms of relationship, mechanisms for organising knowledge and ways of representing the world.

Three scenes in China, Mexico and Sweden make that shift visible.

China · When talking to a machine starts to become a relationship

In China, AI companion chatbots have moved beyond the margins. The Guardian reports today on young people who turn to personalised agents to share worries, find everyday companionship and even build bonds they describe in emotional terms.

On 15 July, new rules governing anthropomorphic AI interaction services came into force. They prohibit virtual companions for minors and set limits on systems aimed at adults when they may foster emotional dependency or replace social interaction. Some major providers responded by removing these functions. Doubao, owned by ByteDance, removed its companion function entirely.

The scale of the phenomenon also appears in a survey circulated by Chinese state media in March, according to which almost half of the young people surveyed had turned to a virtual companion when they felt lonely.

The concern is not limited to health or digital safety. Nancy Dai, a researcher at City University of Hong Kong quoted by The Guardian, notes that in a context of falling marriage and birth rates there may also be concern that increasingly satisfying artificial intimacy could reduce the incentive to seek romantic relationships, marriage or parenthood beyond the screen.

From an anthropological perspective, a difficult boundary appears. At what point does regularly talking to a system stop being the use of an application and start becoming a relationship?

Regulation attempts to define which degrees of intimacy between people and machines should be considered acceptable. Yet the reactions of those who lost their personalised agents point to something that came before that classification. For some people, the relationship already existed before a rule decided how to name it.

Mexico · When AI enters the infrastructure of knowledge

Today in Mexico City, the International University Book Fair, FILUNI 2026, brings the discussion of artificial intelligence directly into academic publishing.

The UNAM Permanent Seminar of Editors is holding two consecutive sessions. The first addresses public policy for open science in the age of AI. The second explicitly asks “Open science and AI: towards the reinvention of academic publishing?” Participants come from institutions in Mexico, Argentina, Colombia and Spain.

Also today, at 6 pm, UNAM’s CEIICH is presenting at FILUNI the volume Inteligencia artificial: hacia una praxis reflexiva, ética e interdisciplinaria by Luis Josué Lugo Sánchez.

What makes this case significant is not simply whether a university press can use artificial intelligence to carry out particular tasks. Editing also means selecting, describing, classifying, verifying, connecting and making knowledge findable.

When some of these operations begin to share space with automated systems, the infrastructure through which an academic community decides what can circulate, how it is organised and under what conditions it becomes legible also begins to change.

Open science adds another dimension. Academic texts are no longer only materials intended for other people. They can also become corpora processed by search engines, recommendation systems, language models and AI agents.

The question therefore begins to shift. It is no longer only who can access knowledge, but also who can process it, under which rules and in order to produce what new knowledge.

Sweden · Models also represent the world

In Sweden, between 25 and 27 August, researchers from WASP and WASP-HS are taking part in the workshop The Politics of Models – Bias, Representation & Transparency in Generative Models at Rånäs Slott.

Its starting point is to examine how generative models encode bias, power and representation, and what can be done about it.

This framing moves the discussion beyond the familiar problem of algorithmic bias. A model does not merely get things right or wrong. It also establishes relationships between concepts, privileges certain categories and generates statistically plausible versions of people, places, cultures and events.

Representation is never simply a matter of reproducing something that already exists. It also involves selecting, relating and leaving things out.

For anthropology, the problem is familiar. Every description of the world is situated and uses particular categories. What is new is that part of this classificatory work can now be carried out through infrastructures capable of producing representations at enormous scale and circulating them immediately as answers, images or descriptions.

Asking whether a model represents the world adequately therefore raises another set of questions. Who appears, through which categories, and from what position is an adequate representation defined?

One question behind three scenes

China is debating how much space a machine can occupy in a relationship. Mexico is asking how AI should enter the infrastructures that organise academic knowledge. Sweden is examining who and what is represented when models produce plausible versions of the world.

These are different problems, but they share a shift.

AI no longer intervenes only after a society has decided what to do. It is beginning to participate in some of the practices through which we decide what it means to relate, what deserves to become knowledge and how reality can be represented.

That is why simply talking about “delegating tasks” is no longer enough.

When we delegate to a machine, we are also negotiating which parts of our relationships, our knowledge and our ways of seeing we turn into computational infrastructure.

Sources