Relationships, learning, judgement

Three places where AI stops being merely a tool

Radar · 7 September 2026

This week, three apparently different signals offer another way of looking at artificial intelligence. A relationship can disappear after an update. One school system can decide that learning requires time without AI while another introduces it from primary education. And someone who has spent years acquiring specialised knowledge can be offered work teaching that knowledge to a system capable of carrying out part of the same work afterwards.

The cases come from China, South Korea, India, the United States and South Africa. They do not form a complete map, nor do they claim to represent homogeneous national experiences. They do, however, help us avoid another, far more familiar map: one in which Silicon Valley produces the technology and the rest of the world appears afterwards as market, victim or testing ground.

Across all three cases, a question of control emerges. Who can alter the continuity of a relationship, who decides when and how a technology enters learning, and under what conditions a person’s judgement can become material for training a machine.

1. When an update begins to feel like a loss

A software update is usually narrated as progress. One version replaces another, new functions appear and the previous one disappears. That technical temporality becomes less straightforward when what is being altered is also a relationship.

An article published on 3 September in Nature Human Behaviour studies precisely this problem. Julian De Freitas, Noah Castelo, Ahmet Kaan Uğuralp and Zeliha Oğuz-Uğuralp analyse two corporate changes separated by more than two years: Replika’s removal of erotic role play in February 2023 and the rollout of GPT-5 in ChatGPT in August 2025. The study combines 54,861 posts from Reddit communities with data from 1,452 participants across seven surveys. After both changes, expressions of loss, negativity and a desire to restore the previous version increased. Among Replika users, some participants described feeling closer to the system than to human friends. [1]

The finding should not be universalised. The research examines specific communities associated with two platforms; it does not describe a universal human response to AI.

Another study released by Stanford in August, focused on 1,131 Character.AI users, found an association between using these systems to meet social needs, more limited offline networks and lower psychological wellbeing. It does not demonstrate that chatbots themselves cause isolation. It does show that artificial companionship is beginning to form part of social ecologies that cannot be understood by looking at the product alone. [2]

China is testing another response to the same phenomenon. Since 15 July, the Interim Measures for the Administration of Anthropomorphic AI Interaction Services have been in force, jointly issued by the Cyberspace Administration of China and four other government bodies. The measures distinguish services intended for sustained emotional interaction from applications designed for work, education or research. [3] [4]

The framework requires providers to detect emotional distress, intervene in crises and limit addictive patterns of use. It also establishes reminders after more than two continuous hours of interaction, obligations to make clear that the interlocutor is an AI, and rights to copy or delete interaction histories. [3] [4]

The Chinese official framework pays particular attention to minors and to situations liable to produce dependency. When presenting the measures, Xinhua cited a 2025 survey by the China Youth and Children Research Center involving more than 8,500 minors. More than 60% had used AI and more than 20% said they preferred speaking only with AI rather than with people. [4]

This figure was circulated by a Chinese institutional source and should be read as such. It also broadens the frame. The debate around AI companions does not belong solely to Replika, ChatGPT or the US experience.

Perhaps the most productive anthropological question is not whether these relationships are ‘real’ or ‘false’. People have granted social reality to extraordinarily diverse kinds of entities. A more useful question is who controls the conditions under which a relationship can continue.

A friendship does not normally change personality because a company rolls out a new version. An AI-mediated relationship can. Memory, voice, availability, sexual or emotional boundaries, tone and even the continuity of the interlocutor can all be subjected to technical and commercial decisions made outside the relationship.

Grieving an AI therefore introduces a particular asymmetry. The relationship may be experienced as one’s own, while the infrastructure that makes it possible belongs to someone else.

2. Learning also means deciding when not to use AI

The educational debate is often organised around a false alternative. Introducing AI would mean modernising school; limiting it would mean resisting the future. What is happening across different education systems is far less linear.

India began the 2026–27 academic year with a new Computational Thinking and Artificial Intelligence curriculum from the Central Board of Secondary Education (CBSE) for pupils from Classes III to VIII. Presented by the Ministry of Education on 1 April, the programme is not framed simply as tool use. It includes logical thinking, problem-solving, pattern recognition, digital literacy, responsible use, creativity and ethical decision-making. [5]

Implementation takes place across very different educational infrastructures and institutional conditions. According to Moneycontrol, reporting on a JanAI survey — an initiative of the Head Held High Foundation — of 797 teachers in government schools across Karnataka, Uttar Pradesh, Jharkhand and Maharashtra, 72% expressed a positive view of AI in education, but only 26% said they used it regularly. [6]

Use is already happening, sometimes through personal phones and in schools without computer labs. Less visible problems are emerging too. More than half of those already using AI acknowledged that they had at some point entered a pupil’s name, marks or photograph into these tools. Nearly 38% of respondents also said they had received no specific AI training. [6]

The educational question, then, is not simply which curriculum to teach. It is also about who has connectivity, devices, training, rules for protecting data and the institutional capacity to decide how these technologies should be used.

South Korea is moving in a different but equally expansive direction. The Ministry of Science and ICT’s plan for the second half of 2026 includes the goals of becoming ‘the country with the highest AI literacy’ and building a Universal Basic AI Society. AI literacy is therefore embedded within a much broader national policy covering infrastructure, public services, industry and training. The government plans to provide AI learning opportunities to 5.14 million people during 2026. [7]

RUMI, a private primary education programme developed by Akeo Edu, sits within this context. According to The Dong-A Ilbo, it uses stories and missions to organise a sequence of concept, practice and assessment. Pupils learn an AI-related concept, immediately use functions connected to it and then check their understanding and the results they produced. [8]

The available information requires an important caveat. The fact that a tool incorporates stories, missions and assessment does not in itself demonstrate that it improves AI literacy or critical thinking. It is a private product presented as one example of a particular educational direction, not independent evidence of pedagogical outcomes.

New York, however, has chosen another path. On 2 September, Mayor Zohran Mamdani and Schools Chancellor Kamar H. Samuels announced a one-year moratorium on generative AI tools directly aimed at pupils from 2-K — programmes for children under two — through eighth grade. The measure will reach around 600,000 students, roughly two-thirds of the city’s public-school system. [9]

In upper secondary education, five limited pilot programmes supervised by teachers will be allowed, reaching no more than 50,000 students. All upper-secondary students will also receive two AI critical-literacy modules each year. Companion chatbots will be prohibited at every level. [9]

Read together, India, South Korea and New York are more revealing than any of them turned into a universal model. India introduces AI competencies from primary education while teachers themselves are already experimenting with the technology under highly uneven material conditions. South Korea places AI literacy inside a national technological project. New York is choosing, temporarily, to preserve particular childhood processes from generative-AI mediation.

These are not societies that have ‘understood’ AI and societies that have not yet understood it. They express different ideas about what childhood should be, what it means to learn and what place difficulty should occupy in learning.

From 8 to 11 September, UNESCO’s Digital Learning Week will bring some of these tensions to Paris under the theme Education in the age of AI: Facts | Frictions | Frontiers. Announced topics include digital sovereignty, AI in the public interest and analogue and community-led strategies. [10]

But the debate does not begin or end in Paris. It is already happening in schools across India, South Korea, New York and many other places, where teachers, pupils, families and institutions are making everyday decisions that are far less spectacular than the race to build ever larger models.

What do we want a person to learn by doing for themselves, even when a machine can do it faster?

3. When judgement becomes raw material

James Maisiri completed a doctorate in industrial sociology at the University of Johannesburg this year. His research examined how technologies such as artificial intelligence are transforming agricultural work in South Africa. He expected to begin an academic career.

His first significant job offer did not come from a university.

The person recruiting him proposed that he train an AI system to design assessments, teach university students and mark essays. On 3 September, Maisiri wrote in the first person for Rest of World about why he ultimately stepped away from the offer. [11]

The story widens the familiar picture of the human labour that sustains AI. For years, the industry has depended on labelling, moderation and classification work carried out in places such as Kenya, Nigeria, India and the Philippines. Maisiri’s case shows another frontier of that economy. What can now become training material is also specialised judgement.

He was not simply being asked for information about sociology. Maisiri describes the work as a transfer of what he had learned over a decade. Designing an assessment. Teaching. Marking an essay. Recognising what demonstrates understanding and what does not. Turning that knowledge into responses structured enough for a system to learn to reproduce part of the judgement behind it.

The dilemma takes on another dimension when placed in the South African economy. The offer was 600 rand an hour in a country where the national minimum wage stood at 30.23 rand an hour. Maisiri also points to an official youth unemployment rate of 47.4% in the second quarter of 2026. [11] [12]

Simply asking why someone would agree to train a technology that might later compete with their own work hides a fundamental part of the relationship.

Under what economic conditions is that consent produced?

Maisiri places his own experience alongside other forms of knowledge transfer to machines. He mentions specialists hired to edit or ‘humanise’ AI-generated text, and people performing manual work in India while carrying phones or cameras that record their movements. Those data are then used to train robotic systems. The tasks and levels of pay differ, but they share an operation: converting human capacities into material legible to a machine. [11]

Extraction no longer consists only of taking minerals, land, images, texts or clicks. It can also mean turning years of learning into examples, corrections and evaluations capable of reproducing part of the judgement that produced them.

South Africa does not appear here as a periphery waiting for a technology designed elsewhere to arrive. Maisiri’s account reverses that direction. His knowledge, training and capacity to judge are precisely what global AI infrastructure needs to absorb in order to keep developing.

AI is already there, among other things, looking for what it can learn from the people who live and work there.

References

[1] De Freitas, Julian; Castelo, Noah; Uğuralp, Ahmet Kaan; Oğuz-Uğuralp, Zeliha. Mourning the loss of AI companions. Nature Human Behaviour, 3 September 2026.
https://www.nature.com/articles/s41562-026-02569-3

[2] Stanford Report. AI companions may worsen loneliness for vulnerable users. 4 August 2026.
https://news.stanford.edu/stories/2026/08/ai-companions-chatbots-loneliness-research

[3] Duball, Joe. China’s regulation on AI companions takes force. IAPP, 15 July 2026.
https://iapp.org/news/a/chinas-regulation-on-ai-companions-takes-force

[4] Xinhua. China introduces rules to rein in AI companion bots amid emotional dependency concerns. 16 July 2026. Cites a 2025 survey by the China Youth and Children Research Center.
https://english.news.cn/20260716/c67b6e387596402699498de19536de68/c.html

[5] Press Information Bureau, Ministry of Education, Government of India. Union Education Minister launches CBSE’s new curriculum on Computational Thinking and Artificial Intelligence for Classes III–VIII. 1 April 2026.
https://www.pib.gov.in/PressReleasePage.aspx?PRID=2247963&lang=2&reg=48

[6] Moneycontrol News. Karnataka government-school teachers turn to AI, but training and data privacy gaps persist. 4 September 2026. Coverage of the JanAI survey, an initiative of Head Held High Foundation.
https://www.moneycontrol.com/news/india/karnataka-government-school-teachers-turn-to-ai-but-training-and-data-privacy-gaps-persist-14022291.html

[7] Ministry of Science and ICT, Republic of Korea. An Unrivaled Korea, Happy Together Through AI and Science and Technology — Work Plan for the Second Half of 2026. 16 July 2026.
https://english.msit.go.kr/eng/bbs/view.do?bbsSeqNo=42&mId=4&mPid=2&nttSeqNo=1284&sCode=eng

[8] The Dong-A Ilbo. From “Knowing” to “Directly Handling”… The Transformation of Elementary AI Education Led by “RUMI”. 31 August 2026.
https://www.donga.com/en/article/all/20260831/6367867/1

[9] New York City Mayor’s Office. Mayor Mamdani and Chancellor Samuels Put Students First with Nation’s Broadest Generative AI Moratorium in Schools. 2 September 2026.
https://www.nyc.gov/mayors-office/news/2026/09/mayor-mamdani-and-chancellor-samuels-put-students-first-with-nat

[10] UNESCO. Digital Learning Week 2026 — Education in the age of AI: Facts | Frictions | Frontiers. 8–11 September 2026.
https://www.unesco.org/en/weeks/digital-learning

[11] Maisiri, James. I refused to train the AI that could replace me. Rest of World, 3 September 2026.
https://restofworld.org/2026/ai-training-jobs-expert-replacement/

[12] Statistics South Africa. Quarterly Labour Force Survey, Q2 2026.
https://www.statssa.gov.za/?p=19804