There is a considerable distance between obtaining answers and building knowledge. Over the past few days, seven signals from very different places have pointed towards the same shift. Artificial intelligence is increasingly being treated as part of the environment, while attention is moving towards the capabilities worth preserving as obtaining an answer requires less and less effort.
In quick succession, we have seen an agreement between trade unions and one of the world’s largest technology companies, the PISA 2025 results, a UNESCO ministerial declaration, a study covering 200 higher education institutions in Latin America and the Caribbean, a Brazilian public policy reaching hundreds of thousands of students, a conference on artificial intelligence and living heritage and, from within the industry itself, a call to slow the pace of frontier models.
These are different developments. Taken together, however, they place agency, judgement, cognitive effort and the ability to decide at the centre of a debate that has for too long focused on what the technology can do.
Microsoft accepts contractual limits on its own tools
Microsoft and two teachers’ unions — the American Federation of Teachers (AFT) and its New York affiliate, the United Federation of Teachers (UFT) — announced the National AI Safety & Privacy Standard for US schools on 9 September. [1][2]
The agreement sets out ten protections that school districts can incorporate directly into their contracts with Microsoft. Once included, they become contractually enforceable obligations. Among them is a ban on using student and education staff data to train AI models, sell those data or repurpose them for other uses. Schools also retain control over how the data are used, stored and deleted.
The standard prevents student tracking, requires human oversight of automated decisions and obliges providers to give families and school communities clear information about how the tools work, what data they collect and what safeguards are in place. [1]
There is a limited exception for certain data needed to protect students. The agreement also prohibits features designed to encourage emotional dependency or artificially prolong interaction beyond what is required for an educational task. The new protections will be available from 1 November. The AFT is also in talks with OpenAI and Anthropic about developing similar commitments. [3]
What makes the move significant is who is involved in setting the limits. They do not come solely from the company or a regulator. Unions representing education workers are claiming a role in deciding the conditions under which AI enters schools.
The OECD finds an uncomfortable association
PISA 2025 assessed 760,000 15-year-olds across 91 countries and economies, the largest edition to date. The results record the lowest OECD averages yet observed in reading and mathematics. Between 2015 and 2025, average reading performance fell by 28 points, a difference the organisation equates to roughly a year and a half of learning. In mathematics, the fall was 22 points, a little over a year. [4]
The deterioration affects precisely some of the reading capabilities that matter most in an AI-rich environment. PISA highlights evaluating information, connecting multiple sources and thinking critically about what is read. It also reports that the share of “hasty readers”, who move quickly through texts but answer inaccurately, almost doubled between 2018 and 2025, from 6.6% to 11.4%. [5]
The relationship with artificial intelligence is more complicated than a simple divide between use and non-use.
Students who report not using AI for specific school tasks — summarising texts, carrying out preliminary research or writing assignments — perform better in science on average than those who do use it. The gap is around 20 points, roughly a year of schooling. Yet students who use AI weekly as general support for their learning achieve results similar to those who do not use it. [5]
There is a third finding worth noting. Among frequent users, performance improves slightly when schools have explicitly taught students how to assess the quality of AI-generated information. [5]
PISA shows associations; it does not demonstrate that using artificial intelligence is causing the decline in performance. Students’ social backgrounds, patterns of use and school contexts also matter. That caution is particularly important because the data themselves show different outcomes depending on what the tool is used for and what kind of support accompanies it.
Andreas Schleicher, the OECD’s Director for Education and Skills, frames the distinction as one between a scaffold and a crutch. Technology can provide scaffolding that supports learning, or it can become a crutch that substitutes for part of the learning process. [6]
The difference lies in the work that remains with the learner. Interpreting, comparing, formulating, revising and judging require effort that a ready-made answer can remove. That saving may be useful, but it may also transfer to the machine part of the cognitive activity that gave the exercise its meaning.
UNESCO has been using a term for that shift in recent days: cognitive offloading. Not merely outsourcing a task, but outsourcing part of the reasoning needed to carry it out.
There is, however, another way to read the problem. Pierre Bourdieu showed that schools do not simply transmit knowledge. They also classify, legitimise certain forms of knowledge and can turn pre-existing social inequalities into apparently natural differences in ability or merit. In L’école conservatrice, published in 1966, he described the education system as one of the most effective mechanisms of social conservation because it sanctions as individual merit a cultural inheritance that is distributed unequally. [15]
Bourdieu did not formulate a critique of PISA in today’s terms. His framework does, however, allow us to see any international assessment system as more than a neutral instrument of measurement. Defining what counts as competence, which forms of knowledge can be compared and which results acquire political authority also produces hierarchies.
Artificial intelligence introduces an ambivalent possibility here. It can reproduce the inequalities embedded in the data, institutions and companies from which it emerges. It could also reduce the cost of accessing capabilities that have until now depended more heavily on economic, cultural or institutional capital. Translation, source comparison, access to specialist literature, learning outside elite institutions or connecting distant intellectual traditions can already be carried out with infrastructure that is far more accessible than it was only a few years ago.
That shift could have epistemological consequences. Narratives about colonialism, knowledge produced in the Global South, feminist critiques or analyses of the social effects of the neoliberal policies associated with the governments of Ronald Reagan and Margaret Thatcher can circulate, connect and be interrogated from places that historically had less power to establish what counts as legitimate knowledge.
There is no guarantee that AI will produce such a rebalancing. The International Monetary Fund and the United Nations have in fact warned that uneven adoption may widen gaps in income, infrastructure and capabilities between people and countries. At the same time, both institutions acknowledge that the technology can raise productivity, expand services and distribute new capabilities if public policy shapes how its benefits are shared. [16][17]
That double possibility matters. A technology capable of deepening inequality may also partially disrupt the mechanisms that reproduce it. In an already profoundly unequal society, any technology that lowers the cost of accessing knowledge, analytical capacity or cultural production may change who has a voice and who retains authority.
It is reasonable to think that such a transformation may be perceived as a threat by those who benefit from the existing distribution. This does not justify attributing a specific intention to PISA, the OECD or AI companies. It does allow us to examine the structural incentives that arise when a technology ceases to be merely a productivity tool and begins to alter the distribution of cultural capital.
UNESCO puts human agency at the centre
UNESCO’s Digital Learning Week took place in Paris from 8 to 11 September. More than 25 ministers and education representatives adopted a joint declaration calling for education to remain a human right and a common good in the age of artificial intelligence. [7]
The declaration was presented on 8 September and made public by UNESCO the following day. Its priorities include deliberative governance, public accountability and respect for the rights of students and education staff.
UNESCO also published The algorithm in the room. Seeing and confronting the implications of AI for the future of education, a collection of 26 pieces structured around six questions. Who will retain agency, autonomy and the ability to supervise; how young people will develop relationally, emotionally and cognitively; what will remain meaningful in pedagogy and assessment; how we will think, ask questions and express ourselves; how education can be governed as a common good; and how AI may reshape our imagination of educational futures. [8]
The shift is revealing. Assessing a technology no longer centres exclusively on what it can do. Relationships, autonomy, expression, judgement and institutions also become part of what must be measured.
Adoption is moving faster than governance in Latin America and the Caribbean
UNESCO IESALC and the United Nations University Institute for the Advanced Study of Sustainability (UNU-IAS) presented a study on the implementation of artificial intelligence in higher education across Latin America and the Caribbean during the same week. [9]
The research covers 200 institutions in 19 countries. Eighty-seven per cent report using AI in at least one area of their activity, while only 26% have a formal AI strategy.
The gap between those figures gives a concrete measure of the mismatch. The tools are already being used in teaching, learning and research, while the institutional mechanisms intended to decide how they should be used are developing much more slowly.
The study also identifies a bottom-up pattern of adoption. Much of the uptake is being driven by teaching staff, researchers and students rather than by strategies defined in advance by institutions. [9]
This reverses the usual sequence of technology policy. Practice comes first and the structure intended to govern it follows. By the time a university starts setting criteria, part of its community is already using external systems, has developed habits and has delegated certain functions to infrastructure the institution itself does not control.
What matters, then, is not only how many universities use AI. It also matters who introduces the tools, who sets the conditions of use and how much capacity the institution retains to supervise them once they are embedded.
Brazil tests a different relationship with the tool
This year’s UNESCO King Hamad Bin Isa Al-Khalifa Prize for the Use of ICT in Education recognised two initiatives: Generation AI, developed by the University of Eastern Finland, the University of Helsinki and the University of Oulu; and Redação Paraná, led by the Education Department of the Brazilian state of Paraná. [10]
UNESCO frames this year’s prize as a response to a growing crisis of cognitive offloading. Generative systems make it easy to outsource activities of thought and critical judgement that until now formed part of the educational process. [10]
Redação Paraná is interesting because it uses artificial intelligence without making that process disappear.
The programme began as a pilot in 2020 and was fully implemented in 2021. It combines a digital platform with a public policy designed to broaden access to writing instruction across the state.
AI provides feedback within structured cycles of writing and revision, while teachers retain a mediating role. The stated aim is to strengthen linguistic competence, argumentation, creativity and higher-order thinking. The programme has already reached close to 850,000 students in around 2,000 schools. [10]
The technology does not remove the act of writing by delivering a finished answer. It intervenes within the process. Learners still have to write, revise, argue and work with teachers’ feedback.
The distinction may look small, but pedagogically it is considerable. Automation can save work, or it can help people do the work. Both involve AI, but they do not necessarily produce the same learning.
Athens takes the discussion into living heritage
On 11 September, the National Gallery – Alexandros Soutsos Museum in Athens hosted Generative Traditions. Living Heritage and Artificial Intelligence. New Horizons – Exploring Cultural Possibilities, organised by the Greek Ministry of Culture in collaboration with UNESCO. The institutional announcement was published the previous day. [11]
The conference brought together specialists in living heritage, artificial intelligence, digital technologies, research and cultural policy. Sessions examined how these tools are changing the documentation, research, safeguarding and transmission of cultural practices in digital environments.
The keynote was given by Hanna Schreiber, holder of the UNESCO Chair on Intangible Cultural Heritage in Public and Global Governance at the University of Warsaw. Her talk was titled Artificial Intelligence and Intangible Cultural Heritage: Mapping terra (in)cognita. [11]
Moving from the classroom to heritage broadens the scope of the problem.
Documenting a practice, transmitting knowledge or sustaining a tradition are not activities whose value can be measured solely by how efficiently they produce an output. Living heritage depends on relationships, interpretation, memory, participation and recognition among the people who practise it.
A generated description may be accurate. A classification may be efficient. An archive may become dramatically more searchable. None of those capabilities, on its own, resolves who interprets, who transmits or who retains authority over what is being documented.
Here too, automation forces a decision about what is worth keeping as a human and communal practice.
The industry itself begins to ask for time
On 12 September, Anthropic chief executive Dario Amodei published We Must Pace the Frontier, an essay calling for a deliberate slowdown in the pace at which frontier-model capabilities improve so that safety research can keep up. [12][13]
Amodei is not proposing an end to training or technical progress. His argument is for enough time between advances to study risks, improve alignment measures and allow third parties to verify them.
His first concrete proposal is to embed external evaluation teams with continuing access comparable to that of internal staff carrying out similar risk-analysis work. Those teams could examine not only finished models but also training processes, incidents and safety practices. Anthropic commits unilaterally to adopting this model and calls on other companies to do the same. [12]
Sam Altman and Elon Musk publicly backed Amodei’s call. Demis Hassabis, head of Google DeepMind, also supported the broader direction of stronger cooperation and safety. [13]
Donald Trump responded forcefully in two stages. On 13 September, speaking to the press in Ireland, he referred to “very negative forces” exaggerating risks that would not materialise and linked any slowdown to the danger of losing the race against China. A day later he went further: in a series of social-media posts he called the idea that AI could “take over the world and destroy humanity” a “HOAX”, spoke of a “SICK conspiracy” against artificial intelligence and data centres, and — in the same rhetorical move — compared those warnings with decades of warnings about climate change, which he also dismissed as a hoax. [14]
This discussion operates at a different scale and starts from concerns distinct from those in education. Its appearance within the same few days is nevertheless revealing.
For years, the race for artificial intelligence has been told through acceleration. More parameters, more compute, more capabilities, shorter cycles. Now some of the organisations sitting precisely at that frontier are beginning to propose another variable: time.
Time to evaluate, understand, supervise and decide before the next capability makes the previous decision irrelevant.
There are technical and safety reasons for doing so. There is also a political dimension worth examining. The people in a position to decide how fast to accelerate and when to slow down belong to an extraordinarily small group of executives, investors and laboratories at the centre of Western technological elites. A conversation about a technology with potentially planetary effects is once again dominated by institutions with an enormous concentration of economic, scientific and political capital.
It is significant that Trump himself has already drawn the same comparison between AI and climate change, albeit in the opposite direction: to dismiss both warnings as a hoax rather than to demand the same level of scrutiny for each. That asymmetry is precisely what deserves closer attention.
There is another asymmetry that is difficult to ignore. Scientific warnings about environmental deterioration have been accumulating for decades. The IPCC published its first assessment in 1990, and its most recent synthesis describes climate change as a threat to human well-being and planetary health, with a rapidly closing window to secure a liveable future. [18][19] The United Nations Environment Programme now warns of a triple planetary crisis of climate change, biodiversity loss and pollution. Global resource extraction has tripled over five decades, while high-income countries consume six times more resources and generate ten times more climate impacts than low-income countries. [20] Under policies currently in place, UNEP projects around 2.8°C of warming over this century. [21]
Against that record, presenting artificial intelligence as a possible existential threat while tolerating for decades a material trajectory that threatens the conditions of habitability should, at the very least, make us sceptical about the hierarchy through which risks are constructed. Not because the danger associated with AI is necessarily false, but because the precautionary principle should operate with comparable intensity in the face of threats already supported by an extensive body of empirical evidence. Climate science has not lacked warnings. What has been missing is a transformation proportionate to them. The Emissions Gap Report 2025 once again finds that current commitments and policies remain far from pathways compatible with the Paris Agreement. [21]
The suspicion, then, should not be directed at prudence itself, but at its selective application. If we accept that a technology deserves limits when it may profoundly alter the human future, the same logic should extend to a system of production and consumption whose climatic, ecological and distributive harms have been documented for decades.
None of this demonstrates that calls to slow AI are intended to protect a particular social order. It does, however, allow another hypothesis to be considered. A pause may make a transformation safer; it may also give markets, regulators and institutions time to accommodate its effects within existing relations of power before new capabilities disrupt them too extensively.
The issue takes on another scale if AI does more than automate jobs and begins to alter who can produce knowledge, interpret evidence or contest historical narratives. A system that makes it easier to investigate colonialism across dispersed archives, compare the effects of economic policies or access academic debates previously restricted by money, language or institutional affiliation may help reduce some asymmetries. Precisely for that reason, any equalising potential may generate resistance among those who have the most to lose from a different distribution of knowledge, income or influence.
It would be excessive to turn that possibility into a single explanation for the current safety debate. But it would also be naïve to separate technological governance entirely from the social structure that produces it. Bourdieu reminded us that institutions capable of classifying also participate in reproducing the world they classify. AI could reinforce that mechanism or open cracks within it.
Learning not to delegate
Microsoft is negotiating limits while continuing to expand its presence in education. PISA finds associations that demand cautious interpretation. UNESCO is trying to place human agency within technological governance. Latin American institutions are adopting tools faster than they are building mechanisms to govern them. Brazil is experimenting with AI without removing the work of writing and revising. Athens carries the issue into the transmission of living heritage. Part of the industry is asking for more time before moving forward again.
They do not constitute a consensus. Nor do they describe a single technology policy. What they share is a tension that becomes visible once automation works well enough.
Learning has never consisted solely of obtaining an answer. It also means sustaining the effort to distinguish, interpret, connect, write and judge. Something similar happens when a community transmits knowledge, when an institution decides which technology to adopt or when a laboratory decides how far it can advance before understanding what it has built.
What is at stake is not only how much thought we delegate. It also matters who decides which capabilities are preserved, which forms of knowledge acquire authority and how much change a society can absorb before its institutions attempt to channel it. AI can reproduce the world from which it comes. It can also make accessible capabilities and knowledge that this world has distributed profoundly unequally.
Answers are becoming extraordinarily cheap.
Learning to use artificial intelligence is also beginning to mean learning what not to delegate, and to whom not to delegate it.
References
-
American Federation of Teachers. AFT, UFT and Microsoft Announce ‘National AI Safety & Privacy Standard’ for Schools to Protect Students, Families and Educators. 9 September 2026.
-
Microsoft Source. AFT, UFT and Microsoft announce ‘National AI Safety & Privacy Standard’ for schools to protect students, families and educators. 9 September 2026.
-
Associated Press. AI in schools leads to pressure on tech companies around data privacy. September 2026.
-
OECD. PISA 2025: Students’ reading and mathematics performance declined sharply across the OECD. 8 September 2026.
-
OECD. PISA 2025 Results (Volume I). 2026.
https://www.oecd.org/en/publications/pisa-2025-results-volume-i_73451bc5-en.html
-
OECD. PISA findings on artificial intelligence use, reading skills and learning. 8 September 2026.
-
UNESCO. Education Ministers call for education to remain a common good in the age of AI at UNESCO’s Digital Learning Week. 9 September 2026.
-
UNESCO. The algorithm in the room: Seeing and confronting the implications of AI for the future of education. 8 September 2026.
-
UNESCO IESALC and UNU-IAS. Implementation of AI in Higher Education in Latin America and the Caribbean. 10 September 2026.
-
UNESCO. UNESCO ICT Prize awards dynamic projects promoting creativity and critical thinking from Finland and Brazil. 9 September 2026.
-
UNESCO Intangible Cultural Heritage. Conference on ‘Generative Traditions. Living Heritage and Artificial Intelligence. New Horizons – Exploring Cultural Possibilities’ in Athens. 10 September 2026.
-
Amodei, Dario. We Must Pace the Frontier. September 2026.
-
Anadolu Agency. Musk, Altman, Hassabis back Amodei’s call to slow pace of AI development. 14 September 2026.
-
Associated Press (Boak, Price, Freking, Castillo, Chan Ho-him, Tang). Trump calls AI risks a ‘hoax,’ says there is a ‘SICK conspiracy’ against AI and data centers. 14–15 September 2026.
https://apnews.com/article/trump-ai-guardrails-data-centers-b85df16775ff7e9611a456b061a0e4b9
-
Bourdieu, Pierre. L’école conservatrice. Les inégalités devant l’école et devant la culture. Revue française de sociologie, 7(3), 1966, pp. 325–347.
-
International Monetary Fund. AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity. 16 January 2024.
-
United Nations Development Programme. The Next Great Divergence: Why AI May Widen Inequality Between Countries. 2 December 2025.
https://www.undp.org/asia-pacific/publications/next-great-divergence
-
Intergovernmental Panel on Climate Change. History of the IPCC.
-
Intergovernmental Panel on Climate Change. AR6 Synthesis Report: Summary for Policymakers. 2023.
https://www.ipcc.ch/report/ar6/syr/summary-for-policymakers/
-
United Nations Environment Programme & International Resource Panel. Global Resources Outlook 2024. 1 March 2024.
-
United Nations Environment Programme. Emissions Gap Report 2025. 4 November 2025.