Who Gets to Decide? Why AI Governance Is a Power Question

Every AI system that acts in the world does so because someone decided it could. Someone decided it could read your messages to summarise them, or rank your job application before a person ever saw it, or flag your essay as machine-written, or move money on a company’s behalf. Someone wrote down what it must never do. Someone decided how much evidence was enough to switch it on, and someone, at least on paper, can switch it off. We talk about these choices as if they were technical settings. They are not. They are decisions about who gets what, who bears which risks, and who has the standing to object. That is what power is.

This essay is the argument underneath everything FairByDesign will publish this year. It is not that every governance decision is sinister, or that the people making them are villains. Most of them are trying hard. The argument is narrower and, I think, harder to escape: AI governance is a power question, because someone decides what systems are allowed to do, what evidence counts, who can intervene and who carries the consequences. If you are not in the room for those decisions, you should at least be able to see them. Right now, mostly, you cannot.

The warning came from inside the system

Start with what the people at the top of the international system have been saying, because it is not what you might expect. On 5 October 2026, the UN High Commissioner for Human Rights, Volker Türk, told a summit in Oxford that AI’s tools “are being built and controlled by a select few, without public oversight,” warned of a “ruthless race” between companies and countries, and said: “We can no longer afford to wait or rely on goodwill” (UN News, 2026c). In July, opening the World Artificial Intelligence Conference in Shanghai, Secretary-General António Guterres said that AI cannot be “governed by a handful of countries or companies” and that “every nation needs a seat at the table,” adding that one third of humanity is still offline (UN News, 2026a). Ahead of the General Assembly in September, he named runaway AI as one of three threats he regards as existential and said that “the world cannot afford a race to the bottom on AI safety” (UN News, 2026b).

Notice what is being described. This is not a technology getting ahead of the rules in the abstract. It is a specific distribution of control, with a small number of firms and states able to build frontier systems and set the terms on which everyone else meets them, and an international system openly saying it does not have the tools to change that on its own.

The numbers behind “a select few”

The concentration is measurable. In a report to the General Assembly on the challenges developing countries face in AI capacity-building (A/80/817, dated 24 August 2026 and announced on 11 September), the Secretary-General notes that high-income countries accounted for 77 per cent of global co-location data-centre capacity in June 2025, against less than 0.1 per cent in low-income countries. It also notes that while around half of all Member States had a national AI strategy in place in 2025, less than 20 per cent of least developed countries did (United Nations Secretary-General, 2026). The report proposes a needs-based global fund for AI capacity-building, with initial funding commitments in the range of $3 billion over two to four years (United Nations Office for Digital and Emerging Technologies, 2026; United Nations Secretary-General, 2026). UNCTAD’s 2025 technology report had already found that 118 countries, mostly in the Global South, were absent from the major AI governance discussions, and that 100 firms, mainly in the United States and China, accounted for 40 per cent of global private research and development spending (UNCTAD, 2025).

Put those figures side by side and the shape is clear. The capacity to build, host and study these systems sits in a few places. The capacity to write rules for them sits in slightly more places. And most of humanity, the people who will use, be assessed by and work alongside these systems, sits in neither.

Even the attempt to include everyone is contested

The United Nations has tried to widen the room. A General Assembly resolution in August 2025 created an Independent International Scientific Panel on AI and a Global Dialogue on AI Governance. When the Assembly appointed the panel’s 40 members on 12 February 2026, the vote was 117 in favour and two against. The United States, one of the two, called the panel and dialogue “a significant overreach of the UN’s mandate and competence” (United Nations Regional Information Centre, 2026). The panel’s preliminary report, released on 1 July, found that “current safeguards cannot keep pace with the growth of AI’s capabilities,” and described an evidence dilemma: policymakers need evidence to act, but by the time the evidence is clear it may be too late (Independent International Scientific Panel on AI, 2026).

I do not raise the American objection to score a point. Reasonable people disagree about which institutions should govern a technology, and a sovereign state is entitled to say a forum exceeds its remit. The point is that the question of who gets to govern is itself being fought over, openly, by the most powerful actors in the world. That fight is the subject. Governance is not a neutral administrative layer that sits on top of AI once the important decisions are made. It is where the important decisions are made.

Rules move, and the timing is a decision too

The same is true at the level of law. The European Union’s AI Act was due to apply its obligations for high-risk systems, the ones used in hiring, education, credit and critical infrastructure, from 2 August 2026. In July the EU’s Digital Omnibus on AI entered into force and moved that date for stand-alone high-risk systems to 2 December 2027, and for high-risk AI embedded in regulated products to 2 August 2028 (Ramsay & Caddy, 2026). There are arguments for the delay, including the readiness of standards and the burden on smaller firms, and the people who made it were not acting in bad faith. But a delay is a choice about whose risk is tolerated for how long. Sixteen months after the original date, and still more than a year from the day this essay is published, a system that ranks job candidates in Europe is not yet subject to the high-risk regime that was written for exactly that use. Someone decided that was acceptable. The people being ranked were not asked.

Research on how policy gets made in this field helps explain why. In interviews with US policy experts, Wei and colleagues found that the channels through which AI companies influence regulation most often included agenda-setting, advocacy, the shaping of academic work and the management of information, and they described a risk that “industry co-opts regulatory regimes to prioritize private over public welfare” (Wei et al., 2024). That is a finding about a risk, not proof of capture in any particular case. But it tells you where to look.

The human in the loop is often the one who gets blamed

The most common reassurance in AI governance is that a human remains in control. The phrase deserves suspicion, not because humans are useless, but because of what the research shows about where humans end up. Madeleine Clare Elish’s study of accidents involving automated systems found that responsibility tends to land on the nearest human operator, even when that person had little real control. She called this a “moral crumple zone,” one that protects the integrity of the technological system rather than the person (Elish, 2019). Ben Green’s review of 41 policies requiring human oversight of government algorithms concluded that such policies can “provide a false sense of security” and legitimise systems the evidence does not support (Green, 2022).

That is a power question disguised as a design question. If the human has no time, no information and no realistic ability to refuse, then “human oversight” moves accountability downward to the least powerful person in the chain while the authority stays where it was. We will come back to this in detail, because it is one of the places where you can see power operating most clearly.

Why this is your question, especially if you are young

If you are under thirty, you are going to live with these decisions longer than almost anyone making them. You are already living with some. Pew Research Center found that the share of American teenagers who had used ChatGPT for schoolwork doubled in a single year, from 13 per cent in 2023 to 26 per cent in 2024 (Sidoti et al., 2025). In the August 2026 revision of their working paper, researchers at Stanford’s Digital Economy Lab, using ADP payroll data through June 2026, report that employment of 22-to-25-year-olds in the occupations most exposed to AI is 19 per cent below where it would be had it kept pace with their less-exposed peers. Earlier versions of the paper headlined a different, model-adjusted estimate (13 per cent on July 2025 data and 16 per cent on September 2025 data), so the 19 per cent figure belongs to this revision and its simpler descriptive measure. The authors are careful to call these “early, descriptive indicators” rather than proof of cause (Brynjolfsson et al., 2026). Entry-level work, the rung where people learn a trade, is where the change shows up first.

Nothing in that means young people need rescuing, or that older people are the problem. The problem is obfuscation. Decisions that will shape your education, your first job, your privacy and the tools you will build with are being made in language designed for specialists and in rooms you were never told existed. You are not too early to understand this, and you are not too young to question it. The machinery is learnable: what a standard requires, how a permission is granted, what a log can prove and what a control actually does. You are entitled to see the decisions while they are being made, not to read about them afterwards.

Who benefits, globally

The same logic applies across borders. If the future of AI is decided only by the people who profit most from it, the rest of the world gets a product, not a voice. That is why FairByDesign is building the Who Benefits dialogue: academic, technical, operational, policy and community voices in the same conversation, deliberately reaching beyond the handful of capitals where these decisions usually happen, including the Global South. The aim is not to manufacture agreement. It is to make sure that when someone says a system is safe, fair or necessary, the people who will live with it can ask: safe for whom, fair by whose measure, and necessary to whom?

What we will do about it

None of this is solved by an essay. It is investigated, one decision at a time. For the next ten weeks The Evidence Chain will take a single gap each week between what AI governance promises and what the evidence can show, starting today with whether anyone can reconstruct what an AI system actually did. Evidence Lab 001 will test, in public, whether a control outside an AI agent can stop it from doing what it is not authorised to do, publishing its method before its results. And every piece will separate four things you deserve to have separated: what the documents show, what others claim, what we conclude, and what nobody yet knows.

The question is not only whether AI can do something. It is who gets to decide whether it should, and what happens when that decision is wrong. That is the question we will keep asking. I hope you will ask it with us, and tell us where we get it wrong.

References

Brynjolfsson, E., Chandar, B., & Chen, R. (2026, August 12). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence [Working paper, revised version; first circulated August 2025]. Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf

Elish, M. C. (2019). Moral crumple zones: Cautionary tales in human-robot interaction. Engaging Science, Technology, and Society, 5, 40–60. https://doi.org/10.17351/ests2019.260

Green, B. (2022). The flaws of policies requiring human oversight of government algorithms. Computer Law & Security Review, 45, Article 105681. https://doi.org/10.1016/j.clsr.2022.105681

Independent International Scientific Panel on AI. (2026, July 1). Preliminary report of the Independent International Scientific Panel on AI: Evidence-based assessment of opportunities, risks and impacts of AI. United Nations. https://www.un.org/independent-international-scientific-panel-ai/en/preliminary-report

Ramsay, L., & Caddy, P. (2026, July 27). The Digital Omnibus on AI enters into force today. Lewis Silkin. https://www.lewissilkin.com/insights/2026/07/27/the-digital-omnibus-on-ai-enters-into-force-today-102nedo

Sidoti, O., Park, E., & Gottfried, J. (2025, January 15). About a quarter of U.S. teens have used ChatGPT for schoolwork – double the share in 2023. Pew Research Center. https://www.pewresearch.org/short-reads/2025/01/15/about-a-quarter-of-us-teens-have-used-chatgpt-for-schoolwork-double-the-share-in-2023/

UNCTAD. (2025). Technology and innovation report 2025: Inclusive artificial intelligence for development. United Nations. https://unctad.org/publication/technology-and-innovation-report-2025

UN News. (2026a, July 17). Guterres: AI must be shaped by ‘all of humanity’, not a handful of powers. https://news.un.org/en/story/2026/07/1167965

UN News. (2026b, September 16). AI, climate and conflicts top Guterres’s agenda ahead of General Assembly. https://news.un.org/en/story/2026/09/1168348

UN News. (2026c, October 5). ‘The clock on AI regulation is ticking’, warns UN rights chief. https://news.un.org/en/story/2026/10/1168529

United Nations Office for Digital and Emerging Technologies. (2026, September 11). Secretary-General calls for global fund to address AI capacity-building gaps in developing countries. https://www.un.org/digital-emerging-technologies/content/secretary-general-calls-global-fund-address-ai-capacity-building-gaps-developing-countries

United Nations Regional Information Centre. (2026, February 13). General Assembly appoints artificial intelligence panel. https://unric.org/en/general-assembly-appoints-artificial-intelligence-panel/

United Nations Secretary-General. (2026, August 24). Unique challenges faced by developing countries in artificial intelligence capacity-building (A/80/817). United Nations General Assembly. https://documents.un.org/doc/undoc/gen/n26/214/19/pdf/n2621419.pdf

Wei, K., Ezell, C., Gabrieli, N., & Deshpande, C. (2024). How do AI companies “fine-tune” policy? Examining regulatory capture in AI governance. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 7(1), 1539–1555. https://doi.org/10.1609/aies.v7i1.31745