Who Governs the Governors?

Standards bodies, auditors, regulators and international panels decide what “safe” means for AI. So who checks them? On the second layer of AI governance, where the rules about the rules get made.

FairByDesign The Power Argument title card: “Who Governs the Governors?”

Somebody decides what “safe” means. Before any AI system is certified, audited, approved or banned, a smaller group of people has already decided which tests count, which risks matter, what documentation is enough and who is qualified to say so. Those people are the governors of AI governance. They sit in standards committees, audit firms, regulators, evaluation institutes and international panels. Most of the time, nobody asks who governs them.

That question matters to you even if you will never read a technical standard, because the standard decides what your employer, your school or your bank has to prove before using a system on you. If the definition of “good enough” is written without you, you inherit it anyway.

Standards: where values become specifications

The EU’s AI Act is a good place to see this, because it openly delegates. Much of what the Act’s requirements mean in practice is left to harmonised European standards, drafted by standardisation bodies. Veale and Zuiderveen Borgesius warned early that this hands substantive, rights-laden choices to private bodies with little fundamental-rights expertise and limited civil-society participation (Veale & Zuiderveen Borgesius, 2021). Laux, Wachter and Mittelstadt sharpened the point. When standards bodies settle normative questions such as what counts as fair or transparent, they act without democratic legitimacy, and standardisation, they note, “tends to exclude non-expert stakeholders and the public at large” (Laux et al., 2024). Their proposed remedy is minimum, public disclosure of testing and documentation by default, so that the value judgements are at least visible.

This is FairByDesign’s analysis. A standard is a decision about power that has been formatted to look like engineering. That does not make standards bad. They are often the only way principles become operational controls. But it means the composition of the room matters, and so does whether outsiders can see what was traded away.

Auditors: who checks the checkers?

The next layer is assurance. When an organisation says its system has been audited, the reassurance is only as good as the auditor’s independence, access and method. Costanza-Chock, Raji and Buolamwini surveyed and interviewed people working in algorithmic auditing and found a young, unstandardised field in which results are rarely made public and the communities affected are seldom involved. They recommend standards for auditors, accreditation, disclosure of results and a role for affected stakeholders (Costanza-Chock et al., 2022). Raji and colleagues, drawing on third-party audit regimes in finance, environmental regulation and health, argue that outside oversight needs deliberate institutional design, including independence, access, accreditation and protections such as incident reporting. They also warn that audits alone will not deliver accountability (Raji et al., 2022).

The deeper warning comes from outside AI altogether. Michael Power’s study of the “audit society” showed how, from the 1980s, verification spread as a political answer to demands for accountability. It often became a ritual that produced comfort rather than insight, and it reshaped organisations around whatever could be audited (Power, 1997). Goodman and Tréhu apply that to AI. Poorly designed, unstandardised or non-independent audits can be “at best meaningless and at worst even excuses harms,” a practice they call audit-washing (Goodman & Tréhu, 2022).

Regulators: close to the people they regulate

Regulators are supposed to be the final check, and most work hard at it. But they need expertise that is concentrated in the companies they oversee, and that dependence creates channels of influence. In interviews with US policy experts, Wei and colleagues found that the channels most often named were agenda-setting, advocacy, influence over academic work and the management of information. They framed the risk as one in which “industry co-opts regulatory regimes to prioritize private over public welfare,” and proposed remedies including in-house government expertise, independent research funding and wider civil-society participation (Wei et al., 2024). Again, this describes a risk, not a verdict on any particular agency. It tells you which questions to ask: who drafted the first version, whose data the regulator relied on, and who was in the meetings.

Panels: global, contested and voluntary

At the international level, the question of who governs the governors is fought out in public. When the UN General Assembly appointed the 40 members of its new Independent International Scientific Panel on AI in February 2026, the United States voted against and called the panel and the associated Global Dialogue “a significant overreach of the UN’s mandate and competence” (United Nations Regional Information Centre, 2026). At the same time, scholars working from the Global South have argued that mainstream “inclusive AI governance” often includes their regions only as a token, without engaging with power or historical inequality (Png, 2022). One side says the global forum has too much reach. The other says it has too little real inclusion. Both are claims about who gets to govern.

What would make the governors governable

The literature does not leave us empty-handed. Across these fields the same conditions keep recurring. The first is independence, meaning that the checker is not paid, staffed or briefed by the checked. The second is access, the ability to see enough of the system to test it. Longpre and colleagues show that company-run access programmes are “an inadequate substitute for independent research access” and call for a legal and technical safe harbour for good-faith researchers (Longpre et al., 2024). The third is disclosure: publishing methods, results and failures so that others can check the checkers. The last is participation, so that the people affected have standing before the definitions are fixed, not after.

None of these conditions is exotic. They are how we already try to govern auditors in finance and inspectors in public health. What is missing in AI is the will to apply them, and that shortfall is itself a decision.

FairByDesign’s place in this is modest and specific. We are not a regulator or an auditor. We are an independent investigator whose job is to read what the governors publish, test what can be tested, and show you where the definitions came from. That only works if we are not funded by the people we examine. In the spirit of this piece, we will also publish the strongest case against our own role. If you think independent scrutiny of this kind is overrated, we want to hear why.

Why this needs independent support: someone has to look across the companies, the auditors, the standards bodies and the regulators and ask what the evidence actually shows, without working for any of them. Give once ($5, $10, $25, $50 or any amount), give monthly ($10 or $25), or become a patron → fairbydesign.org/patrons-2/

References

Costanza-Chock, S., Raji, I. D., & Buolamwini, J. (2022). Who audits the auditors? Recommendations from a field scan of the algorithmic auditing ecosystem. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (pp. 1571–1583). ACM. https://doi.org/10.1145/3531146.3533213

Goodman, E. P., & Tréhu, J. (2022). AI audit-washing and accountability. German Marshall Fund of the United States. https://www.gmfus.org/news/ai-audit-washing-and-accountability

Laux, J., Wachter, S., & Mittelstadt, B. (2024). Three pathways for standardisation and ethical disclosure by default under the European Union Artificial Intelligence Act. Computer Law & Security Review, 53, Article 105957. https://doi.org/10.1016/j.clsr.2024.105957

Longpre, S., Kapoor, S., Klyman, K., Ramaswami, A., Bommasani, R., Blili-Hamelin, B., Huang, Y., Skowron, A., Yong, Z.-X., Kotha, S., Zeng, Y., Shi, W., Yang, X., Southen, R., Robey, A., Chao, P., Yang, D., Jia, R., Kang, D., … Henderson, P. (2024). Position: A safe harbor for AI evaluation and red teaming. In Proceedings of the 41st International Conference on Machine Learning (PMLR Vol. 235, pp. 32691–32710). https://proceedings.mlr.press/v235/longpre24a.html

Png, M.-T. (2022). At the tensions of South and North: Critical roles of Global South stakeholders in AI governance. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (pp. 1434–1445). ACM. https://doi.org/10.1145/3531146.3533200

Power, M. (1997). The audit society: Rituals of verification. Oxford University Press.

Raji, I. D., Xu, P., Honigsberg, C., & Ho, D. E. (2022). Outsider oversight: Designing a third party audit ecosystem for AI governance. In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society (pp. 557–571). ACM. https://doi.org/10.1145/3514094.3534181

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

Veale, M., & Zuiderveen Borgesius, F. (2021). Demystifying the draft EU Artificial Intelligence Act: Analysing the good, the bad, and the unclear elements of the proposed approach. Computer Law Review International, 22(4), 97–112. https://doi.org/10.9785/cri-2021-220402

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

Leave a Reply