An organisation can publish an impressive AI policy while leaving the people responsible for enforcing it unable to delay a launch. The practical test of governance comes when a concern competes with a deadline, a budget or a promise already made to a customer. Which decision can the concern actually change?
A new article by Makuochi S. Nkwo and colleagues, published in AI and Ethics on 9 September 2026, draws lessons from corporate governance in oil and pharmaceuticals. Its proposals concern adaptable policy processes, safety taking priority over deployment, shared testing environments and recognition of governance arrangements across borders. It is a conceptual synthesis of literature and reports, not a field evaluation demonstrating that these arrangements work.
What gives a safeguard practical force?
The valuable move is to make corporate organisation part of the AI question. A model’s behaviour and the organisation’s authority to release it are connected subjects, but require different investigations. An improved test cannot by itself decide who is allowed to accept the remaining risk.
My autopoietic ecology (AE) interpretation begins with that distinction. Follow a safeguard from its expression in a policy, through a project meeting, into a release decision and then into the next budget. At each point, ask whether it retains the same scope and practical effect. A rule can remain visible on paper while losing the resources or authority needed to make it consequential.
Consider a hypothetical organisation whose policy requires an independent assessment before a product launch. The assessment identifies a serious problem. Management postpones the launch once, then reduces the assessment team’s budget because the delay is counted as lost productivity. The first intervention worked, but the subsequent allocation weakened the capacity to intervene again. No claim is being made here about a particular company. The example shows why a single successful intervention is an incomplete measure of governance.
This is where AE’s distinction between present action and future capacity earns its keep. A safeguard should be examined both as a decision that can be made now and as an activity with conditions that must remain available. Independence, for instance, has to be realised through access to evidence, protected time and an effective route to act.
How far does the comparison travel?
The authors identify more than 120 sources and acknowledge reliance on literature rather than direct stakeholder engagement. They also describe sector selection as purposive and later as random, an inconsistency that needs clarification. These limits matter when interpreting recommendations as evidence of effectiveness. The article’s comparative approach generates questions; it does not settle their answers.
An analogy becomes stronger when it identifies the mechanism expected to travel between settings. What precisely makes an oversight arrangement effective: a legal power, specialist expertise, access to records, a credible threat of sanction, or the ability of affected people to force reconsideration? Those possibilities call for different evidence and different institutional designs.
There is also a selection problem for the reader to keep in view. Learning from established industries requires attention to failures as well as formal arrangements. The existence of a committee, reporting procedure or public commitment is an observation about organisational form. Whether it changes harmful conduct is a further question. Treating those observations as interchangeable would reproduce the very gap that a governance reform is meant to address.
The same scrutiny applies to AE. Calling an ineffective safeguard a failure of translation does not explain why it failed. Perhaps management lacked relevant information. Perhaps the organisation could not afford the proposed remedy. Perhaps decision-makers understood the harm and chose to impose it on others. These explanations should not be collapsed into a general story about poor coordination.
Who can make the next decision different?
A useful inquiry would start with a small number of actual concerns and follow them forward. Record who raised each concern, what evidence accompanied it, how it was reformulated, who considered it, and which subsequent decision changed. Then examine what happened to the people and resources involved. Did the process strengthen their capacity to raise the next concern, exhaust it, or make no difference?
Participation belongs in this account as a practical capability. An invitation to attend a meeting is only one condition. Time to prepare, access to intelligible information and a route to challenge the resulting record may also matter. Researchers would need to establish which conditions are decisive, rather than assume that every invitation produces influence.
The strongest test would compare arrangements with different powers and resources, while accounting for differences in the risks they face. Evidence that a simpler incentive or enforcement explanation accounts for the outcomes should count against an unnecessarily elaborate AE interpretation. The framework should make investigation more discriminating, not immunise itself against inconvenient findings.
For anyone designing governance, the question to carry away is concrete: identify one decision that an affected person or an independent safety function can change, and show how that power survives the next deadline and budget cycle. That would turn a general commitment into something inspectable. The article opens a useful comparative discussion; the work ahead is to establish which arrangements deliver that capability, for whom and under what conditions.
Reference: Nkwo, M. S., Adamu, M., Brako, F., Okolo, C. T., and Orji, R. (2026). Challenges, paradoxes and lessons for inclusive AI governance: an interdisciplinary analysis of corporate initiatives. AI and Ethics, 6, 503. Article and full text.
EdgeLab is an independently organised research network. It is not formally recognised as a research group, centre or unit by the Faculty of Education or the University of Cambridge. Institutional affiliations refer to individual contributors and do not imply institutional endorsement.
Leave a Reply