On the surface, AI governance sounds technical: audits, documentation, benchmarks, red-teaming. Underneath, it’s a question as old as politics: who gets to decide what other people’s tools may do?

When we talk about “controlling AI,” we are really talking about controlling people through the systems they increasingly rely on. That is the core tension of AI governance. It’s not just safety; it’s power.

Control over what, exactly?

Control operates at multiple layers. Capabilities, deployment, behavior, and ecosystem structure each encode different political choices.

The word “control” hides several different levers:

1. Control over model capabilities
   What can this system do? What tasks is it allowed to perform? What knowledge and tools can it access?
2. Control over deployment
   Where is the model used? In hiring, credit scoring, medical triage, advertising?
3. Control over behaviour
   What values and policies are baked into the system: what it’s allowed to say, what it must refuse, whose laws it follows?
4. Control over ecosystem structure
   Who is permitted to build and deploy powerful models: a few frontier labs, open-source communities, nation-states, some mix of all three?

When regulators draft AI laws, when companies write usage policies, when engineers add more guardrails, they are deciding whose preferences prevail at each of these layers.

Alignment as politics-by-other-means

Alignment is often framed as technical: how do we get models to reliably follow instructions, avoid harmful outputs, and stay within desired boundaries?

But the crucial missing sentence in many alignment discussions is: aligned to whom?

Right now, the answer is often: aligned to the values and risk appetites of a handful of companies, filtered through legal advice and the anticipated preferences of large regulators.

That is understandable. It is not obviously legitimate.

A model that refuses certain political content because it is trained to minimize regulatory risk in the US and EU is not neutral. It encodes a political compromise, enforced through infrastructure. A model that refuses to help union organizers with strategy, but happily helps employers with “change management,” is not misaligned. It is aligned—to someone.

Good governance starts by admitting that.

The new regulatory triangle

Historically, power in tech flowed between three poles:

• States — law, regulation, procurement
• Firms — platforms, infrastructure, markets
• Publics — users, citizens, workers, consumers

AI complicates that triangle in at least two ways:

1. Concentrated technical capital
   Training frontier models now requires billions in capex, access to specialized chips, and deep engineering talent. That stacks power toward a small cluster of firms and states with the resources to build and host them.
2. Global externalities
   A misaligned content policy in one model can shape discourse in dozens of countries simultaneously. A biased hiring model can propagate discrimination at scale across industries.

Regulators are starting to respond. The EU AI Act attempts to claw back some control by imposing obligations on general-purpose AI models and high-risk systems, with phased deadlines and explicit enforcement powers.

In the US, agencies like the FTC are signalling that deceptive AI practices and unfair automated decisions can be treated under existing consumer protection and civil rights law.

Yet the deep question remains: what mix of centralized and decentralized control should we accept?

Centralization vs pluralism

There’s a plausible case for central control: a small number of well-regulated labs, strong safety processes, tight export controls, and high barriers to building dangerous systems. You want that for nuclear material, for aviation standards, for certain types of cryptography.

There’s an equally plausible case for pluralism: many smaller models, open-weight systems, strong local customization, and the ability for communities, firms, and states to run systems aligned with their own legal and cultural frameworks.

Push too hard toward centralization and you risk creating informational sovereigns—corporate or state entities whose value judgments and failure modes quietly shape the lives of billions. Push too hard toward decentralization and you risk fragmentation, race-to-the-bottom safety practices, and the diffusion of capabilities that could enable coordinated harm.

There is no purely technical solution to this balancing act. It is a constitutional question.

Control with people, not just over systems

A humanistic philosophy of AI control starts in a different place: not “how do we constrain AI?” but “how do we protect and empower human agency in a world with pervasive AI?”

That requires at least four commitments:

1. Right to explanation and contestation
   People affected by AI-mediated decisions should be able to understand the basis of those decisions and contest them in human-readable terms.
2. Meaningful human override
   There must be places in systems where humans can say “no”—to a model’s recommendation, to a deployment, to a policy. Not as a checkbox, but as a genuine veto.
3. Pluralism in model behaviour
   Where possible, people and institutions should be able to choose between differently-aligned systems, within the constraints of law. A world with one official model is a world with one official worldview.
4. Shared governance over infrastructure
   As intelligence becomes infrastructure, governance should not be left solely to corporate boards or a handful of government agencies. Standards bodies, civil society, labour organizations, and academic institutions all need a seat at the table.

Control in an age of uncertainty

The honest answer is that we are navigating all this under uncertainty.

We do not know the upper bounds of current model architectures. We do not know how quickly new capabilities will emerge as we scale inputs. We do not know how resilient our institutions will be to cascades of synthetic misinformation, automated persuasion, or high-frequency decision loops.

In that context, governance must be both humble and firm:

• Humble, in acknowledging what we do not know, and in avoiding overconfident predictions—either utopian or catastrophic.
• Firm, in setting boundaries where stakes are high, insisting on evaluation and transparency, and refusing to outsource fundamental public choices to closed systems.

We should be deeply wary of anyone who claims that AI governance is apolitical, or that “the math will save us.” The math matters. So do the laws, the incentives, and the hands on the levers.

In the end, control is not something we apply to AI. It is something we exercise—or fail to exercise—over the institutions that build, deploy, and profit from it.

---