AI governance has become a market.

In the last twelve months, the number of vendors selling AI ethics scoring tools, governance platforms, responsible AI certifications, and trustworthy AI audits has multiplied. The websites are clean. The language is confident. The visual identity is sharp. Words like sovereignty, alignment, ethics, and governance are being deployed at a pace the field has not seen before.

Underneath a substantial portion of this market activity, there is very little actual governance happening.

I want to give you a way to tell the difference. After almost twenty years operating inside institutions where AI deployment decisions are being made (across humanitarian response, public sector communications, large-scale community engagement, and current AI governance work inside one of the largest charter school networks in the country), I have come to believe the distinction between real AI governance and AI governance theater can be surfaced with three questions.

Question one. Does the governance work happen before deployment, or after?

Governance theater operates downstream of the AI system. It scores model outputs. It evaluates whether responses align with stated values. It produces audit reports and certifications. The AI is already deployed. The communities affected by it are already affected. The governance product is essentially a quality assurance check, dressed in ethics language.

Real governance operates upstream. It asks who made the decision to deploy this system. It asks who was in the room when that decision was made. It asks who was not in the room. It asks what would have had to be true for the answer to be no.

If a governance approach can only score systems after they are running, it is auditing, not governing. Both functions can exist in an organization, but they should not be confused with each other. An audit cannot stop a deployment. Governance can.

Question two. Does it surface power, or does it obscure power?

Governance theater is heavy on values language. Ethics, principles, alignment, trust, responsibility. The language is universal and uncontroversial. It rarely specifies, with precision, who holds the authority to make binding decisions about the AI system in question.

Real governance is explicit about decision rights. It names who has the authority to deploy. Who has the authority to pause. Who has the authority to decommission. What evidence counts as a reason to stop. Who decides what evidence counts. Whether the affected community has any standing in that process, and what their standing actually permits them to do.

A useful test: read the governance framework, audit tool, or platform's marketing copy. Can you tell, from what is published, who actually holds the power to stop the AI system from operating? If you cannot answer that question after reading their materials, the framework is obscuring power, not surfacing it.

Question three. Can affected communities actually enforce a no?

This is the question that separates everything.

Almost every AI governance framework currently on the market describes participation, consultation, stakeholder engagement, or community input as part of its approach. Very few of them give affected communities the authority to stop deployment. The participation is performed. The decision-making stays with the institution. The community provides input, and the input is taken into consideration, but the input is not binding.

This is what I have come to call the translation layer.

The translation layer is the governance infrastructure that enables communities most affected by AI systems to exercise actual decision-making authority over the development, evaluation, and deployment of those systems. It distinguishes between participation and power. Existing AI governance frameworks frequently conflate the two.

Three components define the translation layer in operational terms.

Participation infrastructure. Standing bodies with reserved authority for affected communities, not advisory panels, not consultation processes, not town halls.

Evidence redefinition. Treating the testimony, observation, and analytical work of affected communities as expertise rather than anecdote.

Enforcement authority. Real power over evaluation metrics, equity constraints, and the timing and conditions of deployment, including the authority to pause or decommission tools.

Without all three components, the translation layer does not exist. A participatory council without enforcement authority is consultation. Evidence redefinition without participation infrastructure is rhetoric. Enforcement authority granted to bodies that do not include affected communities is the institutional layer absorbing the language of inclusion while retaining the substance of control.

Why this matters now.

The risk in the current moment is not that AI governance fails. The risk is that AI governance succeeds rhetorically while failing structurally. The market is filling up with products that use the language of ethics and governance to sell something else. If enough vendors do this, AI governance as a category becomes meaningless, the same way ESG as a category became meaningless after a wave of commercial vendors hollowed it out.

Mission-driven organizations, philanthropies, civic institutions, regulators, and policymakers are making decisions right now about which AI governance approaches to adopt. The decisions are consequential. The wrong choice does not just waste budget. It puts affected communities in the position of being governed by a system that claims to protect them but does not.

The three questions are diagnostic. They do not require expertise in machine learning. They do not require legal training. They do not require access to a vendor's proprietary engine. They require the discipline to look past the language and ask, in concrete terms, who decides, on whose terms, and with what enforcement.

Real AI governance is harder, slower, and less aesthetically appealing than the market is currently selling. It requires institutions to give up power they currently hold. It requires affected communities to have real authority, not symbolic representation. It requires the willingness to be told no by people who have historically been talked over.

That is the work. Everything else is theater.

Dupé Ajayi is the founder of The Ajayi Effect, an AI governance and ethics platform. She is the author of the forthcoming Cambridge Journal of Artificial Intelligence piece "The Translation Layer: What AI Ethics Refuses to Build."

The Translation Layer™ is an analytical framework developed by Dupé Ajayi. Researchers, journalists, and practitioners are welcome to engage with, cite, and build on the framework. Attribution requested.