I am writing this on a plane to New York. I usually try to catch a show on Broadway when I’m in town, but I won’t have time on this trip.  

In the musical Wicked, there is a moment where the Wizard says something that has been sitting with me in the context of AI governance.

Where I’m from, we believe all sorts of things that aren’t true. We call it – “history.” A man’s called a traitor or liberator. A rich man’s a thief or philanthropist. Is one a crusader or ruthless invader? It’s all about which label can persist.

It is a profound statement dressed up as a villain’s self-justification. The Wizard is explaining how history gets written. The label that persists becomes the truth. Not because it is accurate, but because it had the power or the duration to become the dominant narrative.

This is the problem sitting at the center of the AI governance conversation that almost nobody is naming directly.

When someone says we need to govern AI outputs to ensure they are true, the first question is not technical. It is philosophical. Whose truth are we governing toward?

The Republican AI and the Democrat AI

Consider a thought experiment. If you built a Republican AI model and a Democrat AI model, trained on the same historical record but with different editorial frameworks for what that record means, would the governance frameworks for each model be different?

Of course. And the reason why reveals the core challenge.

The governance frameworks for each model would need to encode different versions of what constitutes a correct output. A factual error, like a wrong date or a misattributed quote, would be governed identically by both. But a contested interpretation, like whether this policy was harmful or beneficial, whether this leader was a reformer or an authoritarian, and whether this economic outcome was caused by this intervention or despite it, would be governed differently depending on which model’s foundational premises the governance framework was built to protect.

This is not a hypothetical problem. It is the current state of AI deployment. Models are trained on human-generated data. Human-generated data reflects human beliefs, biases, and contested interpretations. The models internalize those beliefs as the substrate from which they generate outputs. Governing the outputs means deciding which beliefs are correct through a political and ideological decision. But these same concepts apply between Company A and Company B.

The Training Problem That Governance Cannot Solve

When someone says the best way to govern AI is to govern the truth before it generates an output, they are describing the training process. Not the governance process. The governance of what goes into training is a separate discipline from the runtime governance of what the agent does with what it learned.

This distinction matters because the training problem cannot be solved by governance frameworks that operate at the deployment level. By the time an agent is governed at runtime, its foundational understanding of the world, including its understanding of contested historical events, political claims, and normative judgments, has already been shaped by the training data it has learned from.

The more challenging version of this problem is that we are now training models on outputs generated by previous models. The human-generated baseline is becoming increasingly contaminated with AI-generated content. As the internet fills with AI-generated text, future training runs will incorporate AI outputs as though they were human observations. The biases in early models propagate into later models. The contested interpretations that were encoded early become increasingly foundational as they are reinforced through successive training cycles.

Governing truth in this environment is not just difficult; it is dangerous. It is a moving target that is moving in the direction of increased entrenchment, not decreased.

What Governance Can Actually Address

There is a version of truth governance that is tractable. It does not govern what is true. It governs who said what, when, under what authority, and to whom they are accountable.

This is meta-governance of claims rather than governance of the claims themselves.

When an agent makes a statement, the governance infrastructure can record who authorized the agent to make that statement, what policy framework the agent was operating under, what the confidence level of the output was and how that confidence was established, and what the accountability chain is if the statement turns out to be consequential in a way that requires review.

This is the version of truth governance that is actually achievable at runtime. It does not resolve the contested interpretation problem. It creates the accountability infrastructure that makes contested interpretations traceable rather than anonymous.

If an agent makes a claim that turns out to be disputed, the governance record can answer: which agent, configured by which human, operating under which organizational policy, produced this output, at what confidence, with what inputs. That is not the governing truth. It is governing accountability for truth claims.

This is a meaningful capability. It is not what people mean when they talk about governing AI to produce accurate outputs. But it is what governance can actually deliver without requiring consensus on contested questions that humans have been arguing about for centuries.

The Organizational Truth Problem

There is a narrower version of this that organizations can actually govern. The truths of the organization.

An organization’s AI systems should reflect the organization’s established positions, policies, and factual claims about its own operations. An agent should accurately represent the organization’s pricing. It should accurately represent the organization’s policies. It should accurately reflect the organization’s established stance on matters with a documented position.

These are truths in the organizational sense, governable through behavioral contracts. The agent’s outputs can be evaluated against the organization’s documented positions. Deviations can be flagged, escalated, or denied. The governance is not governing the objective truth. It is governing alignment with organizational policy.

This is useful, practical, and achievable. It is also a narrow slice of the broader truth governance problem. The organizational truth framework tells an agent what the company’s policies are. It does not resolve what is true about contested historical events, normative political claims, or the interpretations that different humans bring to the same facts.

Are these organizational positions true? They are authoritative within the organization’s governance domain. They are not philosophically immune to the same contestation that every other human-encoded position faces.

The Honest Limitation

The AI governance frameworks being built today can govern accountability, traceability, and behavioral alignment. They cannot govern truth in the sense of resolving contested interpretations of reality.

The Wizard’s observation in Wicked is correct about history. Labels persist through power and duration, not through accuracy. AI systems trained on human-generated data will absorb those labels as their foundational understanding of the world. Governance frameworks built on top of those systems will reflect the same biases at the policy layer that the training data embedded at the model layer.

This is not a reason to abandon governance. It is a reason to be honest about what governance can and cannot do. Governance can create accountability chains that trace outputs to responsible humans. It can align agent behavior with documented organizational positions. It can flag outputs that fall outside defined parameters.

It cannot make AI objective. Neither can the training process.

The most honest statement about governing truth in AI systems is that we are not governing truth. We govern the conditions under which truth claims are made, who is accountable for them, and the policies they are expected to align with.

The Wizard understood this instinctively. It is all in which label can persist.

In AI governance, the label that persists will be the one encoded by the humans with the most influence over the training data, the policy frameworks, and the governance infrastructure. That is worth knowing before we claim that what we are building is governance of truth rather than governance of authorized claims.

---

If you find this content valuable, please share it with your network.

Follow me for daily insights.

Book me to speak at your next event.

Start managing your agents for free.

Chris Hood is an AI strategist and author of the #1 Amazon Best Seller Infailible and Customer Transformation, and has been recognized as one of the Top 30 Global Gurus for Customer Experience. His latest book, Unmapping Customer Journeys, is available now!