A policy manifesto on AI governance

I have read most of the AI governance frameworks. I have written a fair number of the definitions and strategies that the field now uses. So let me state this plainly, because I believe it. The greatest flaw in AI governance today is not a missing control, a weak standard, or a runtime regulator. The problem with AI governance today is a handful of words.

We are trying to govern systems we have not correctly defined, and you cannot govern what you cannot define.

Governance has always been about language. Get the definition wrong, and every rule built on top of it inherits the mistake.

There are three words that determine whether AI governance is even possible, and we have broken all three in different ways.

• Autonomy – is used incorrectly on a daily basis.
• Agent – there is yet to be a shared definition or agreement on.
• AI – has been condensed to mean the name of a single product.

These words have corrupted our basic understanding of systems and, thus, also corrupted our ability to govern AI adequately.

Autonomy. The word used against its meaning.

Autonomy comes from autonomos. Self-rule. A thing that gives itself its own law. Automation comes from automatos. Self-moving. A thing that, once set in motion, keeps moving toward an end it did not choose. A clock is self-moving. It is not self-ruling. It keeps perfect time and cannot reset itself when the season changes. A thermostat runs your house and cannot decide what temperature your house should be. Every system we are shipping today is self-moving. None of it is self-ruling. It is automation. It is not autonomy.

To govern is to impose law from the outside. Governance assumes the thing being governed does not govern itself. That is the entire premise of the activity. So read the sentence slowly. If a system were autonomous, it would govern itself, and there would be nothing left for us to govern. The moment we call a system autonomous, we have declared it beyond our authority. Then we write a rulebook for it anyway. The vocabulary and the mission are at war, and almost no one has noticed.

If you want to watch that war lost in real time, read the law. The European Union’s AI Act defines an “AI system,” in Article 3, as a machine-based system “designed to operate with varying levels of autonomy.” That word is the hinge on which the entire regulation swings. Every obligation, every risk tier, every penalty attaches to that definition. And the Commission’s own guidance explains what it means by autonomy: some degree of independence from human involvement. That is automation.

They reached for the word that means self-rule to describe a system running without a human clicking every step, and then built the machinery of governance on top of it. In the Act, the more “autonomous” a system is, the more it falls under governance. Read against the word’s meaning; this is upside down. Autonomy is the one property that would make governance unnecessary, and the law has turned it into the trigger that summons the regulator.

To simplify this argument, if systems were autonomous (which means governance), then we wouldn’t need an Act to define how to govern them.

This fix has already been made in the auto industry. SAE International, in its J3016 standard, formally deprecated “autonomous vehicle.” The taxonomy the whole industry now cites is levels of driving automation, zero through five. Automation, not autonomy. Their reasoning is almost a manifesto in miniature: these vehicles depend on external systems to function, so “autonomous” is misleading.

The engineers recognize the difference between automation and autonomy.

The marketers reached for the exciting one.

You can tell who is who by the vocabulary they use.

The more accurate word is “heteronomous systems,” or “heteronomous AI.”

Agent. The word no one has defined.

Look at what “agent” means long before we reached the era of “AI.” An agent acts on behalf of a principal. That is the entire concept, in law, in economics, in plain speech. An agent represents someone. An agent is loaned authority by someone. An agent who serves no one is not an agent; he is an actor with his own ends. “Agent” is a heteronomous word at its root. It presupposes a principal.

So “autonomous agent” is an oxymoron the instant you slow down and read it. Autonomous means it rules itself. Agent means acts for another. We have minted a phrase for a thing that answers to a principal and answers to no one at the same time. The phrase survives mainly because of intellectual laziness.

Then there is the harder failure. There is no shared definition of an AI agent. Ask ten vendors, and you will get ten definitions, each one describing the product the vendor is selling. That is a sales sheet dressed as a definition. You cannot govern a moving goalpost. You cannot write a rule against a word that reshapes itself to fit whoever is holding it and expands to match a quarterly roadmap. A definition authored by the party being governed is not a definition worth governing by. Before we regulate agents, we have to agree on what an agent is, and that agreement cannot belong to the people selling them.

AI. The word condensed into a product.

Ask a thousand people what AI is. Most will say ChatGPT, or Claude, or the assistant on their phone. Artificial intelligence has quietly collapsed into the name of a few consumer products from the last few years. That is the third error, and it may be the costliest, because it decides where governance even points. AI is decades of techniques.

Expert systems, optimization, computer vision, and classical machine learning. The credit model that rules on your loan, the actuarial system that prices your insurance, the filter that reads your resume before a human sees it. None of it has a chat window. All of it shapes lives. When the public hears “AI” and pictures a chatbot, the law follows the public. We write careful rules for the conversational surface people can see, and we leave the vast machinery of automated decisions in hiring, credit, housing, and policing sitting quietly outside the frame. Shrink the word, and you shrink the reach of the law. Define AI as a chatbot, and you will govern chatbots while everything that matters runs unwatched.

Two failures, in opposite directions. Autonomy and agent inflate, dressing mechanism as selfhood. AI deflates, shrinking a whole science to a single app. One makes systems sound more worthy of governance than they are. The other hides most of them from view. Both leave the law aimed at the wrong object. “Loop engineering” and the rest of the buzzword parade are just further instances, each one taking a word that means will, selfhood, or self-rule and bolting it onto a mechanism that has none, because the true word does not sell and does not thrill.

So here is the policy, and I mean it as standards, not slogans.

1. Define before you govern. Fix the meaning of the load-bearing terms in the open, through standards bodies, before any rule that leans on them. A regulation is only as sound as its definitions.
2. Name systems by what they do. It is automation. Reserve “autonomy” for a system that authors its own law, and until one exists, retire the word from technical and legal use.
3. Grade the machine, not the myth. Where independence from human oversight needs to be measured, adopt levels of automation, as aviation and automotive already have. Degrees of human-out-of-the-loop are measurable. Selfhood is not.
4. Assume the human. Every deployed system is heteronomous. Stop treating “human in the loop” as a feature to add, and start requiring that the human who is already there be named, visible, and answerable.
5. Refuse vendor-authored definitions. A word used to govern an industry cannot be defined by the firms selling into it. “Agent” has no settled meaning today, so it has no business anchoring a rule today. Settle the meaning in the open first.
6. Govern the decision, not the interface. A system’s risk comes from what it decides about people. Whether it talks is beside the point. Point the rules at consequential automated decisions wherever they run, and hardest where there is no chat window to draw the eye.

There is a fair objection here, and I will meet it. This sounds like AI literacy, and literacy is slow, diffuse, and hard to enforce. Agreed. That is the very reason definition is the place to start. You cannot make a hundred million people stop and think about what a system can and cannot do. You can fix the words they inherit before they think at all. Correct the definition, and ensure the understanding follows the definition. That is a policy.

If the system were truly autonomous, why would you need to govern it at all?

---

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!