Axioms and Semantics: Formalizing AI Behavior
The belief that AI systems can be adequately described through natural language specifications is the foundational mistake of modern AI governance.
We have spent years writing policy documents, safety guidelines, and alignment frameworks in English—or French, or Mandarin—as though linguistic precision could constrain mathematical objects. It cannot. An AI system is not a person who reads your values statement and chooses to honor it. It is a function operating within a formal space. If you want to specify its behavior, you must speak in the language of that space: axioms, semantics, and formal proof.
This is not a technical nicety. It is the difference between intention and guarantee.
The Thing Everyone Gets Wrong
The standard approach treats AI behavior as something that emerges from training, then gets refined through alignment techniques, then gets bounded by policy. Each layer is assumed to add safety. But this is sequential thinking applied to a mathematical object. You cannot layer natural language constraints onto a neural network the way you layer regulations onto a human organization. The network does not parse your safety guidelines. It has already been instantiated. Its behavior is determined by its weights, its architecture, and the input it receives.
When we say an AI system "should not" do something, we are making a claim about desired behavior. When we want to ensure it does not, we need a formal specification: a set of axioms that define the space of permissible outputs, and a proof that the system's semantics—the actual mathematical meaning of its operations—cannot violate those axioms.
This distinction matters because informal specifications create the illusion of control while leaving gaps. A language model trained to "be helpful, harmless, and honest" has absorbed statistical patterns from text. Those patterns may correlate with helpfulness in common cases. They will not hold at the boundary. They will not hold under adversarial input. They will not hold when the system encounters a scenario its training data did not adequately represent. At that point, you have no guarantee. You have only hope.
Why This Matters More Than People Realize
The cost of this gap is compounding. As AI systems move into domains where behavior matters—medical diagnosis, infrastructure control, financial decision-making—the informal specification becomes a liability. A hospital deploying a diagnostic AI cannot rely on the vendor's assurance that the model "has been tested for fairness." It needs a formal proof that the model's decision boundary does not encode protected characteristics, or a formal specification of the conditions under which it may.
This is not about perfect safety. It is about knowable safety. Formal systems allow you to specify what you actually require, then verify whether a given system meets it. They make the boundaries explicit. They expose assumptions. They force you to articulate what "correct behavior" means in mathematical terms rather than leaving it as an intuition.
The secondary cost is architectural. Systems designed without formal specification in mind are harder to formally specify afterward. They are built for performance on benchmarks, not for provable properties. Retrofitting formal guarantees onto a system designed for empirical optimization is expensive and often impossible. The time to think about axioms is before you train the model, not after it has already learned to approximate your values through gradient descent.
What Actually Changes When You See It Clearly
Once you accept that AI behavior must be formally specified to be reliably constrained, the entire approach to system design shifts. You begin with axioms: the irreducible properties your system must maintain. You design the architecture and training process to preserve those properties. You prove—not test, but prove—that the semantics of the system cannot violate them.
This is harder than writing a safety document. It is also the only approach that scales. As systems become more capable and more autonomous, informal assurance becomes noise. Formal specification becomes the only language in which guarantees can be made.
The question is not whether this is possible. It is whether we will do it before the cost of not doing it becomes catastrophic.