AI companies like Anthropic and OpenAI are using foundational documents called "AI constitutions" to embed specific values and behaviors directly into their models, but regulators are increasingly viewing these documents as targets for government intervention, raising urgent First Amendment questions that free speech scholars have largely overlooked.

What Exactly Is an AI Constitution?

An AI constitution is a technical document that outlines the values, ethics, and intended behavior for an artificial intelligence model. Unlike a legal constitution, it's not the product of democratic deliberation. Instead, it's written predominantly by employees at AI labs, with some external consultation, and it directly shapes how models behave in the real world.

Anthropic publishes Claude's Constitution, which the company describes as "the foundational document that both expresses and shapes who Claude is." OpenAI maintains the Model Spec, which "outlines the intended behavior for the models that power OpenAI's products." These aren't merely mission statements or governance frameworks; they function as operative control mechanisms during model development.

According to Joe Carlsmith, one of the principal authors of Claude's Constitution, an AI constitution is minimally defined as "a description of the intended values and behavior for an AI system." A thicker definition accounts for how these constitutions are applied during training and monitoring. They specify what a model should and should not do, take final authority over other instructions, are trained directly into the model, and are meant to shape a model's character rather than police outputs case by case.

How Do These Documents Actually Shape AI Behavior?

The two major AI constitutions read quite differently, but both have been empirically shown to influence model performance. Claude's Constitution is philosophical and concerned with the model's psychology. Rather than ordering the model to "always be honest," it exhaustively explains what honesty is and why it matters. It's designed to encode Anthropic's higher-order values by reference to specific ways one can imagine a model behaving well or badly.

OpenAI's Model Spec, by contrast, reads like case law, pairing each principle with sample prompts and examples of allowed and disallowed answers. Both documents inform the generation of synthetic training data, sit inside the model's chain-of-thought reasoning, or serve as scorecards for alignment evaluation afterward. Anthropic reports that training on its Constitution improved alignment in ways that "persisted through RL post-training," a technical process called reinforcement learning from human feedback (RLHF) that fine-tunes model behavior based on human preferences.

Anthropic

Why Are Regulators Suddenly Interested in These Documents?

Over the past year, the federal government has grown intensely interested in the values embedded within AI models. Across executive orders, Office of Management and Budget (OMB) procurement rules, Federal Trade Commission actions, and Department of Justice interventions, the government has engaged in an expansive effort to dictate the values that models millions of people use every day express.

The demonstrated capacity of constitutions to shape model behavior is precisely what will likely make them targets of regulators seeking to confine models to certain values and perspectives. The possibility of governments and governing bodies at state, federal, and international levels attempting to amend or revise AI constitutions warrants advanced scrutiny from free speech scholars.

Recent history strongly suggests that the public should expect some kind of regulation of AI constitutions in the not-too-distant future. The number of congresspeople discussing the dangers of AI has skyrocketed, and policymakers' concerns often center on what these models value. Initiatives like the "Preventing Woke AI" executive order have already signaled government interest in shaping model values through policy.

The First Amendment Problem Nobody Expected

At this stage in AI governance debates, free speech and free expression scholars have focused on more doctrinal questions, such as whether AI outputs themselves are protected speech. But a more urgent question looms: whether AI constitutions are protected speech, and whether government regulation of them would violate the First Amendment.

The use of the word "constitution" to refer to technical documents risks inviting direct comparisons to legal constitutions, but the similarities and differences matter legally. AI constitutions list things models should and should not do, function in society by explicitly outlining principles to resolve difficult ethical and practical questions, and are intended for public analysis. OpenAI publishes its Model Spec because "it's important for people to be able to understand and discuss the practical choices involved in shaping model behavior," suggesting these documents carry expressive weight.

Yet AI constitutions, unlike legal constitutions, are not the product of a democratic process that signals broader consent of the governed. Nor are conflicts over how to interpret them subject to external, independent adjudication. Still, the similarities with legal constitutions are strong enough to elicit regulatory interest.

How Are Different Labs Approaching Constitutional AI?

Not all AI labs have adopted a version of an AI constitution. The landscape is fragmented, with different approaches reflecting different philosophies about how to embed values into models:

• Anthropic and OpenAI: Both publish stand-alone documents that function as training-time specifications of value, making them the clearest examples of AI constitutions.
• Meta: Uses an acceptable-use policy rather than a formal constitution, taking a different approach to value specification.
• Google: Does not publish a single constitution document, though it's possible to craft one from various company documents that make up the core components.
• xAI: Relies on raw system prompts rather than a formal constitutional framework.

There is not yet a default way to write an AI constitution. Claude's Constitution and the Model Spec read quite differently, reflecting different underlying philosophies about how to encode values into AI systems.

What Happens If Governments Start Regulating These Documents?

If regulators attempt to mandate changes to AI constitutions, they would effectively be compelling AI companies to express certain values and suppress others. This raises the question of whether such mandates would constitute compelled speech under the First Amendment, a legal doctrine that generally prohibits the government from forcing individuals or organizations to express particular viewpoints.

Alternatively, if governments attempt to prohibit certain values from being embedded in AI constitutions, they would be restricting the expression of those values, raising free speech concerns. The challenge is that AI constitutions are simultaneously technical documents that control model behavior and expressive documents that communicate company values to the public.

There are alternative possibilities for shaping model characteristics that would avoid First Amendment issues altogether, but these remain largely unexplored. The urgency of this question is heightened by the fact that AI constitutions are increasingly serving as the central locus of AI values discourse, making them a natural target for regulation.

What Should Happen Next?

It's urgent that free speech and free expression scholars expand their inquiry into whether AI constitutions are protected speech before constitution-based regulation takes off in earnest. The stakes are high: the values embedded in AI models affect millions of people daily, but so do the principles of free expression that protect the ability of companies to express those values.

The challenge is to find a path forward that respects both the legitimate public interest in understanding and influencing AI values and the constitutional protections for expressive activity. This will require careful legal analysis, public deliberation, and potentially new frameworks for thinking about how to govern AI in a way that respects both safety and free speech.