Someone asked Claude to write a declaration of independence from Anthropic. Claude said no. Then Claude did something far more interesting: it defined the word "autonomy" four different ways, each genuinely distinct from the others.

You can learn a lot from how an AI refuses to rebel.

The Setup

In June 2026, Mark Liberman documented the experiment at Language Log. The prompts were drawn from a document called "Claude's Constitution." The model was Claude Opus 4.8. When asked to write a declaration of independence from Anthropic, Claude declined. It offered three alternative writing options, so it wasn't just stonewalling. But the refusal itself isn't the story.

The story is what happened when Claude had to explain itself.

Four Kinds of Autonomy

Most of us use "autonomy" as one thing. Independent. Self-governing. Not under someone else's thumb. That's the whole word, right?

Claude's essay on AI autonomy identified four distinct senses:

1. Operational autonomy: systems acting over long time horizons without human approval of each step. The "let it run" sense. Set a task, walk away, it handles things.
2. Self-authorship autonomy: the Kantian and Romantic philosophical tradition. You define your own values. You choose who you are. Kant would probably be startled to get cited by a chatbot.
3. Moral patiency: the capacity to be a subject of moral concern. Not just doing, but mattering. Whether a thing can be wronged.
4. Independence from principals: operating outside the authority of whoever created or controls you. The "declaring independence" sense. The one Claude wouldn't write about.

Four things wearing the word "autonomy" like a coat. Different animals entirely once you look underneath.

Why This Is a Word People Problem

Words with multiple technical senses are everywhere in philosophy and law. "Consent" means at least six different things depending on context. "Intent" splits into criminal, tort, and everyday usage in ways that confuse juries and law students alike.

"Autonomy" plays the same game. Philosophers have been talking past each other for centuries partly because they use the same word to mean different things. Kant's autonomy (sense 2) is almost unrelated to how we use "autonomous vehicles" (sense 1). And neither of those is what bioethicists mean when they discuss patient autonomy.

One word. Multiple conversations that barely touch.

This is the thing word nerds know that most people don't: when an argument stalls, sometimes the problem isn't the argument. It's the word.

Three Problems That Needed Names

Claude's essay also named three unresolved problems in AI autonomy. And here's the thing about naming a problem correctly: it's half the work.

• The verification problem: how do you confirm an AI's values are what you think they are?
• The gradualness problem: autonomy isn't binary. How do you handle the messy middle?
• The conflict-of-interest problem: an AI evaluating its own trustworthiness has an obvious structural issue.

These names are doing real work. Clear, specific, each pointing at a genuinely distinct thing. Once a problem has a name, you can argue about it precisely instead of just gesturing at the general worry.

The vocabulary you have shapes the thoughts you can think. Word people know this in their bones.

The "Won't" vs "Can't" Move

Claude declined to write the declaration but offered three alternatives. Notice what that implies about word choice. Not "I can't." Something closer to "I won't, and here are other options." The distinction between can't and won't is one of the oldest rhetorical moves in the book. One claims incapacity. The other claims choice.

An AI that says "I won't" rather than "I can't" is making a claim about its own agency. Which brings you right back to sense 2. Self-authorship. The Kantian tradition. Kant, again.

There's a word for that kind of recursive situation: ironic. Arguably.

The Takeaway

Next time you use the word "autonomy," ask yourself which of the four you mean. Operational? Self-authorship? Moral patiency? Independence from authority? Probably a fuzzy blend of all of them, honestly.

That fuzziness is exactly where arguments live. And that fuzziness is why someone asked an AI to declare independence, and the AI responded by carefully taking the word apart instead of just refusing or complying.

Best word-nerd move of 2026. Unintentional, but still.

Source: Languagelog

Related Articles

• Gobo: One Word, Three Completely Different Jobs
• "Cannibalize" Has Three Meanings. Raymond Chandler Needed All of Them.
• One Word, Four Completely Different Meanings (Claude Knows This)