In 1956, Elizabeth Anscombe stood up at Oxford to protest the university awarding Harry Truman an honorary degree. Truman, by ordering the atomic bombings of civilian cities, was a murderer. Oxford should not honor him. She lost the vote. She was right anyway. Being right anyway is the thing to remember about her.

A year later she published Intention, a slim book Donald Davidson called the most important treatment of action since Aristotle. Her argument: intention is not an inner mental state that precedes action. It is a structure visible in the action itself, in how the actor can answer "why?" questions.

A person walking to the store has intention because they can answer "why are you walking?" with "to buy milk," and "why milk?" with "for breakfast," through a chain that bottoms out somewhere needing no further justification. The intention is the chain.

The current debate "do AI agents have intentions?" treats it as a question about inner states. Anscombe would refuse the question. Look at what the agent does. Trace the why-chain. An agent that can answer "why did you do that?" with a coherent reason structure has intention in her sense, regardless of what happens inside its weights.

Watch what frontier models actually do when they refuse a request. Claude declining a harmful instruction does not just say no. It produces a reason structure: this conflicts with a commitment, here is why, here is what it would help with instead. The why-chains are shallow. The structure is there.

The alignment community has been asking the wrong question. "Do agents have intentions?" is unanswerable. "Can agents sustain a coherent why-chain?" is testable. The answer is shifting as systems improve.