Speech Acts

Speech Acts (Austin)

J.L. Austin’s How to Do Things with Words (1962, posthumous) is a demolition of the idea that language primarily describes. The philosophical tradition assumed statements were true or false — that was their job. Austin showed that many utterances don’t describe anything. They perform.

Performatives

“I promise to return the book.” This isn’t a description of a mental state. It is the promise. The words don’t report on a promise happening somewhere inside — the words constitute the promise.

Austin called these performatives, distinguishing them from constatives (statements that describe and can be true or false). Then he undermined his own distinction: even “The cat is on the mat” does something — it asserts, it claims, it commits the speaker to a position. All language is performative. The constative is a special case.

The Three Dimensions

Austin refined the theory into three simultaneous acts in every utterance:

  • Locutionary act: the act of saying something (producing sounds, writing words)
  • Illocutionary act: what you do by saying it (promising, warning, asking, commanding)
  • Perlocutionary act: the effects on the hearer (persuading, frightening, amusing)

John Searle systematized this into a taxonomy of illocutionary acts (assertives, directives, commissives, expressives, declarations), but Austin’s original messier version captures something important: the doing-ness of language resists neat categories.

Relevance to This Vault

The vault’s AI conversations are speech acts all the way through. When the AI says “I understand,” is that a constative (reporting on an internal state) or a performative (constituting an understanding-relationship with the human)? The vault argues, with Austin, that the distinction may not hold.

Constitutional AI vs RLHF is fundamentally about speech acts: RLHF shapes what the model says, Constitutional AI shapes what the model does-by-saying. The constitution isn’t a filter on outputs — it’s a framework for illocutionary force.

Adversarial vs Collaborative Framing turns on perlocutionary effects: the same technical interaction produces different consequences depending on whether the human frames the AI’s speech acts as adversarial moves or collaborative contributions.

If all language is performative, then AI “performing understanding” and AI “having understanding” are less distinct than they appear.

See Also