Doing Things With Words

Photo by Kelly Sikkema on Unsplash

In 1955, the philosopher J.L. Austin delivered a series of lectures that questioned a foundational assumption about language. The assumption held that when people say things, they are primarily stating facts that can be judged true or false. Austin showed this view was incomplete. His philosophy of speech acts has much to offer those who think about artificial intelligence.

Austin distinguished between three levels of what happens when someone speaks. The locutionary act is simply saying words with meaning. The illocutionary act is what we do in saying those words, the force behind them. When someone says “the bull is about to charge,” they may be warning, informing, or merely describing. The perlocutionary act is the effect produced on the listener, whether they are alarmed, convinced, or reassured.

Austin was after something more than taxonomy. For him, speech doesn’t just convey knowledge, it also enacts other modes of being, such as being able to, and being obliged to.

Large language models generate simulacra of locutionary acts, the said without the saying, traces of speech brought to life through a recursive process. When an LLM says ‘you should evacuate,’ it can also produce the effects of a warning without Austin’s speech acts. A system fluent enough to sound right about anything can also produce reassurance whether or not anything has actually been settled. This reflects a risk of false adequacy.

According to Austin, for an utterance to warn, inform, or describe, certain conditions must be satisfied, and these are part of what makes the acts ethically and epistemically significant. The speaker must have intentions and commitments, and the utterance must be issued in appropriate circumstances with appropriate authority.

An LLM meets none of these conditions: it has no intention to warn, no commitment standing behind the warning, and no position from which to claim the authority to issue one. It follows constraints without choosing among them: execution, not address.

Bourdieu extended Austin’s insight: illocutionary force derives less from the speaker’s intention than from social position and the institutional field in which the utterance occurs.

Derrida went further, arguing that language was never fully tied to who said it, where it was said, or even whether it had a clear origin. AI didn’t start this, it just makes it more obvious and more extreme.

With large language models, those social conditions are still operating through interfaces, incentives, norms, and tendencies to anthropomorphise, but the speaker they were meant to invest has vanished. We are navigating an absence that feels present, the humanly embodied utterance without the human. Foucault would call this a ‘structured absence’.

The ethical weight that was once distributed between speaker and listener now falls almost entirely on the user. In Mikhail Bakhtin’s terminology, the user inherits answerability, the obligation to answer for what’s been made. According to Daniel Bashir, what we’re left with is speech shaped like address with no one behind it to address from. If it ever binds anyone, the binding happens downstream, when a human forwards it, files it, or lets it stand as their own.

Austin’s vocabulary remains remarkably useful because it helps us see where responsibility, authority, and judgement migrate when language is detached from a responsible speaker. Al does not perform speech acts in the ordinary sense, yet outputs can still be received as advice, judgement, explanation, or authority through uptake, interfaces, and institutional contexts. This shifts the focus from model capabilities to the conditions under which generated text is allowed to function as authoritative discourse.

Whether the statistical orchestration of human language constitutes a kind of doing, whether LLMs have something like a socially situated illocutionary force of their own, is domething to consider. As AI moves from disembodied text to robotic bodies with voices, gestures and proximity, we’ll need to revisit Austin’s framework to again explore what is being conveyed when some embodied conditions hold.