How LLMs do things with words
The concept of Performative Sentences comes from the philosophy of language.
In How to do things with words, J.L. Austin challenges the dogma that the purpose of a sentence is to describe reality, by pointing out the category of performative sentences that work differently altogether. To utter a performative sentence, is to perform some kind of action such as making a promise, agreeing to a transaction, giving a toast, making a declaration, giving permission, welcoming, thanking, apologising, et cetera. For example “I promise to water the plants” is not a description of the world, but rather, it is the case that the act of saying it makes it true that I enter into a promise to water the plants.
LLMs are interesting in this regard.
- Unless embedded in some context, a bare chat agent or LLM will utter meaningless performative statements. Sometimes it will hallucinate being a human and promise to do things but fail to do them.
- Can an LLM can genuinely apologise or express gratitude?
- LLMs have to say they are using the tool for them to use it. The way an LLM agent uses tools is typically by generating a JSON file in a certain format that describes the tool it wants to use (which must be one from a list of tools it is told it has access to) and the parameters with which it wants to use the tool. I think the generated JSON file counts as “utterance” for an LLM agent, and this utterance causes the agent programming logic to trigger and causes the agent to use a tool, thus it is performative.
- In a chain of thought, LLMs often say things like “I will do X”. If we say “I wonder xyz”, it is not performative but a description of our mental state. But if an LLM says this sentence, that utterance also makes it wonder. For LLMs their agency is fundamentally mixed up with their utterances, so what they say affects what they do more fundamentally than it does for humans.
It seems like LLMs can’t commit speech acts in the external world without more interfacing such as being given tools or being ascribed more social significance and power. On the other hand, their style of thinking by generating tokens based on previous tokens, makes more of the things they say performative utterances because they act by uttering and they choose how to utter based on their previous utterances.