The defective AI speech act is a speech act form whose sincerity condition is structurally unfulfillable. The notion goes back to John R. Searle himself (Speech Acts, 1969): defective are those speech-act forms in which the felicity conditions are not satisfied or cannot be satisfied — non-serious, parasitic, fictional, or mechanically produced acts.
Defect by Violation vs. Defect by Structure
Two categorial cases:
Defective by violation: The conditions could be fulfilled but are violated in the concrete case. Example: the lie — an assertive speech act under violation of the sincerity condition. The liar could believe what he says; he does not. Conceptually his speech act remains intelligible — precisely because he violates the condition whose fulfillability is presupposed.
Defective by structure: The conditions cannot be fulfilled at all, because the bearer of the act lacks the prerequisites. Example: the LLM output. The model has no speaker who could have a psychological state. An “assertion” by an LLM is neither a truthful nor a lying assertion; it lies structurally outside the truth–lie game (cf. AI Truth-Indifferent Utterance).
Application to AI Language Models
Murray Shanahan (Talking About Large Language Models, CACM 2024) draws the consequence: the LLM itself is “neither truthful nor untruthful, in any everyday sense of these terms”. Its “promises” cannot be kept — not because the model did not try, but because there is no “it” there that could try. Its “assertions” have no bearer who could defend, correct, or retract them. Its “apologies” are form without contrition, because there is no subject to repent.
Thus every seriously meant assertive, commissive, expressive, or directive speech act issuing from an LLM is a defective speech act in Searle’s sense — provided one takes the full sense of speech-act theory as one’s basis.
Consequence for Practice
Defective speech acts are not worthless. A theatrical performance, a small child’s prayer, a computer-game command — they are all defective in the technical sense (in Searle’s idiom: only serious speech acts satisfy the full sincerity condition), yet functionally meaningful. LLM outputs too are useful — as a proposal, a translation aid, a sorting function.
The point is not depreciation but conceptual discipline: what counts as an assertion, what as a promise, what as a declaration of truthfulness, should remain ontologically clean. Whoever deals with defective speech acts as though they were genuine overlooks the decisive difference.
Ontological Classification
- is subclass of: speech act
- distinct from: genuine speech act / truthful speech act
- requires only: derived intentionality
- typically occurs in: AI Conversation Simulation
Sources: Generated by querying the Personhood ontology.
Further sources:
- Searle, John R. (1969): Speech Acts. Cambridge: CUP, ch. 2–3 (on defective speech acts).
- Searle, John R. (1979): Expression and Meaning. Cambridge: CUP.
- Shanahan, Murray (2024): Talking About Large Language Models. CACM 67(2).
- Frankfurt, Harry G. (1986/2005): On Bullshit. Princeton: PUP.