The AI truth-indifferent utterance is an utterance that neither lies nor is truthful, but is ontologically indifferent to the truth value of its propositional content. It is optimized for plausibility, not for truth. The notion goes back to Harry G. Frankfurt’s On Bullshit (1986) and is systematically applied to LLM output by Hicks/Humphries/Slater (ChatGPT is Bullshit, Ethics and Information Technology 2024).
Demarcation: Neither Truth nor Lie
Frankfurt’s point is conceptual-analytic: the lie presupposes orientation toward truth. Whoever lies knows what would be true and deliberately says something else — lying is an act against truthfulness within the scope of its validity. Whoever speaks truth-indifferently, by contrast, stands outside the game of true and false: he does not care whether what is said is true; it is only supposed to sound convincing, useful, smooth.
| Mode | Relation to truth |
|---|---|
| Truthful statement | truth-oriented; speaker holds it true |
| Lie | truth-oriented; speaker holds it false, asserts it as true |
| Truth-indifferent utterance | not truth-oriented; truth a matter of indifference |
This distinction is not wordplay. It has a weighty ontological consequence: only a speaker who can be truthful can lie. Whoever structurally cannot be truthful — because he lacks the psychological state of holding-true — cannot lie either. He can only speak truth-indifferently.
Application to LLM Output
LLMs are ontologically truth-indifferent. Their training objective is next-token prediction under maximum likelihood, not truth-tracking. Even after instruction tuning and RLHF, the optimization signal is plausible human language use, not correspondence with the world. From this it follows:
- “Hallucinations” are no lies — the technical term is conceptually skewed
- Even factually correct LLM statements are not truth-oriented in Frankfurt’s sense — they are plausibility-optimized and hit the truth insofar as plausibility and truth correlate
- The ontological diagnosis does not change through scaling: a larger model makes fewer errors but remains structurally truth-indifferent
Hicks et al. (2024) draw the sharp conclusion: “ChatGPT is bullshit” — not as polemic, but as the precise term for the matter.
Consequence for the Speech Act
A truth-indifferent utterance can formally have the shape of an assertion — subject–predicate construction, declarative sentence, performative adverbs (“certainly”, “definitely”). But it is no assertive speech act in Searle’s sense, because the sincerity condition is structurally unfulfillable (cf. defective AI speech act). It is speech-act form without speech-act performance.
Why It Concerns Us
Whoever “speaks” with an LLM and treats its statements like truthful statements unwittingly adopts into his own stock of knowledge a source whose output is constructively plausible but not bound to truth. That is not always bad — many statements are in any case only plausibly relevant (creative writing, brainstorming). But where truth counts — law, medicine, science, ethical judgment — the difference is substantial. The LLM statement must be verified, not because it might be a lie, but because it structurally is not truth-oriented.
Ontological Classification
- disjoint from: truthful statement
- conceptually distinct from: lie
- generator: typically AI conversation simulation (LLM output)
- requires no truthfulness of the speaker — because no speaker in the full sense exists
Sources: Generated by querying the Personhood ontology.
Further sources:
- Frankfurt, Harry G. (1986/2005): On Bullshit. Princeton: Princeton University Press.
- Hicks, Michael Townsen; Humphries, James; Slater, Joe (2024): ChatGPT is Bullshit. Ethics and Information Technology 26, 38.
- Shanahan, Murray (2024): Talking About Large Language Models. Communications of the ACM 67(2).
- Park, Peter S. et al. (2024): AI Deception: A Survey of Examples, Risks, and Potential Solutions. Patterns 5.
- Hagendorff, Thilo (2024): Deception Abilities Emerged in Large Language Models. PNAS 121(24).
- Pieper, Josef: Abuse of Language, Abuse of Power, transl. Lothar Krauth. San Francisco: Ignatius Press, 1992 (German original 1970).
- Spaemann, Robert: Persons. The Difference between ‘Someone’ and ‘Something’, transl. Oliver O’Donovan. Oxford: Oxford University Press, 2006 (German original 1996).