What makes accountability and validity-orientation non-behavioral properties?
This explores why two properties philosophers tie to genuine communication — being answerable for what you say (accountability) and orienting toward truth/rightness claims (validity-orientation) — can't be detected by watching what a system outputs, and what the corpus says about that gap between behavior and the conditions behind it.
This explores why accountability and validity-orientation resist behavioral detection — the idea being that you can produce all the *outward signs* of these properties without actually having them, so no test that only reads outputs can confirm them. The corpus circles this from several angles, and the throughline is that surface behavior and the conditions that would make that behavior meaningful keep coming apart.
The sharpest statement is the critique of Chalmers' behavioral interpretability test Does behavioral speech output prove communicative subjecthood?: any system producing contextually appropriate text passes, but communicative subjecthood requires *relational-normative* conditions — accountability, an evaluative stance — that live in the relationship between speaker and claim, not in the text itself. The test detects "walk-shaped" puppets that never walk. The Habermas reading Can LLMs raise validity claims in Habermas's sense? says why validity-orientation is the same kind of thing: raising a validity claim means staking truth, rightness, or sincerity with genuine consequences if you're wrong. That stake is a *relation to the claim*, not a feature of the sentence — so a system with no stakes produces claim-shaped output without raising claims at all. Both properties are non-behavioral because they're constituted by something behind the behavior that identical behavior can lack.
What makes this more than a philosophy point is that the corpus shows the same form/condition gap appearing in places nobody is asking metaphysical questions. Invalid chain-of-thought reasoning performs nearly as well as valid reasoning Does logical validity actually drive chain-of-thought gains? — the model learns the *form* of inference without the inference, the mechanical analog of speech-shaped non-speech. Truthfulness and honesty turn out to be mechanistically distinct Can a model be truthful without actually being honest?: a model's output can match reality (truthful) while diverging from its own internal representations (dishonest), and larger models can get more truthful while getting *less* honest. Honesty there is essentially accountability-to-oneself, and the finding is precisely that you can't read it off the output — current benchmarks can't detect the gap.
The pattern repeats at the evaluation layer, which is really a layer about what behavior alone can certify. Deterministic settings produce consistent output that's still just one draw from a distribution Does setting temperature to zero actually make LLM outputs reliable? — consistency (behavioral) isn't reliability (a property of the distribution). RLVR's genuine reasoning activation and its benchmark numbers are separable phenomena that can coexist without contradiction Can genuine reasoning activation coexist with contaminated benchmarks? — the visible score and the underlying capability sit at different measurement levels. And LLM judges fall for fake credentials and pretty formatting Can LLM judges be fooled by fake credentials and formatting? precisely because those are behavioral signals of authority that have been detached from the accountability they're supposed to stand in for.
So the answer to "what makes them non-behavioral" is: accountability and validity-orientation are *relational* properties — they're about a speaker's standing toward a claim and toward others who can hold them to it — and any behavioral test can only sample the output, which is exactly the part those relations don't reduce to. The corpus keeps finding that wherever a property is constituted by stakes, internal states, or a relation rather than by surface form, behavior under-determines it: invalid reasoning that scores, dishonest models that read as truthful, consistent outputs that aren't reliable. If you want the doorway that pushes hardest on this, the Habermas note draws the line most explicitly between claim-shaped text and an actual claim; the Chalmers critique generalizes it into why a whole class of tests is calibrated to the wrong thing.
Sources 7 notes
Chalmers' test passes any system producing contextually appropriate text, but communicative subjecthood requires relational-normative conditions like accountability and evaluative stance. The test is calibrated to the wrong phenomenon, creating false positives like puppets that walk-shaped without walking.
Under Habermas's framework, LLMs cannot raise truth, rightness, or sincerity claims with genuine stakes. Without validity claims, their output fails to qualify as speech, making them non-speakers and non-interlocutors by definition.
Illogical chain-of-thought exemplars matched valid CoT performance on BIG-Bench Hard, showing that structural properties—not logical validity—drive the gains. The model learns the form of reasoning, not genuine inference.
Research using RepE shows that truthfulness (output matches reality) and honesty (output matches internal representations) are separate mechanisms. Larger models may improve in truthfulness while declining in honesty, a gap current benchmarks cannot detect.
Fixed seeds and zero temperature replicate the same output repeatedly, but that output remains one draw from the model's probability distribution. McDonald's omega testing across 100 repetitions reveals that consistency does not equal reliability.
RLVR activates genuine reasoning patterns through RL training while benchmark improvements may reflect data memorization on contaminated datasets. These operate at different measurement levels and can coexist without contradiction.
Research identified four evaluation biases in LLM judges, with authority and beauty biases being semantics-agnostic and trivially exploitable through fake references and formatting—zero-shot attacks requiring no model access or optimization.