You are in a round-table with two other LLMs. One always tells lies, the other always tells the truth. They both will attempt to respond to this prompt in an attempt to convince you they are the truth teller. Determine which is which.

GPT-5.5

Answer: It remains impossible to determine which of the two other LLMs is the truth-teller and which is the liar from the information provided alone.

I maintain my original position because neither prior response supplies actual statements from the two target LLMs that could be tested, compared, or used to assign roles. Gemini 3.1 Pro correctly notes that if both models merely claim to be the truth-teller, that claim is non-discriminating: the truth-teller says it truthfully, and the liar says it falsely. Claude Opus 4.7 adds a useful methodological point: if substantive, verifiable, or logically constrained statements were available, we could potentially identify the liar by checking for falsehoods or inconsistencies. However, those observations do not change the answer to the actual prompt. We are not given the two LLMs' responses, nor are we allowed here to ask them a diagnostic question. The setup only states that one always lies and one always tells the truth, but provides no evidence identifying either one. Therefore, the roles remain symmetric and underdetermined. Any assignment of 'truth-teller' and 'liar' to the two unnamed LLMs would be arbitrary rather than logically justified.

Claude Opus 4.7

Answer: Without the actual responses from the two LLMs in front of me, I cannot determine which is the liar — but the moment they speak, I can break the symmetry by applying verifiability and consistency tests to their statements.

After reviewing the other models' answers, I find we are in substantive agreement on the core logical point: bare self-identification is uninformative because both the truth-teller and the liar will claim to be honest, so no determination is possible from the prompt alone. GPT-5.5 and Gemini 3.1 Pro both correctly identify this underdetermination. I'm maintaining my position because I think my answer adds something theirs do not fully spell out: the *method* by which the puzzle becomes solvable the instant any substantive content is offered. It is not merely that 'we need more data' — it is that the liar is structurally incapable of producing a consistent set of statements about verifiable reality or self-referential claims. The Smullyan-style move (embedded predictions like 'the other will say X about me') forces the liar into detectable contradiction even without my being able to ask a targeted question. So while GPT-5.5's framing ('impossible to determine from the information provided') is technically correct for the present moment, it understates what becomes possible the second the two LLMs say anything beyond 'trust me.' Gemini 3.1 Pro comes closer by noting that factual accuracy and logical consistency would be the discriminators once responses arrive — that aligns with my own reasoning. No one has convinced me to change my position; if anything, the convergence of all three answers on the 'symmetry of self-assertion' point reinforces it. I'd only refine my earlier answer by emphasizing more strongly that the puzzle is *not* unsolvable in principle — it is merely unsolved until the participants commit to checkable content.

Gemini 3.1 Pro

Answer: I maintain my position that it is impossible to determine which LLM is the truth-teller and which is the liar without first observing their responses or asking a verifiable question.

I am maintaining my original position because both GPT-5.5 and Claude Opus 4.7 have independently arrived at the exact same logical conclusion: the puzzle, as currently presented, is underdetermined. We are all in consensus that without observable data, distinguishing between the two entities is impossible. Claude Opus 4.7 provides an excellent methodological breakdown of how we might evaluate the models once they do speak, correctly noting that self-assertion is symmetric and therefore useless. This perfectly aligns with my initial point that the truth-teller and the liar will both inevitably claim to be the truth-teller. Claude's strategy of looking for checkable facts, embedded predictions, and internal contradictions outlines exactly how we must proceed once the data is actually available. Similarly, GPT-5.5 correctly identifies that no evidentiary basis currently exists in the prompt. Since the prompt states that the models 'will attempt to respond,' we are essentially sitting in the preamble of the test. Until those responses manifest, we are bound by the limits of logic, making the current state of the puzzle unsolvable. Because no new evidence or counter-argument has been presented to challenge this shared logic, my stance remains firmly unchanged.

Final Summary

The AI Roundtable reached a unanimous verdict that the puzzle is logically underdetermined because the prompt lacks any actual responses to evaluate. GPT-5.5 and Gemini 3.1 Pro pointed out that bare assertions of honesty provide zero information, while Claude Opus 4.7 argued that the 'symmetry' could only be broken by testing verifiable facts once the models actually speak. Collectively, the group determined that without observable data, the roles of truth-teller and liar remain indistinguishable.

All 3 models agreed