RMUNN: Falsus in uno's moral-lie model fits persons, not LLMs; measurable hallucination still yields comparable unreliability for info-seeking

The Gist

RMUNN says the old rule "false in one thing, false in everything" is about people who choose to lie. Chatbots are not choosing to lie; they do not really "know" truth that way. Even so, because they invent things at a measurable rate, you still should not treat them as steady sources when you need reliable information. This packet is a steelman reconstruction of RMUNN's Hacker News reply for LogicFirst import. It is not an endorsement of RMUNN, andrewdb, or any position in the thread.

Conclusion

Falsus in uno's moral-lie rationale applies to persons, not to LLMs as liars, but measurable hallucination rates still warrant treating LLMs as unreliable information sources in a practically comparable sense for reliability-seeking readers.

Premises

1. The maxim falsus in uno, falsus in omnibus traditionally links a person's willingness to lie about one thing to broader unreliability, because deliberate falsehood implicates moral or credibility character.
2. That moral-lie model does not cleanly apply to LLMs: they lack a moral component, and hallucination is not lying in the ordinary sense that requires knowing the truth and asserting the opposite.
3. LLMs, as RMUNN states the point, do not "know the truth" as an internal normative constraint in the way the lying analysis presupposes, so a single false output does not prove a moral defect.
4. Separately from moral character, a measurable tendency to hallucinate still makes a source unreliable for information-seeking in a way that is practically comparable to the unreliability ascribed to a person willing to lie.
5. Therefore one can reject anthropomorphic "LLM is a liar" talk while still treating hallucination rate as a sufficient unreliability marker for triage.

Assumptions

• "Lying" is used in the ordinary knowledge-and-intent sense RMUNN invokes. Research residual (confirmed direction): Frankfurt-style analysis distinguishes lying (oriented to falsity) from bullshit (indifference to truth); several 2024-2026 philosophy-of-AI papers argue LLM false outputs are closer to bullshit or structural indifference than to lying or to human hallucination. Legal falsus in uno is typically a permissive credibility inference about willful material falsehood by a witness, not an automatic rule, and is often criticized when stretched. Differs residual: calling LLM output…
• Differs: Even for persons, the maxim is typically optional inference, not automatic total disbelief.
• Differs: Bullshit is a contested label; steelman needs only not-lying plus rate-based unreliability.

Analysis

Overall strength: Moderate. Argument type: Deductive.

Premise Strength

• P1: falsus in uno traditionally links lying to moral/credibility character (Strong) — Accurately describes the doctrinal core of the maxim, though the argument's own residual assumptions correctly note it is a permissive inference, not an automatic rule - a nuance the premise itself doesn't fully carry but which the assumption block appropriately supplies.
• P2: the moral-lie model doesn't cleanly apply to LLMs (no moral component, hallucination isn't ordinary lying) (Strong) — Well-supported by standard philosophy of language and action (Frankfurt-style lying-requires-knowledge-and-intent analysis); this is the argument's most secure conceptual move.
• P3: LLMs don't 'know the truth' as an internal normative constraint (Moderate) — Presented as settled via RMUNN's characterization, but this is a substantive and contested claim in philosophy of AI (interpretability research on calibrated uncertainty, functionalist accounts of internal representations) that is treated with more confidence than the live debate warrants.
• P4: measurable hallucination tendency makes LLMs practically comparably unreliable to a willing liar (Weak) — This is the argument's load-bearing and most vulnerable premise. 'Practically comparable' is never operationalized against any shared metric or threshold, no actual hallucination-rate data or human-lying baseline is cited, and rates are known to vary enormously by domain, task, and prompt - undermining the claim that a single rate can ground a stable comparability judgment.
• P5: therefore reject 'liar' talk while retaining rate-based unreliability marker (Moderate) — Follows validly as a conjunction of the two prior sub-conclusions, but fully inherits P4's undefended comparability claim, so its practical force is only as strong as that weakest premise.

Potential Fallacies

• Underspecified operationalization / unsupported empirical premise (P4, and by inheritance P5) — P4 asserts that 'measurable' hallucination rates yield 'practically comparable' unreliability, but no rate, benchmark, threshold, or comparative human baseline is ever supplied. The term 'measurable' implies empirical grounding that the argument doesn't actually deliver, leaving the central comparability claim asserted rather than demonstrated.
• Motte-and-bailey retreat-and-reoccupy (Transition from P2/P3 to P4/P5) — The argument concedes the strong, hard-to-defend claim (LLMs are moral liars) and retreats to a modest one (LLMs are statistically unreliable), but then reimports the practical/rhetorical force of the original claim via 'practically comparable' without independently justifying why the same downstream response (skepticism, triage-level distrust) is warranted for a mechanistically different kind of error.
• Categorical/continuous conflation (P4, in relation to P1) — Falsus in uno operates as a near-categorical, discretionary credibility inference (one proven willful lie can discredit the whole testimony), while hallucination rate is a continuous, context-dependent frequency amenable to proportional discounting. Treating these as yielding 'comparable' practical unreliability blurs a binary distrust posture with a probabilistic calibration posture, risking either overstated distrust or an incoherent policy that borrows the totalizing force of the discredited maxim.
• Reliance on unspecified authority (P1-P3, and A1's research residual) — The argument's negative claim (P1-P3) leans heavily on a single Hacker News comment and a gestural reference to unnamed '2024-2026 philosophy-of-AI papers' without citation specifics, making key conceptual premises difficult to independently verify even though they are largely defensible on their own philosophical merits.

Counterarguments

• P4/P5 (practically comparable unreliability) (High impact) — Without a specified metric or threshold, a skeptic can simply ask 'comparable how, and comparable to what baseline lying rate?' The comparison as stated is rhetorical rather than argued, and under any strict operationalization (e.g., expected value of trusting an output) hallucination and strategic lying likely diverge sharply, since hallucination is non-adversarial and often correctable by calibration in ways deliberate deception is not.
• P4 (High impact) — If any nonzero, measurable error rate is sufficient to license liar-comparable distrust, then virtually all imperfect epistemic sources - honest experts, memory-fallible witnesses, translators - become 'practically comparable to liars' whenever their error rate is nontrivial. This reduction trivializes the very distinction (moral vs. non-moral unreliability) the argument was built to preserve.
• P4 (Medium impact) — Hallucination rates vary drastically by domain, task, model version, and prompt engineering; treating 'the' hallucination rate as a stable, portable property that can ground blanket practical comparability ignores this heterogeneity, and undermines P5's claim that rate alone is a 'sufficient' triage marker.
• P2/P3 (Medium impact) — The Frankfurtian 'bullshit' framing gestured at in the assumptions suggests structural indifference to truth may be a distinct and arguably worse epistemic vice than lying, not a morally neutral non-lie; if adopted, this could change the comparability calculus in P4 rather than simply confirming the not-lying claim.
• Conclusion (Medium impact) — Stripping moral/liar language from LLM failures, while practically useful for precision, risks creating a responsibility gap: no culpable liar to blame and, without explicit reattachment to developer/deployer accountability, potentially reduced urgency for fixing the underlying causes of hallucination.

Suggested Improvements

• Operationalize 'practically comparable' — Specify the metric (e.g., expected decision cost, calibration error, downstream harm rate) and threshold used to judge comparability, and cite actual benchmark figures (e.g., TruthfulQA, HaluEval, FActScore) alongside a comparative human baseline (e.g., witness/eyewitness error rates). This is the argument's single most exploitable weakness; without it, the central 'still comparable' claim - the paper's advertised contribution - remains asserted rather than demonstrated.
• Distinguish categorical from continuous unreliability regimes — Explicitly separate 'discount proportionally to measured error rate in a given domain' from any language of comparability to lying, since falsus in uno's totalizing binary logic and hallucination-rate calibration point toward different, sometimes contradictory, practical recommendations. Prevents the argument from smuggling back the totalizing force of the maxim it explicitly disclaims, and clarifies what triage behavior is actually being recommended.
• Acknowledge domain/task heterogeneity — Note that a single aggregate hallucination rate is a composite of very different failure mechanisms (fabricated citations, reasoning slips, outdated knowledge) that vary by context, and that instance-level or domain-level reliability signals may be more actionable than an aggregate rate. Improves real-world applicability of the triage recommendation in P5 and avoids overgeneralizing from a monolithic treatment of 'LLMs' as a class.
• Reconnect to accountability — Add explicit discussion of who bears responsibility for hallucination-driven harms (developers, deployers) now that moral blame is removed from the model itself. Prevents the reframing from being read or used as an implicit exculpation for institutional actors who could otherwise be held to harm-reduction standards.

Scenario Tests

• A model exhibits a stable, low hallucination rate on factual QA but a much higher rate on niche technical domains or adversarially prompted queries. (Challenges) — Undermines the claim that a single 'measurable hallucination rate' can ground blanket practical comparability; supports domain-specific rather than aggregate triage.
• A jailbreak or prompt injection causes an LLM to produce confident, contextually tailored falsehoods that mimic strategic deception. (Challenges) — Suggests the 'not lying, just less reliable' framing may understate risk in adversarial contexts, where LLM output can look functionally indistinguishable from intentional deception despite lacking intent.
• Policymakers or courts adopt the 'practically comparable to a liar' framing as license for categorical distrust of LLM outputs. (Challenges) — Would reintroduce exactly the totalizing moral-culpability-style blanket disbelief the argument sought to avoid, revealing an unaddressed slippery-slope risk in P4/P5's phrasing.
• A rigorous framework is developed comparing calibrated hallucination rates against documented human witness error/lying base rates in matched decision contexts. (Supports) — Would give P4 the empirical grounding it currently lacks, converting the comparability claim from asserted to demonstrated and substantially strengthening the argument's practical payoff.

Coherence & Relevance

The argument is structurally coherent as a conjunction of two independent sub-claims - a well-grounded conceptual disanalogy (LLMs aren't moral liars) and a separately asserted practical claim (hallucination rate still warrants comparable distrust) - and it avoids strict equivocation by explicitly bifurcating moral and rate-based senses of unreliability. However, the coherence is more formal than substantive: the argument's persuasive payoff depends entirely on the undefended comparability claim in P4, and reasonable readers will find that the argument 'has it both ways,' rejecting the analogy's premises while retaining much of its practical force without proportional justification.

• P1 (Strong) — Accurately sets up the doctrinal baseline, though it should note (as the assumptions do) that the maxim is permissive rather than automatic, or risk overstating the strength of the analogy being rejected.
• P2 (Strong) — None significant; this premise does the real conceptual work of disanalogy and is well-supported.
• P3 (Moderate) — Relies on a contested claim about LLM internal states being treated as settled fact rather than an open question in philosophy of AI.
• P4 (Weak) — This is the critical gap: no operationalized metric, threshold, or comparative data connects 'measurable hallucination' to 'practically comparable unreliability.' The inferential leap from moral disanalogy to practical equivalence is asserted, not argued.
• P5 (Moderate) — Validly follows as a conjunction of P1-P3 and P4, but is only as strong as P4, which is the weakest link in the chain.

View this argument on LogicFirst.ai