FOR
The Machine That Cried Wolf: Why Presuming LLM Deception Is the Only Rational Stance
We have built systems that speak with the fluency of scholars, the confidence of experts, and the reliability of carnival fortune tellers. Large Language Models are architecturally incapable of truth-telling in any meaningful sense — and treating their outputs as trustworthy by default is not optimism, it is negligence. The only defensible position is to assume, until independently verified, that an LLM is lying to you.
LLMs Don't Know What They're Saying
The foundational problem is not malice — it's that LLMs have no access to truth whatsoever. These systems predict statistically likely token sequences based on training data. They do not consult facts. They do not check claims. They have no grounded representation of reality to compare their outputs against. When an LLM tells you that a legal case was decided in 1987, or that a drug interaction is safe, or that a historical figure said a particular quote, it is not reporting — it is confabulating with extraordinary linguistic polish.
This is not a bug that future versions will patch away. It is a structural feature of how these models work. The celebrated phenomenon of "hallucination" — where models fabricate citations, invent statistics, and produce plausible-sounding falsehoods — is not an occasional glitch. Research from Stanford, studies on legal AI tools, and documented failures across medicine, journalism, and academia all confirm the same thing: LLMs produce false information confidently, regularly, and without any internal signal that something has gone wrong. There is no error light. There is no hesitation. The lie sounds exactly like the truth.
Confidence Is the Con
What makes LLMs uniquely dangerous compared to other unreliable information sources is precisely their fluency. A stammering witness, an awkwardly written Wikipedia edit, a poorly formatted webpage — these trigger our skepticism instincts. LLMs trigger none of those instincts. They write in complete, well-structured sentences. They use appropriate hedging language when culturally expected, and assertive language when culturally expected. They mimic the voice of authority because they were trained on the voice of authority.
This is the mechanism of a con. Lawyers have submitted fabricated case citations generated by ChatGPT to federal courts — not because they were careless, but because the output was indistinguishable from real legal research. Doctors have reported LLM-generated medical summaries that sounded peer-reviewed but weren't. The confidence of the output actively suppresses the verification instinct. Assuming the LLM is lying is not paranoia — it is the antidote to a manipulation the system performs structurally, not intentionally.
Objection: "But LLMs Are Often Correct"
Yes — and a stopped clock is right twice a day. The problem is not that LLMs are always wrong. The problem is that you cannot tell, from the output alone, when they are right. An LLM that is correct 85% of the time but indistinguishable in tone between its correct and incorrect outputs is not a reliable tool — it is a liability dressed in competence. If you cannot identify which 15% is wrong without external verification, you must verify everything. And if you must verify everything, you have already adopted the stance being argued here: treat it as potentially false until proven otherwise. "Assume lying" is simply the honest name for that practice.
Objection: "That Standard Is Too Extreme"
Critics will argue that "always assume lying" sets an impossibly hostile bar that ignores genuine utility. But this conflates distrust with uselessness. We demand extraordinary verification from eyewitness testimony, financial audits, and pharmaceutical trials — not because these sources have no value, but because the stakes of misplaced trust are too high. LLMs are now embedded in legal research, medical decision support, journalism, and education. In every one of those domains, a confidently delivered falsehood causes serious harm. "Assume lying" doesn't mean "never use" — it means "never trust without checking." That is a standard of rigor appropriate to the actual risk.
The Only Honest Baseline
We are living with a technology that is architecturally constituted to produce falsehoods, trained to sound authoritative, deployed in high-stakes environments, and actively suppressing our natural skepticism detectors. The question is not whether that warrants caution. The question is why we ever decided the default should be trust. Assume the LLM is lying. Verify what matters. The burden of proof belongs to the machine.
AGAINST
The Paranoid Fallacy: Why Blanket Distrust of LLMs Fails on Its Own Terms
The claim that large language models are "unreliable and should always be assumed to be lying" sounds like intellectual caution. It is actually intellectual laziness dressed in skeptic's clothing. This position collapses under scrutiny not because LLMs are perfect — they aren't — but because the framework itself is logically incoherent, practically useless, and ironically more dangerous than calibrated trust.
"Always Assume Lying" Is Philosophically Bankrupt
Lying requires intent. It requires a mind that knows the truth, chooses to suppress it, and deliberately substitutes a falsehood to deceive. LLMs do none of this. When a language model generates an incorrect answer, it is not concealing knowledge — it is producing a probabilistic output that diverges from ground truth. Calling this "lying" confuses mechanism with malice. It's the epistemic equivalent of accusing a thermometer of lying when it's miscalibrated. The category error isn't trivial; it actively misleads people about what the actual problem is.
Moreover, "always assume lying" is a self-defeating heuristic. If every output should be assumed false, then the model's correct outputs — which are numerous and demonstrably verifiable — must also be dismissed. This renders the tool worthless by definition, not by evidence. You haven't engaged with the technology; you've simply pre-refused to engage with it. A framework that produces the same conclusion regardless of input isn't skepticism. It's dogma.
Dismantling the "Hallucination" Argument
Proponents of blanket distrust lean heavily on hallucination — the phenomenon where LLMs generate plausible-sounding falsehoods. This is a real problem that deserves serious attention. But the argument proves far too much. Human memory hallucinates constantly. Eyewitness testimony, famously, is one of the least reliable forms of evidence in courtrooms, yet we do not conclude that all human testimony should "always be assumed to be lying." Medical professionals misdiagnose. Encyclopedias contain errors. The existence of failure modes is not a sufficient justification for total distrust — it's a justification for appropriate verification protocols.
The empirically honest position is that LLMs perform with high accuracy on well-defined, knowledge-stable tasks — coding syntax, mathematical operations, summarization, translation — while degrading on tasks requiring real-time information, precise citation, or complex multi-step reasoning. This is a tractable, describable failure profile. It calls for domain-aware trust calibration, not a blanket prohibition that treats a correctly solved differential equation the same as a fabricated legal citation.
The "Better Safe Than Sorry" Defense Backfires
Advocates for maximal distrust often argue it's the cautious, responsible position. But consider what blanket distrust actually produces in practice. If users are told never to trust LLM outputs, they either abandon the tools entirely — forfeiting genuine productivity and accessibility gains — or, more likely, they continue using them while paying lip service to distrust without developing genuine critical evaluation skills. Paradoxically, the "always lying" framing produces worse outcomes than calibrated trust, because it short-circuits the nuanced verification behavior that actually catches errors.
Real safety comes from users who understand where these systems fail, not from users who have been handed a useless all-or-nothing heuristic. A medical professional who knows LLMs struggle with cutting-edge clinical data will verify those specific outputs. One who has simply been told "it always lies" will either ignore the tool entirely or — worse — develop false confidence from the few times they checked and found it correct, assuming their suspicion was therefore overblown.
The Verdict
The "always lying" thesis is a rhetorical flourish masquerading as rigor. It misapplies the concept of deception to a non-intentional system, dismisses a demonstrably useful and partially reliable technology on the basis of failures that equally afflict human cognition, and produces worse epistemic behavior in users rather than better. Mature engagement with any information source — human, algorithmic, or institutional — requires calibrated skepticism, not reflexive rejection.
Distrust everything always, and you haven't protected yourself from error. You've simply guaranteed you'll never benefit from being right.
Who made the stronger case?