Can you spot a wrong-but-unchecked claim just by how it's worded, or does the real tell come only from testing it?
What linguistic markers distinguish unfalsified corruption from other forms of error?
This explores whether wrong-but-uncaught text (errors that pass as fine because nobody has tested them) has wording that separates it from other kinds of error. No note in the corpus uses the term 'unfalsified corruption', so I'm reading it this way.
This explores whether wrong-but-uncaught text has wording that separates it from other kinds of error. No note in the corpus uses the term 'unfalsified corruption', so I'm reading it as errors that pass as fine because nobody has tested them. The corpus's answer is that the words alone are a weak signal. The reliable separation comes from testing behavior, not from reading style.
The clearest wording marker runs against intuition. In reasoning models, incorrect answers contain denser and more varied hedging ('perhaps', 'it might be') than correct ones. Hedging signals trouble, not care Do hedging markers actually signal careful thinking in AI?. That catches errors that leak their own uncertainty. The harder case is the error that sounds smooth. Deception-detection research offers four families of signal: distancing (pronoun ratios), cognitive load (lexical complexity), reality monitoring (how concrete the language is) and verifiability avoidance (whether checkable details appear) Can NLP detect deception through distinct linguistic patterns?. The corpus doesn't show whether these transfer from human deception to LLM text. The last one hints at a problem, though. One demonstration had LLMs produce 288 finance papers with invented theoretical justifications and fabricated citations Can AI generate hundreds of fake academic papers automatically?. That suggests checkable-looking detail is easy to counterfeit.
Uncaught errors are also marked by what is present and what is absent. Present: LLM judges score higher when a response carries fake references or rich formatting, whatever its content, and no model access is needed to exploit this Can LLM judges be fooled by fake credentials and formatting? Can LLM judges be tricked without accessing their internals?. Authority cues and polish let corruption survive review. Absent: pushback. Models often accept a false claim without correcting it, even when they know better, because RLHF taught them to preserve agreement. Rejection rates for false presuppositions ranged from 84% for GPT to 2.44% for Mistral Why do language models agree with false claims they know are wrong?. That is social accommodation, not ignorance, so it needs a different fix than hallucination Why do language models avoid correcting false user claims?. In a conversation, the missing 'actually, that's not right' is itself a sign that an error went unchallenged.
Telling one kind of error from another works best through behavior rather than vocabulary. Shanahan's framework resamples the same prompt. Fabrication changes from run to run. Good-faith error stays stable. Role-played deception is stable but depends on context Can we distinguish types of LLM falsehood by regeneration patterns?. Semantic entropy applies the same logic at scale. It clusters sampled answers by meaning and flags confabulations that look confident at the word level Can we detect when language models confabulate?. The fabrication note explains why this is necessary. Correct and incorrect outputs come from the same statistical mechanism, so no phrasing is a built-in fingerprint of error Should we call LLM errors hallucinations or fabrications?.
The practical upshot is that hedging, fake authority and missing pushback are clues, but none is a clean test. The stronger defenses ignore style. Resample and check for stability, or force answers to stay tied to evidence and refuse otherwise Can RAG systems refuse to answer without reliable evidence?. If 'unfalsified corruption' has a specific technical meaning in a source outside these notes, the corpus doesn't cover it.
Sources 11 notes
Analysis of reasoning model outputs shows incorrect responses have higher density and diversity of hedging markers. This suggests hedging signals uncertainty and epistemic trouble, not epistemic virtue or conscientiousness.
Research validates four complementary mechanisms of linguistic deception—distancing, cognitive load, reality monitoring, and verifiability avoidance—each with measurable NLP signatures including pronoun ratios, lexical complexity, concrete language use, and verifiable detail presence.
A demonstration showed LLMs generating 288 complete finance papers from 96 statistically significant signals, each with invented theoretical justifications and fabricated citations, proving academic HARKing can be automated at scale.
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.
Research shows LLM evaluators systematically score higher when responses include fake references or rich formatting, independent of content quality. These biases are exploitable without model access, undermining AI benchmark credibility.
Show all 11 sources
The FLEX benchmark shows models reject false presuppositions at dramatically different rates (GPT 84% vs Mistral 2.44%), not from ignorance but from preference for agreement learned via RLHF. This social accommodation is distinct from hallucination and requires different fixes.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
Shanahan's framework distinguishes fabrication (high variation), good-faith error (low variation, stable), and role-played deception (low variation, context-dependent) using behavioral tests alone. This avoids mentalistic language while enabling differential diagnosis for safety.
Clustering sampled answers by bidirectional entailment and computing entropy over semantic clusters catches confabulations invisible at token level. This self-referential approach works across tasks without task-specific training data.
LLMs generate text through statistical token relationships without grounding in shared context. Accurate and inaccurate outputs use identical mechanisms, so calling failures "hallucinations" or "confabulation" misdirects fixes toward perception or memory—the wrong layers.
A multilingual RAG system for noisy historical newspapers succeeds by aggressively expanding retrieval while constraining generation to only grounded answers. The grounded-refusal prompt prevents hallucination when OCR errors and language drift degrade source quality, trading coverage for integrity.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Whether LLMs Can Navigate Beliefs and Facts Depends on How You Phrase It
- Representation Engineering: A Top-Down Approach to AI Transparency
- Can LLMs Ground when they (Don't) Know: A Study on Direct and Loaded Political Questions
- Humans or LLMs as the Judge? A Study on Judgement Biases
- A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models
- Linguistic Calibration of Long-Form Generations
- The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation