LLM rhetorical-pattern tells in prose
Parent: Writing and Documentation · Published reference · snapshot 2026-09-23 · skill writing-expert/references/rhetorical-pattern-tells.md
↓ Facts as markdownall context files
Source-anchored research for LLM rhetorical-pattern tells in prose. Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Definitions
- Burstiness is defined as how much writing patterns and perplexities vary over a whole document; the claim is that humans vary naturally while LLMs choose next words by the same rule and so show low burstiness. (medium confidence) [source]
- Hype/promotional language is a studied problem independent of LLMs: biomedical NLP work formalizes automatic detection of promotional phrasing that overstates research significance. (low confidence) [source]
- Tricolon is not proof of AI authorship: corporate communications, grant applications and student essays show the same habit, and prompting can remove the tell; the signal is cumulative with other tells. (medium confidence) (verified as of 2026-02-13) [source]
How it works
- Hypothesised causes: prevalence in training text, and human RLHF reviewers rating the construction highly because it signals nuance and insight. Nobody, including model makers, is sure; OpenAI's model-behavior staff prefer the label 'contrastive phrasing'. (medium confidence) [source]
- Reader-side harm: cognitive-psychology work cited by The Conversation says negation does not erase the negated concept; readers process it first and often lose the negation tag, so the pattern plants the strawman it denies. (medium confidence) [source]
- Suggested fix: lead with what the thing is, drop the defensive strawman, and use a contrast only when it is specific and non-obvious. (medium confidence) [source]
- Research on word overuse finds focal words (delve, tapestry, testament, pivotal, underscore, showcasing, crucial, paramount, unwavering) spiked in scientific English and are overused by ChatGPT; human-feedback learning is implicated because participants prefer text containing such words. (medium confidence) [source]
- Watch-list participles named by the Wikipedia guide: highlighting/underscoring/emphasizing, reflecting/symbolizing, contributing to, cultivating/fostering, enhancing. It treats the tell as stronger when the verb's subject is an inanimate fact or event. (low confidence) (verified as of 2026-09-23) [source]
Measurements and reference values
- Style-based AI detection is unreliable: OpenAI's retired classifier caught 26% of AI text and falsely flagged 9% of human text; a Stanford study found non-native writing is misclassified. Promotional phrasing alone should not be treated as proof. (medium confidence) [source]
Comparisons and alternatives
- Promotional-tone watchwords in the same guide: 'boasts a', 'vibrant', 'rich', 'profound', 'nestled', 'in the heart of', 'groundbreaking', 'renowned', 'diverse array'; output reads like a travel guide or advertisement rather than neutral encyclopedic prose. (medium confidence) (verified as of 2026-09-23) [source]
- Reinhart et al. (CMU/NJIT, arXiv 2410.16107, Oct 2024) found instruction-tuned models use present participial clauses at 2-5x the human rate: GPT-4o 527%, GPT-4o Mini 481%, Llama 3 70B Instruct 261%, 8B Instruct 224%, versus base models near human levels (94-102%). (low confidence) [source]
Facts and statements
- Wikipedia's Signs of AI writing lists negative parallelism ("Not only X but Y", "It is not just X, it's Y", "no X, no Y, just Z") as an LLM tell. (high confidence) [source]
- Wikipedia states LLMs use negative parallelism about three times more frequently than humans; the article rests on a single source, Will Oremus's 2026 Atlantic piece, and carries a single-source maintenance notice. (medium confidence) (verified as of 2026-09-23) [source]
- The Atlantic (via AI Weekly summary) reports Barron's counted the phrase in Fortune 500 filings rising from 50 in 2023 to over 200 in 2025. (low confidence) (verified as of 2026-09-23) [source]
- Humans use the device too (listicles, ad copy, speeches); Wikipedia says isolated instances are not conclusive and it must be corroborated by other signs. The tell is frequency and repetition within a text. (high confidence) [source]
- Across six LLMs from three families, human news text shows more scattered sentence-length distributions, more varied vocabulary and dependency types than LLM text; differences between humans and LLMs exceed differences between LLMs. (low confidence) [source]
- GPTZero stopped relying on perplexity and burstiness as primary detection signals in autumn 2023, moving to a deep-learning classifier; the stats remain one of seven indicators. (medium confidence) (verified as of 2026-09-23) [source]
- Perplexity/burstiness are relative to a specific scoring model, so results shift with each model release; closed models expose no token probabilities, forcing open-model proxies. (low confidence) [source]
- Text heavily present in training data (e.g. the Declaration of Independence) scores as low perplexity and is misflagged as AI; non-native English writing also scores low perplexity/burstiness (2023 Stanford TOEFL study), so the metric produces false positives. (low confidence) [source]
- Wikipedia's AI-writing guide lists significance-inflation markers: 'stands/serves as', 'is a testament', 'crucial/pivotal role', 'underscores its importance', 'reflects broader', 'setting the stage for'; LLMs attach broader symbolic meaning to arbitrary facts, e.g. a name change 'highlighting the enduring legacy'. (medium confidence) (verified as of 2026-09-23) [source]
- The Wikipedia guide says GPT-4-era models were more blatantly positive than newer ones, which use subtler positivity, so lexical lists date quickly. (low confidence) (verified as of 2026-09-23) [source]
- Wikipedia's Signs of AI writing guide says chatbots attach a present-participle (-ing) phrase at sentence end to add superficial analysis of significance, recognition or impact, sometimes with vague attributions; on Wikipedia such comments are usually synthesis or unattributed opinion. (medium confidence) (verified as of 2026-09-23) [source]
- Base Llama 3 models show human-like participial rates, so the excess is introduced by instruction tuning, not pretraining. (low confidence) [source]
- The pattern is only a potential sign, not proof: the Wikipedia guide says such patterns also occur in human writing (editorials, fan fiction) and human style is increasingly influenced by LLMs. (medium confidence) [source]
- Stylometric detectors relying on surface features show high false positives and poor domain transfer (Pudasaini et al., arXiv 2603.23146), so a participial-phrase count alone should not be used to accuse text of being AI-written. (low confidence) (verified as of 2026-09-23) [source]
- Wikipedia's Signs of AI writing guide says LLMs overuse the rule of three in forms from adjective-adjective-adjective to short phrase, short phrase, and short phrase, often to make superficial analyses look more comprehensive. (medium confidence) [source]
- Proposed explanation: LLMs reproduce a technique that is effective in human rhetoric (e.g. 'I came, I saw, I conquered') but lack the judgment to hold back, so saturation dulls the effect. (medium confidence) [source]
- Heuristic: if nearly every enumeration in a paragraph comes in threes, the text is likely AI output; humans break the pattern with two, four, or long lists. (medium confidence) [source]
Children
- Awards-and-recognition boilerplate (frontier)
- Contrastive negation and false-dichotomy framing
- Excess-vocabulary lists (frontier)
- Four- and five-item list extension (frontier)
- Human style convergence from LLM exposure (frontier)
- Inanimate-subject participles (frontier)
- Perplexity as detection signal (frontier)
- RLHF-induced rhetorical overuse (frontier)
- Rhythm uniformity and burstiness
- Sentence-length variance metrics (frontier)
- Significance inflation and promotional puffery
- Trailing participial analysis phrases
- Travel-brochure register (frontier)
- Tricolon and rule-of-three lists
- Tricolon of near-synonyms (frontier)
- Vague attribution to third parties (frontier)
Frontier under this node: Awards-and-recognition boilerplate, Excess-vocabulary lists, Four- and five-item list extension, Human style convergence from LLM exposure, Inanimate-subject participles, Perplexity as detection signal, RLHF-induced rhetorical overuse, Sentence-length variance metrics, Travel-brochure register, Tricolon of near-synonyms, Vague attribution to third parties