LLM Writing Tropes.md
LLM-generated writing is increasingly easy to spot because models lean on a narrow set of overused rhetorical patterns—like “it’s not X, it’s Y” pivots, faux-profound metaphors, emotional qualifiers (“genuinely,” “honestly”), and car-ad-style hype. Commenters link these tics to instruction tuning and reinforcement learning that push models toward attention-grabbing, high-polish prose, which becomes grating when repeated at scale and bleeds into human writing online. Many see value in tools that flag such tropes or use LLMs only as draft generators, while warning that overconfident AI-detection and collapsing writing diversity could erode trust and further homogenize web content.
Perceived LLM Writing Tropes
- Commenters recognize recurring constructions: “It’s not X — it’s Y,” car‑ad / movie‑trailer tone, faux profundity, overuse of words like “genuine,” “honestly,” “quietly,” and metaphor-heavy phrasing (“tapestry,” “camaraderie”).
- Structural tells include: title–subtitle headlines with colons, parenthetical “(why this matters)” endings, overlong README files, pros/cons tables for trivial points, “The [Noun] [Noun]” headers, “This changes everything,” and heavy first/second person (“We’ve all been there,” “Your first instinct…”).
- Different models have distinct quirks (e.g., “I’ll shoot straight with you,” “Fair enough,” snarky personalization, tech metaphors for everything), but converge on similar persuasive, flattened styles.
Why Tropes Arise (Speculated in Thread)
- Several participants blame instruction tuning / RLHF and mode collapse: base models are said to be more stylistically varied, whereas post‑training pushes toward a single “highly rated” style.
- Human raters and chat-oriented training may reward dense, “helpful” language, leading to nominalizations, longer words, melodrama, and repeated rhetorical tricks.
- Some suggest polluted training data (SEO slop, AI‑generated text, corporate/press‑release style) and lack of diverse incentives.
Detection, Overdiagnosis, and Limits
- Many think raw LLM output is easy to spot today due to bad/flat writing and trope overuse, but expect this to get harder.
- Others warn of a self-reinforcing “AI witch hunt”: people call any polished or em‑dash‑using text “AI,” with no way to verify authorship.
- There’s concern about the information ecosystem being flooded with mediocre auto-generated content, but some argue that many human posts are already low value.
User Strategies and Prompting Challenges
- Negative instructions (“don’t do X”) often fail or backfire; models seem to fixate on mentioned tropes (“pink elephant” effect).
- Suggested mitigations: treat LLM output as a first draft, then manually edit or use a separate “editor agent” to strip tropes; specify a human writing style instead of lists of bans.
- Some users adjust model “personality” settings (e.g., less warmth/enthusiasm) or self-host media and tools to avoid AI‑slop and engagement-optimized platforms.
Attitudes Toward AI-Assisted Writing
- One camp sees AI-written prose as aesthetically bad, manipulative, or even deceitful if presented as personal writing.
- Another camp prioritizes clarity and usefulness over provenance, viewing LLMs as acceptable collaborators or superior to much human output, while acknowledging broader societal and incentive problems.