LLM Voice

Original Article ↗ Hacker News Discussion ↗

LLM-generated prose is increasingly recognizable by a bland, over-padded “AI voice” that many readers find exhausting, uninformative, and subtly off, even when they can’t fully articulate why. Commenters worry that as more writing is delegated to models—especially in public-facing content—human style, originality, and even the cognitive benefits of writing will erode, drawing analogies to ultra-processed food and “Sysco slop” in restaurants. Some see a legitimate role for AI in translation, bureaucratic documents, coding, or drafting for people with weaker language skills, but argue that personal communication and thinking-intensive work should remain firmly human if we still want writing to be worth reading.

Perception of “LLM Voice”

  • Many find LLM-generated prose trite, bloated, low–information density, and oddly exhausting to read.
  • Readers describe a “gut feel” or uncanny valley: sentences are grammatical but don’t “fit” human minds or conversational expectations.
  • Recurrent complaints: generic cadence, repetitive structure, over-explaining, faux-punchy constructions (“it’s not X, it’s Y”), sentence-length paragraphs, and shallow content stretched into long essays.
  • Some worry they’re unconsciously imitating LLM style in their own writing.

When LLM Writing Is Acceptable (or Not)

  • Widely seen as fine for: bureaucratic reports, TPS-style documents, forms, design docs, legal/scientific boilerplate, and translation—contexts where voice doesn’t matter and the “audience” is often a process.
  • Many argue it’s disrespectful to use LLMs for personal essays, blog posts, fiction, or emails where a human “I” is speaking, especially without disclosure.
  • Strong norm proposed: okay for private use or drafting; problematic when publishing or representing work as fully one’s own.

Food Analogy & Ethics of Effort

  • Repeated analogy: home-cooked meals vs frozen TV dinners / Sysco slop.
  • LLM text is likened to serving $3 microwave burritos at a dinner party or restaurant without saying so.
  • Principle offered: readers shouldn’t be asked to invest more effort in reading than the writer put into writing.

Thinking, Learning, and Delegation

  • Many equate writing with thinking and fear that outsourcing prose means outsourcing thought and comprehension.
  • Concern that heavy reliance on LLMs erodes discipline, depth of reasoning, and mutual understanding in reviews, code discussions, and technical work.
  • Others counter that it’s a tool: you can still think deeply if you choose, and the real issue is user intent and work context.

Impact on Internet Content

  • Strong sentiment that the web is being flooded with “AI slop”: SEO spam, AI-written blogs, product pages, support answers, even tech support emails.
  • This drives people to older sources, comments, or pre-2023 content, and makes real human writing harder to find.
  • Some see this as an extension of earlier SEO-driven enshittification, now greatly accelerated.

Code vs Prose

  • Many are comfortable with LLM-generated code but not prose.
  • Reasons: code can be tested objectively; prose depends on voice, subtlety, and genuine perspective.
  • Still, some warn that overuse in coding harms abstractions, naming, and maintainability—i.e., the communicative aspects of code.

Detection, Style, and Research

  • People report both explicit “tells” and a more diffuse sense of wrongness.
  • One cited study claims LLMs use more inanimate subjects, rarer n‑grams, longer dependencies, and structures optimized for information density rather than human communication.
  • Others note that base models are less slop-like and that instruction tuning and RL layers (often guided by LLM judges) may be what impose the homogenized “chatbot” voice.

Norms and Suggested Practices

  • Use LLMs as critics, fact-checkers, translators, and brainstorming partners, but write final prose yourself if voice matters.
  • If you do publish LLM-assisted text, many advocate clear labeling and/or sharing the underlying bullet points instead of inflated essays.
  • Some intentionally keep typos and idiosyncrasies as authenticity signals and call for tools that critique without suggesting replacement wording, to avoid “infection” of voice.