Ask HN: How do you make AI writing usable?
A Hacker News user asks whether there are methods to make AI writing less formulaic, citing issues like contrastive negation and unnecessary lists, and questions if such "slop" is the new status quo.
Background
- "LLMisms" are distinctive patterns common in AI-generated text — contrastive negation (starting with "Not only..., but...") and bloated lists are the most notorious. Readers familiar with ChatGPT, Claude, or Gemini will quickly spot these tics.
- "Desloppify" is HN slang for removing the bland, generic tone ("slop") that LLMs default to. The complaint is that even state-of-the-art models produce writing that feels formulaic and soulless.
- No reliable fix exists yet. Users try elaborate system prompts ("avoid bullet points", "write like a cynical journalist"), temperature tweaks, or post-editing by hand — but the models' training data and RLHF fine-tuning strongly pull them toward a polished, risk-averse middle style.
- The deeper question here: is this a solvable engineering problem (better prompting, better fine-tuning) or an inherent limitation of next-token prediction? If the latter, "desloppified" AI writing may remain a contradiction in terms.