Humanising LLM Outputs Is Dumb

Most people responding were not defending today’s default LLM writing. They were fed up with it. The dominant complaint was that Claude, ChatGPT, and similar models now produce long, jargon-heavy, metaphor-soaked text that is harder to read than a direct answer. Several people said this is now a real productivity drag in coding workflows. The useful distinction they landed on is that “humanized” output and “readable” output are not the same thing. People want terse, plain, technical language. They do not want faux empathy, invented metaphors, third-person use of their name, or engagement bait. The sharper point that emerged is that this style problem likely comes from post-training, not from the raw pretraining corpus. Commenters with experience on open-weight models said the weird tics show up much more strongly in instruction-tuned variants than in base models, which fits the idea that RLHF and task tuning are teaching a specific house style. That also explains why model vendors differ. Some people found Anthropic especially bad here, while others said Gemini or GPT become similarly flowery once the product adapts to your preferences. Where the discussion settled was practical. People are already building two-pass workflows where one model or step does the work and another step compresses or translates it for human consumption. Others want this exposed as a product feature so formatting sits outside the core reasoning path. A few went further and said frontier labs should optimize first for precise machine-facing output, then let downstream tools render it for humans however they want. The broad conclusion was not that natural-language interaction was a mistake. It was that mixing reasoning, presentation, and relationship-signaling into one chatty surface is starting to break down for serious technical use.