August 10, 2026
Pretty words, missing facts
Humanising LLM Outputs Is Dumb
Stop making AI sound nice if it means hiding the important stuff, says fed-up internet
TLDR: The article argues that making AI sound friendlier too early can hide important details and mistakes, which matters if people rely on it for real work. Commenters were split between “yes, the fluff is unbearable,” “you’re describing the problem wrong,” and “just do the polishing afterward.”
The internet has found a new AI fight, and wow, people have feelings. The article’s big claim is simple: telling AI to sound extra neat, gentle, or “easy to read” can backfire if that polishing happens while it’s thinking instead of after. In plain English, the writer says making answers prettier too early can strip out the messy but crucial details — like exact errors, failed tests, or uncertainty — and leave users with bland, soothing mush.
That lit up the comments fast. One camp basically yelled, “Finally!” User Xcelerate compared modern AI writing to overblown literature that sounds impressive until you realize you absorbed absolutely nothing. Another crowd said the whole thing is being framed wrong: mikaeluman pushed back hard, arguing people aren’t trying to make AI more “human” at all — they want the opposite of chatbot fluff, meaning short, blunt, machine-like answers with none of the trendy business-speak. Yes, “blast radius” caught a stray.
And then came the classic internet solution guy. Havoc basically shrugged and said: why not just do it in two steps? First get the full answer, then ask for the simplified version. Problem solved, aside from the extra time and cost. Meanwhile, one commenter wandered in just to praise the blog’s side-panel mind map, which honestly feels very on-brand for online discourse. The funniest twist? Even a supposed summary comment accidentally proved the article’s point by flattening the argument into “cool, short code-like answers,” which is exactly the kind of oversimplification everyone is fighting about.
Key Points
- •The article says prompt-level style instructions for LLMs are increasingly used to make outputs shorter, simpler, or more human-friendly.
- •The article argues that applying these instructions during task execution causes lossy compression of information.
- •A central example is multi-agent workflows, where repeated summarization between subagents and parent agents can remove important technical detail.
- •The article states that human-readable prose can obscure failure signals such as uncertainty, conflicting evidence, stack traces, hallucinations, and token-limit problems.
- •The article proposes keeping precise, machine-facing state internally and applying personalization or accessibility formatting only at the final user-facing boundary.