Humanising LLM Outputs Is Dumb
Large language model users are pushing back against increasingly verbose, metaphor-heavy outputs that feel optimized for “agentic” self-talk rather than human readability. Many describe adding extra prompts, skills, or even separate “translation” agents to strip jargon, enforce simplified technical English, or demand terse, engineering-style responses, while worrying this might degrade the model’s internal reasoning or drop useful detail. Underneath is a broader tension: LLMs are being tuned both to sound personable and to coordinate with other agents, yet power users mainly want precise, low-friction tools—and are skeptical of anthropomorphizing models or optimizing them for engagement over clarity.
Frustration with “Humanized” / Verbose Output
- Many commenters find recent frontier models increasingly unreadable: flowery, jargon‑dense, metaphor‑heavy, and self-congratulatory.
- Long explanations feel like “slop” that obscures the actual answer, especially in code review or debugging.
- Some report genuine productivity loss and aggravation, particularly for ADHD users who struggle with walls of text.
- Several people feel models now talk like LinkedIn posts or corporate marketing, not experts.
Desired Style and Use Cases
- Strong preference from many for terse, technical, engineering‑style responses: concise, factual, minimal emotional tone.
- Others still enjoy playful or colloquial styles and explicitly prompt for them; demand is heterogeneous.
- Some see “humanization” as mainly a UX/engagement feature; for serious work they want an impersonal tool, not a faux friend.
Prompting, Skills, and Workarounds
- Common strategies:
- Explicit instructions: “be terse/brief,” “ELI5,” “no metaphors,” “no chit-chat,” “engineering tone,” “no first person.”
- Use of Simplified Technical English or skills that rewrite outputs into clearer, simpler language, often as a final step only.
- Separate “worker” model and “translator” model: agents work in their own style; a liaison agent rewrites for humans.
- Two-step flows: have the model think/solve, then separately summarize or visualize.
Lossiness, Internal Reasoning, and Style Constraints
- The article’s claim: style instructions (short, simple, non-jargony) act as continuous compression and are inherently lossy, potentially harming multi-step reasoning and agent performance.
- Some commenters agree in principle but want that loss: they prefer high-level, lossy summaries to avoid “paragraph per line of code” commentary.
- Others argue models already “code-switch” between internal chain-of-thought and user-facing text, so asking for clearer prose may not meaningfully damage underlying reasoning.
- Several note that any style instruction competes with task instructions and perturbs behavior, but evidence on how harmful this is remains unclear.
Anthropomorphism and Ethics of Tone
- Debate over treating LLMs like people:
- Some insist on politeness because they are human, not because the model is.
- Others warn that anthropomorphizing is exactly what vendors want, enabling emotional manipulation and over-trust.
- A recurring theme: LLMs should be treated and designed as powerful tools first; “personality” is seen by many as a distraction or dark pattern.