The push to make large language models sound more human has drawn sharp criticism from developers and researchers. A recent discussion, captured by the phrase Humanising LLM Outputs Is Dumb and the resulting Comments on Hacker News, argues that anthropomorphizing AI creates more problems than it solves. The core complaint is straightforward: treating probabilistic text generators as conversational partners misleads users about their capabilities and limitations.

What You Need to Know

The debate centers on whether making LLM outputs feel human improves usability or merely conceals the technology's mechanical nature. Critics contend that human-like responses encourage over-reliance and obscure the fact that models lack true understanding. Proponents of raw, less filtered outputs argue that transparency builds better user trust and safer deployment.

The Anthropomorphization Trend

Many companies now train LLMs to adopt conversational tones, use emojis and mimic empathy. This design choice aims to make interactions feel natural, reducing friction for casual users. But the approach has a downside. When a model says "I think" or "I feel," users may attribute consciousness or intent that does not exist. The Humanising LLM Outputs Is Dumb critique points out that this framing sets expectations for coherence and reliability that the underlying technology cannot consistently meet.

Risks of Human-Like Responses

Humanizing outputs also carries practical risks. Users might trust a model's confidence on factual questions, even when it generates false information. In high-stakes fields like medicine or law, such misplaced faith can lead to harmful decisions. The Comments on the original thread highlight examples where politeness or apology phrases from a model made users think the AI was "admitting error" rather than simply producing a plausible-sounding next token.

Critics propose a more transparent approach: let models state uncertainty plainly, avoid personality-driven language and signal their probabilistic nature. This does not mean making outputs unusable, but rather stripping away performative social cues that add no informational value.

  • Over-trust: Users may rely on human-like LLM outputs for critical decisions without verification.
  • Misattribution: Emotional language in responses can create false sense of AI consciousness or intent.
  • Undermined debugging: Anthropomorphized outputs make it harder to identify and fix model errors.

Why This Matters

The debate is not academic. As LLMs are embedded into search engines, customer service and educational tools, the way outputs are presented shapes user behavior. If the industry continues down the path of humanizing outputs without safeguards, the gap between perceived and actual capability will widen. This could erode trust when models inevitably fail. Developers, product managers and regulators must decide whether conversational gloss is worth the cost of clarity. The Humanising LLM Outputs Is Dumb argument, supported by many in the Comments community, offers a clear alternative: treat AI as a tool, not a person.