Humanising LLM Outputs Is Dumb

The cultural shift towards AI tools is evident on platforms like X, viral GitHub repositories, and Hacker News. One noticeable trend involves prompting models with instructions like having ADHD or generating outputs in simplified technical English. While the desire for more concise and easy-to-understand language is understandable, the reporter argues that humanizing LLM outputs is the wrong approach.

The reasoning behind this perspective is that applying these instructions after the model has finished working is problematic. By asking the model to continuously compress its output into a lower-bandwidth format, important details may be lost, and the output may still read nicely without the reporter realizing what has been omitted.

The ASD-STE100 Simplified Technical English example highlights this issue, as it aims to make documentation unambiguous for humans, but an LLM is not a human technical writer, and the raw output often contains the most information-dense representation available.

Furthermore, when agents communicate with one another, humanization becomes even more convoluted. A subagent may generate a human-readable summary of its findings, which is then processed by the parent agent to create another summary for the user. While tests may pass, there is a valuable lesson to be learned here. Humanization can mask failures such as conflicting evidence, unresolved branches, stack traces, and uncertain assumptions.

In other systems, like databases, compilers, and APIs, the highest-fidelity representation is maintained until the last moment before being transformed into a human-readable format. LLM tooling, however, appears to be doing the opposite.

The reporter emphasizes that this does not mean accessibility or personalization should be disregarded. If users desire three-line answers or outputs in simplified technical English, they should be accommodated. However, the reporter suggests that these instructions should be applied at the end, rather than as operating instructions within the model.

Agents should retain detailed state, and subagents can exchange schemas, diffs, exact errors, confidence levels, and provenance. Only then should the outputs be compressed for the user's benefit.

The viral skills demonstrated by users might actually point towards the correct future. Prompting models to speak in a language that is easy for humans to understand, such as "talk to me like I have ADHD," is perfectly reasonable as a renderer but not as an operating instruction. The ideal solution would involve agents operating in precise, machine-facing state, with the warm, concise, and human-friendly version generated only at the boundary. Thus, the viral repositories represent a bug report rather than the end state.