● Covered by 1 source · 1 reportLow impact1 negative
Humanizing LLM Outputs Leads to Lossy Information and Hides Failures
Key points
- Humanizing LLM outputs compresses information lossily.
- This practice can hide failures and critical details.
- LLMs should maintain high-fidelity representations until human consumption.
- The approach is compared to other systems that preserve raw data.
The Problem with Humanized LLM Outputs
A growing trend involves instructing Large Language Models (LLMs) to produce outputs in human-friendly formats, such as using simplified technical English or avoiding jargon. While this aims to make LLM interactions more palatable, it is argued that this approach is fundamentally flawed.
The issue arises because these instructions are applied during the generation process, not after. This forces the LLM to continuously compress its output into a lower-bandwidth format, leading to a loss of information that often goes unnoticed because the output still reads smoothly.
Lossy Compression and Hidden Failures
Using styles like ASD-STE100 Simplified Technical English, while designed for human clarity, becomes problematic when applied to LLM agents. An agent is not a human technical writer, and its raw output often represents the most information-dense state. Applying style rules alongside task instructions results in a lossy compression of the original data.
This humanization also effectively hides failures. LLM agents can fail in useful, albeit 'ugly,' ways, such as conflicting evidence, unresolved branches, or stack traces. Human prose tends to smooth over these issues, presenting them as 'considerations' rather than clear indications of problems, making it harder to identify and address underlying issues like hallucinations or token window limitations.
Analogy to Other Systems
Other technical systems, such as databases, compilers, and APIs, operate on the principle of maintaining the highest-fidelity representation of data for as long as possible. Data is only transformed into a human-readable format at the boundary where a human consumes it, such as a dashboard or user interface.
Current LLM tooling, by contrast, is increasingly reversing this principle, transforming information into a human-friendly format too early in the process. This premature transformation sacrifices the richness and accuracy of the raw data, hindering effective debugging and understanding of the LLM's internal workings.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
The daily brief
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Today's brief
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's brief
New every morning, and the back catalogue is archived by date.
Primary sources
GitHub ayghri/i-have-adhdReporting from
The practice of instructing Large Language Models (LLMs) to humanize their outputs, such as using simplified language or specific styles, is criticized for being counterproductive. This approach compresses information in a lossy way and can obscure critical details or failures within the LLM's operation, making it harder to debug or understand the underlying process.