There are no lossless transformations of natural-language text

The post argues that natural-language cannot be transformed without loss, meaning any rewrite, whether by a human or an LLM, will change the original meaning. This stems from the idea that a model lacks the detailed mental representation of the author’s intent. Consequently, the notion of a "lossless" AI rewrite is fundamentally flawed.

For engineers who incorporate LLMs into their documentation workflow, the practical implication is clear: they must stand behind each sentence after AI assistance. If a reviewer asks for clarification, the engineer cannot deflect responsibility by blaming the model. This requirement enforces personal ownership of the content and prevents misunderstandings caused by subtle meaning changes.

Adopting this stance incurs a cost in time and effort. Teams will need to allocate additional review cycles to compare AI-generated drafts against the original intent, effectively reducing the speed advantage that LLMs promise. The policy does not forbid AI use, but it mandates a verification step before any AI-edited text is shared.

The limitation becomes apparent when the rewrite is performed by an entity without a deep understanding of the author's goals. In such cases, information loss is inevitable, making AI-only transformations unsuitable for precise technical specifications or legal language. Engineers must therefore limit AI assistance to contexts where minor semantic drift is acceptable.

Overall, the guidance reshapes how software teams approach AI-augmented writing: the tool can assist, but the final responsibility rests with the human author. This ensures that documentation remains an accurate reflection of the engineer’s thoughts, preserving clarity for downstream readers and reviewers.