There's No Lossless Rewrite: Why AI-Polished Docs Still Need an Owner
Simon Willison flagged a sharp internal policy from engineer Sophie Alpert: every AI-assisted rewrite loses something, so you're still on the hook for every sentence you ship.
By TRAGenX Desk
Simon Willison spotted a short, pointed piece worth passing around: Sophie Alpert's internal policy on when it's acceptable for engineers to use AI to help with writing. The title states the thesis plainly — there are no lossless transformations of natural-language text.
The claim: rewriting is never neutral
Alpert's argument is simple and hard to argue with. Every rewrite or rephrase changes meaning, even subtly. If the entity doing the rewriting — an LLM, an editor, anyone — doesn't share your exact, detailed mental model of what you were trying to say, some of that intent gets flattened or replaced on the way out. That's true whether a human ghostwrites your memo or claude -p rewrites your PR description.
The rule: you still own every sentence
You must stand behind every idea and every sentence in your docs. It is your responsibility to make sure that the entire document is representative of your own thoughts before you share it.
— Sophie Alpert
Willison calls out the practical consequence: if a reviewer asks "what did you mean by this line?", it's not acceptable to answer "oh, AI wrote that, ignore it." That response wastes the reader's time twice over — once reading a sentence you don't actually stand behind, and again waiting for you to produce the version you meant.
Why this matters beyond prose
The policy is framed around writing, but the logic doesn't stop at docs. Anyone shipping AI-assisted output — code reviews, commit messages, incident postmortems, even generated trading strategy notes — is making the same bet: that they reviewed closely enough to close the gap between what the model produced and what they actually mean.
- Treat AI drafts as a first pass, not a final answer — read every line as if you'll be asked to defend it.
- If you can't explain a sentence (or a code change) without reference to "the AI wrote it," you haven't reviewed it closely enough to ship it.
- This applies symmetrically to AI-assisted code: a PR you can't walk a reviewer through line-by-line is a liability, however clean the diff looks.
None of this is an argument against using AI for writing or code. It's an argument against outsourcing the judgment. The tool can draft; the accountability doesn't transfer.
FAQ
Frequently asked questions
- What does "no lossless transformation of natural-language text" mean?
- It means any rewrite or rephrasing — by a human editor or an LLM — inevitably shifts meaning unless the rewriter shares your exact, detailed understanding of what you intended to say. Some nuance is always lost or altered in translation.
- Is Sophie Alpert's policy against using AI for writing?
- No. It sets a condition for acceptable use: whoever publishes a document is responsible for standing behind every sentence in it, regardless of whether AI helped draft it. The policy targets unreviewed AI output, not AI assistance itself.
- How does this apply to AI-assisted coding, not just writing?
- The same logic holds: a generated commit, PR description, or function is not exempt from ownership just because a model wrote the first draft. If you can't explain why a line does what it does, it isn't ready to ship.
Sources