You Can't Blame the AI for What Your Document Says

An AI tool can smooth out your writing in seconds. It cannot take responsibility for what the final version says. That is the line software engineer Sophie Alpert draws in a short internal policy she published on using AI to write at work:

“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. If a reviewer asks, ‘What did you mean by this line?’, it’s not acceptable to reply with ‘Oh sorry, AI wrote that, just ignore it.’”

Who Alpert is

Alpert currently works in engineering at Clay, a company that builds tools for sales and marketing teams to automate research and outreach. Before that, she led the React core team at Facebook, the small group responsible for one of the most widely used tools for building websites and apps, and later held VP of Engineering roles at Notion and Humu. She has spent years writing internal documents that other engineers rely on to make decisions, which is the exact situation her policy is written for.

What she gets right, and where it’s incomplete

Alpert’s central claim is that language does not survive editing intact: “There are no lossless transformations of natural-language text, every rewrite and rephrase changes the meaning of your writing, and if this is done by an entity that doesn’t have the most detailed mental representation of what you personally were trying to communicate, information will be lost.” An AI model rewriting your draft was never in your head. It cannot know which word you chose on purpose and which one you’d change if you reread it. That gap is real, and it explains why AI-smoothed writing so often reads as confident but says less than the original draft.

She also makes a case that goes beyond risk management: “writing is thinking,” and skipping the writing step by handing a rough idea to an AI tool means skipping the thinking too. A document is proof you worked through a problem, not just a deliverable to produce quickly. That argument holds up in her own examples: brainstorming with AI or having it proofread a draft you already wrote keeps the thinking with you. Handing it a bullet list and asking for a finished document does not.

Where the policy is thinner is on where exactly the line sits. Alpert lists acceptable uses, brainstorming, drafting assistance, and proofreading, but a lot of everyday AI writing use falls in between those categories and a full rewrite. If you ask an AI tool to tighten three sentences you wrote yourself, have you crossed into territory where you no longer have “the most detailed mental representation” of the result? The policy sets the right standard for accountability. It does not give you a bright line for when you have quietly stopped meeting it.

Why it’s notable

Most guidance on AI-assisted writing focuses on catching factual errors or hallucinations, cases where the model states something false. Alpert’s argument works even when everything the AI writes is technically true. The risk she describes is losing meaning you intended but the AI didn’t know to preserve, which no fact-check will catch, because the sentence isn’t wrong. It just isn’t fully yours anymore, and you’re the one who has to answer for it in a meeting or a review.

What it means for builders

If your team uses AI to draft pull request descriptions (the write-up explaining a code change before it gets merged into the project), incident writeups, or product specs, the test Alpert proposes is simple to apply: could you defend every sentence in a live conversation without checking the original prompt? If a teammate asks what a paragraph means and your honest answer is “let me check what I asked the AI,” the document has already failed its job.

The practical fix isn’t banning AI from your writing process. It’s using it earlier, for brainstorming and structure, and later, for proofreading, rather than in the middle, where it quietly rewrites the parts you actually meant to say yourself. Before you publish anything AI touched, read it as if you’ll be asked to explain any line of it on the spot. If you can’t, it isn’t ready.

End of article