如何使用 AI 校對，避免事實被改動

The biggest risk in AI proofreading is not a missed typo. It is the AI quietly rewriting names, numbers and quotes while it "tidies up" the text. The fix is simple to state: let the AI point out problems, and keep the actual rewriting and the fact-checking as human steps.

Plenty of freelance writers now run drafts through an AI before delivering them to clients. For catching typos and slips, it is fast and often more consistent than reading by eye. But anyone who does this regularly also runs into edits nobody asked for. This article walks through those failure patterns and a workflow that keeps them under control.

Why invented facts are worse than typos

Hand a model a draft with "please make this easier to read" and it will do more than fix typos. It will rewrite whole sentences into something plausible. Models like ChatGPT and Claude are not red-pen tools; they are built to generate natural-sounding text, so smoothly reconstructing a passage comes more easily to them than returning only the lines they changed.

During that reconstruction, changes like these can slip in:

• Official names get swapped for common ones (where "Inc." sits in a company name, the capitalization of a product name, and so on)
• Numbers and dates get rounded, or filled in with a value that "fits" the context
• Text inside quotation marks gets paraphrased and no longer matches the original
• Extra explanations that were never in the draft get added, looking like helpful context

These rewrites tend to increase as the scope of the request grows. Asking for proofreading together with "soften the tone" or "cut this in half" makes it much more likely that factual parts move too. It is safer to ask for proofreading on its own.

A missed typo is easy to recover from: someone points it out, you apologize, you fix it. But if a number or a quote changes and the piece ships that way, you have delivered content that is factually wrong. That kind of mistake damages trust with clients and readers, and because the rewritten text reads naturally, a quick skim won't catch it. That is why invented facts are more dangerous than typos.

A three-step workflow where the AI only flags issues

The core idea is to restrict the AI to the role of reviewer. The workflow has three steps.

Step 1: tell it not to rewrite. Ask something like: "Do not rewrite the text. List each problem as a bullet in the form original / suggested fix / reason." Then decide on each suggestion yourself. It takes longer, but the facts in the draft can no longer drift on their own, and reading the reasons also teaches you a lot about your own writing habits.

If the project has a style sheet (a list of house rules such as "email" vs "e-mail"), include it with the prompt. Consistency checks become far more accurate. Without it, the AI nudges you toward whatever spelling is most common in general, which may contradict the client's rules.

Step 2: check the diff. When you do want the AI to rewrite, never save its output over the original. Always compare the two. Word's Compare feature or any text diff tool will show only what changed. Looking at a diff even once makes it clear how many unrequested edits get mixed in.

Step 3: verify facts against the source. Names, numbers, dates, quotes and URLs get checked against the original material, whether or not the AI flagged them. This step cannot be delegated. AI sometimes proposes "plausible mistakes" with full confidence, and the absence of a flag is not evidence that a fact is correct.

This may look like overkill, but not every piece needs the full treatment. For a blog post under your own name, step 1 is often enough. For deliverables that contain prices or contract terms, step 3 deserves real time. A good rule is to scale the effort to how much damage a wrong fact would cause.

One more practical tip: don't paste a long draft in one go. As the input grows, the feedback on the later sections can get noticeably thinner.* Splitting the text by section, or into chunks of a few thousand words at most, keeps the quality even.

* Models can only take in a limited amount of text at once, and attention to the middle of the input tends to drop as you approach that limit. The context window glossary entry explains the mechanism concisely.

Quick reference: what to delegate and what to keep

As long as a human owns the bottom half of this table, the fact-related items, you can lean on AI heavily for the top half without serious accidents. For many writers, proofreading time does shrink with AI, but it does not drop to zero. The time saved is better spent on fact-checking. A proofreading AI works best not as a replacement for a proofreader, but as a first reader that rarely skims past anything.

FAQ

Which AI is best for proofreading?
Claude is often rated highly for natural-sounding prose, and many writers use it for review. But any model can make edits you didn't ask for. A workflow where the AI only flags issues and you verify the facts yourself matters more for quality than the choice of model.

Is it safe to put a client's confidential draft into an AI proofreader?
First check whether your contract or NDA allows entering the material into external services. Even if it does, confirm in the terms of service that your plan or settings exclude your inputs from training. If neither condition can be met, another option is to proofread with a local LLM that runs entirely on your own machine, for example through a desktop app like osFoundry.

Can I leave typo checking entirely to AI?
AI is usually more consistent than the human eye at spotting typos, so it is well worth using. Still, read the piece through once before delivery. The point is less to catch what the AI missed and more to confirm that each fix you accepted still fits the surrounding sentences.