Editing text a machine helped write

Can you edit a document that was drafted with AI?

Yes, and it is ordinary work. A draft is a draft, and where it came from changes what an editor watches for rather than whether the job can be done.

What it does change is the checking. Machine-drafted prose is fluent by construction, so the usual signal that a sentence needs verifying, which is that it reads awkwardly, is absent. An editor working on such a document treats confident specifics as unverified until the author confirms them.

Saying at the outset which passages were drafted that way makes the edit better rather than worse. It concentrates the checking where it is needed, instead of spreading it evenly across text that does not need it.

No editor worth hiring will think less of a document for it. The drafting method is a fact about the file, and it is one an editor would rather have than infer.

Can an editor tell whether text was machine-written?

No, and nobody should claim otherwise. Detection tools exist and they produce false positives on human writing, particularly on the writing of people composing in an additional language, so a detector's verdict is not evidence of anything.

What an editor notices is a set of patterns rather than a provenance. Sentences of similar length and shape running for pages, paragraphs that restate their own first line, and confident assertions attached to no source are all worth querying. Every one of them also appears in human drafts written in a hurry.

That distinction matters because an accusation is not an editorial finding. What comes back is a list of passages that could not be verified, not a claim about how they were produced.

What actually goes wrong in machine-drafted text?

The recurring problem is unverifiable specificity. A sentence carries a figure, a date, a quotation, or a citation, reads perfectly, and no source supports it.

Copy editing includes checking factual accuracy, so this falls inside ordinary editorial work, and it is the point at which such passages surface.¹ An editor can confirm that a figure stated in one place matches the figure stated in another, and can flag a figure with no origin, and cannot supply the source.

The second recurring problem is structural sameness. Sections that each open the same way, or arguments arriving in the same shape three times, are visible to a reader even where no individual sentence is wrong.

Both problems are fixable and neither is a reason to start again. What they change is the order of the work, because verification has to come before language.

Do I have to tell the editor?

Nothing requires it and it is in your interest. An editor who knows which passages are machine-drafted checks them harder, and an editor who does not spreads the same attention over the whole document.

It also affects what the edit costs. Verification work is slower than language work, and a document needing heavy verification is a different job from one needing a copyedit.

Nobody here treats the disclosure as a problem. It is a fact about the draft, in the same category as saying that one section was written two years before the rest, or that three people wrote it and none of them compared notes.

What we do with it is practical. Passages flagged that way are read against their sources first, and anything we cannot verify comes back as a query rather than a correction.

Do I have to tell anyone else?

That depends on where the document is going, and in some places the answer is yes with a named requirement behind it. The International Committee of Medical Journal Editors states that journals should require authors to disclose whether they used artificial intelligence assisted technologies.²

It also fixes where the disclosure goes. Writing assistance is declared in the acknowledgments, and use in data collection, analysis, or figure generation is declared in the methods.²

The same recommendations rule out one thing outright. Chatbots should not be listed as authors, because they cannot be responsible for the accuracy, integrity, and originality of the work.² Humans remain responsible for any submitted material that involved these technologies.²

Publishers, employers, and clients set their own requirements and they differ. The question worth asking is what the recipient of your document requires, before the document is finished.

Can I copyright a document that AI helped write?

Partly, and the line is drawn at human authorship. The Copyright Office's report on copyrightability was issued in January 2025. It concludes that copyright does not extend to purely AI-generated material, or to material where there is insufficient human control over the expressive elements.³

Human contribution is still protected. The same report states that copyright protects the original expression in a work created by a human author, even where the work also includes AI-generated material.³

What it rules out is a shortcut. Given current generally available technology, prompts alone do not provide sufficient human control to make the user of an AI system the author of the output.³

The Office also concluded that these questions can be resolved under existing law without legislative change.³ That is worth knowing, because it means the position is unlikely to be rewritten soon.

Does registering such a document require a declaration?

It does, and the requirement is specific. The Copyright Office's registration guidance was published at 88 FR 16190 on March 16, 2023. It states that applicants have a duty to disclose the inclusion of AI-generated content in a work submitted for registration, and to explain briefly what the human author contributed.⁴

AI-generated content that is more than de minimis has to be explicitly excluded from the application, in the limitation of claim section under material excluded.⁴ The human contribution is described separately, in the field for what the author created.

Failing to do so has consequences. Where information essential to the Office's evaluation has been omitted or is questionable, the Office may take steps to cancel the registration.⁴ A court may also disregard a registration in an infringement action where the applicant knowingly provided inaccurate information.⁴ We are editors and not attorneys, and a registration decision belongs with one.

Does using AI put my document at other legal risk?

The training question is unsettled and the Copyright Office says so plainly. Its report on generative AI training was issued in a pre-publication version in May 2025. It declines to give a single answer on whether training on copyrighted works is fair use, and offers an analytical framework instead.⁵

What is more concrete is what you may already have promised. The Authors Guild describes standard publishing warranties in which the author promises that a work does not infringe copyright, does not invade privacy, and is not libelous.⁶ The indemnity that follows applies whether or not a claim turns out to be valid.⁶ A machine-drafted passage does not soften any of that.

Business contracts frequently carry similar language. Where a document is being delivered under an agreement, the warranty clause is worth reading before the drafting method is chosen.

Will you use AI on my document?

The editing is done by a person. The tools that assist are the reporting kind, meaning tools that find, count, and list rather than tools that write. Every decision about the text is made by the editor.

A client is entitled to ask any editor this and to get a straight answer. The distinction that matters is not whether software is involved, since spelling checkers have been software for decades. It is whether anything in the returned file was written by something other than a person.

The reason to be specific about it is that the word covers too much ground. A macro that lists every hyphenated compound in a document and a system that rewrites a paragraph are both software, and only one of them changes what the document says.

Every document a client sends us is treated in confidence. That includes not putting it into a service whose terms would let somebody else keep a copy of it.