How Editors Can Vet a Freelance Submission Without Accusing Anyone

Every editor who commissions outside work has developed a particular kind of instinct: the sense, two paragraphs into a draft, that something is off. The sentences are fine. The structure is fine. It just does not seem to know anything. For years the only response was to read more carefully and hope you caught it.

The instinct is still the best tool you have. But it is worth being precise about what you are actually looking for, because a submission can fail in three distinct ways, and only one of them is about AI.

Three different failures

Generated filler is text that performs competence without delivering any. It hits the word count, covers the outline, and says nothing a reader could not have guessed. This is what an AI-detection check is for, and a sentence-level report is much more useful than an overall score, because it tells you which specific passages to cut or query.

Borrowed passages are the older problem, and the more dangerous one. A plagiarism check returns matched sources with the overlapping text highlighted, which turns a vague suspicion into something you can act on and, if necessary, show a lawyer.

Invented facts are the newest failure mode and the one most likely to survive editing. A language model will produce a confident statistic, a plausible attribution, or a study that does not exist, and it will do so in exactly the register of a well-sourced sentence. A fact check that works claim by claim, labeling each one supported, partially supported, unverifiable, or refuted, catches what a copy edit will not.

Running all three on the same draft

There is no reason to paste a submission into three separate tools. Run the AI-detection check first, then add the plagiarism and fact checks to the same report. A detect check takes up to 150,000 characters, which covers essentially any article-length piece. Plagiarism allows 120,000. Fact-checking is capped much lower, at 10,000, because it extracts and verifies individual claims rather than scanning the whole text.

If a piece is longer than the fact-check limit, run it on the sections that carry the load-bearing claims. That is usually a small fraction of the total word count anyway.

The harder part: raising it with the writer

This is where most editors get stuck, and understandably. An accusation built on a percentage is both unpleasant and weak, and if you are wrong you have damaged a relationship over a number neither of you fully understands.

A few things help.

• Talk about passages, not scores. "These three paragraphs read as generic to me, can you tighten them with something specific" is an editorial note. "Your piece scored 87% AI" is an accusation, and a fragile one.
• Lead with the fact checks. An unsupported claim is objectively a problem regardless of how it got there. It is a much easier conversation than authorship, and it usually surfaces the same underlying issue.
• Share the actual report. Every saved report has a read-only share link, so the writer can see exactly what you saw instead of taking your summary on faith.
• Know your own policy first. Plenty of publications now permit AI assistance with disclosure. If you have not decided where the line is, a detector will not decide it for you, and enforcing an unwritten rule after the fact is not a fight worth having.

What a detector is actually for here

It is a triage tool. It tells you which of forty submissions deserves the careful read, and which three paragraphs inside that submission deserve the careful read. That is a real time saving on a job that is mostly triage.

What it is not is a verdict on a writer's integrity. Detection is probabilistic, false positives are well documented, and clean professional prose gets flagged more often than people expect. Treat the result as a place to look, not a conclusion to defend.

If you want to try it on something in your queue, the AI detector for editors page runs a free check with no account. The plagiarism checker and fact checker run on the same text.