How can I screen submissions for AI without wrongly accusing writers?

Never treat a probability as a verdict. Use a short workflow: screen with a tool, look for process evidence such as drafts and notes, talk to the writer, and record what you decided and why.

What should you not do when screening for AI use?

Do not act on a score alone, do not ask writers to prove a negative, do not demand that someone rewrite to sound less like a machine, and do not enforce a policy you have never published.

What evidence shows how a piece of writing was produced?

Drafts, version history, notes, sources and an oral follow-up distinguish authorship far more reliably than any classifier output. That is why process evidence should come before a score in any decision.

Blog · 28 September 2026

How to screen freelance submissions for AI: without wrongly accusing your writers

Every editor now has the same two fears: publishing AI-written work, and rejecting a human writer because a number said so. Both are avoidable, but only with a workflow, not with a score.

The trap: treating a probability as a verdict

AI detectors output a statistical estimate. Two error rates matter, and they pull in opposite directions:

False positives, real human writing flagged as AI. This is the one that costs you a contributor.
False negatives, AI writing that passes. This is the one that costs you credibility.

You cannot minimise both at once. Every detector has a dial, and where you set it decides which mistake you make more often. Most tools hide that dial. Ours does not, and the difference is large enough to be worth stating plainly:

Threshold setting	False positives on real human text	AI detected
strict, flags only clear cases	0.9%	91.3%
balanced, flags borderline cases too	5.0%	97.8%

Measured on 5,035 held-out human samples from 12 sources (0.6% when each source is weighted equally) and 92 AI samples, with strict train/test separation.

If you are screening submissions, use the strict setting. The balanced setting is for self-checking, where a soft flag is a useful prompt to look closer. It is not appropriate for triage, because a 5% false-positive rate across 200 submissions is ten human writers accused.

A four-step workflow that survives contact with reality

1. Triage at the strict threshold: and only the strict threshold

Run a fast pass at a conservative setting. Your goal is not to find every AI submission; it is to produce a short list worth a second look without generating noise you will have to apologise for.

2. Look at why it flagged, not just that it did

A useful detector returns the individual features behind the verdict. On our tool you get the weighted contributions. The ones that matter editorially:

Paragraph uniformity, AI paragraphs tend to be suspiciously even in length
Sentence-length regularity, human prose is bursty; machine prose is metronomic
Repetition of openers and transitions, "Moreover", "Additionally", "Furthermore" in a pattern
Simile and cliché density, the tell-tale "shimmering tapestry" register

A single high score with none of these structural markers is a much weaker signal than a moderate score with four of them together. This is where human judgement earns its place; you are reading for shape, not for a percentage.

3. Check the process trail before you check the writer

Most writers have one, whether they think of it that way or not: draft history, version files, notes, research bookmarks, earlier pitches, comments. Asking for process is a normal editorial conversation. Asking a writer to prove their humanity against a tool you chose is not, and it is a good way to lose them.

4. Make the human decision, and be able to explain it

Write down what you saw and why it mattered: flagged at strict threshold, structural uniformity, repetitive openers, and then, whatever you decided. If a writer challenges you, you want to be holding reasons, not a number from a vendor you cannot explain.

Four things you should not do

Do not reject on a single score. No detector, including ours, produces evidence of authorship. Treat a flag as a question, not a finding.
Do not demand a "0% AI" result as a condition of payment. That is asking a writer to be responsible for your tool's error rate.
Do not assume plain, careful prose is machine prose. The most-studied failure mode in this field is that detectors penalise simple, clear writing, and non-native English writing in particular. If a submission reads cleanly and simply, the base rate says it is probably a person.
Do not use it to make a decision with legal or similarly significant effects without human review. This is not just good practice; it is in our licence terms.

If you want this inside your own workflow

The free checker covers occasional use. If you are screening at volume, or you want detection inside your own editorial tool or platform, with your text never leaving your infrastructure, the offline pack runs on your machine: standard library, Python 3.8+, no API key, no word limits, no uploads.

Tier	Price	What it covers
Personal	$19 one-time	One person, unlimited runs, offline
Commercial	$99 one-time	Embed it in one of your own products or services, up to 5M words/year
OEM	from $999/year	Multiple products, white-labelling, no volume cap

See the product & pricing

Or email support@prosechecker.com if you want to talk through a screening workflow first.