AI Detection for Publishers: Screening Submissions Without Making Enemies
Published April 15, 2026 · 8 min read · For editors, editorial directors, and submissions managers
AI-generated submissions are now routine across every category of publishing: news, magazines, books, online media, and academic journals. The challenge is screening them accurately enough to protect editorial standards without creating a detection-first culture that alienates legitimate contributors.
Check a submission now
Paste any text into Airno. 8 detectors, per-signal breakdown, results in seconds. No account required.
The problem publishers are actually facing
The economics of AI content creation have completely changed the submission math. A freelancer who previously could write three articles per week can now use AI assistance to produce twenty. That volume is showing up in slush piles, query inboxes, and open submission windows at a scale most editorial teams were not designed to handle.
The problem is not AI assistance in general. Many writers use AI tools for research organization, outlining, or grammar polish. The problem is unattributed wholesale AI generation: content where the writer did not do the reporting, did not form the arguments, and did not write the prose, but is presenting it as their own work.
The distinction matters for detection strategy. Lightly AI-assisted human writing will produce low-to-medium detection scores. Wholesale AI-generated content, even after light editing, typically produces high scores across multiple detection signals. A well-calibrated ensemble detector can distinguish between these cases more reliably than a single-model tool.
AI risk by submission type
Not all submission categories carry equal AI risk. Understanding which types are most susceptible helps you allocate screening resources.
| Submission type | AI risk | Why |
|---|---|---|
| Short-form news articles | Very High | High-volume, low-pay assignments incentivize AI use |
| Listicles and roundups | Very High | Formulaic structure is easy for AI to replicate |
| Product reviews | High | Especially review aggregations without first-hand testing |
| How-to guides and tutorials | High | AI produces competent but unverified technical content |
| Opinion and essays | Medium | Harder to fake convincingly; more voice-dependent |
| Reported features | Low | Requires sources, quotes, reporting; AI cannot fake these |
| Fiction and creative writing | Medium | AI fiction exists; voice and originality remain key signals |
How AI detection works for editorial screening
A reliable AI detector analyzes multiple independent signals and combines them. Airno runs 8 parallel detectors: a statistical model, a fine-tuned DeBERTa-v3 neural classifier trained on 38,400 samples, a 314-pattern linguistic corpus, a frequency distribution analyzer, a coherence and burstiness model, a CNN artifact detector, a metadata inspector, and an adversarial pattern checker.
For editorial screening, the per-signal breakdown is the critical feature. When you see "85% AI confidence" alongside a breakdown showing 6 of 8 detectors converging on that score, you have defensible evidence. When you see one elevated signal and seven neutral ones, that is a much weaker case.
The other signal to check is what the content lacks rather than what it contains. AI-generated articles rarely have specific named sources, original quotes, verifiable first-person observations, or accurate hyperlocal detail. A listicle or how-to guide may not require these. A reported feature absolutely does.
A practical screening workflow
Pre-submission
Set expectations in contributor guidelines. State that submissions may be screened for AI content. This discourages casual AI use without creating adversarial dynamics.
Receipt triage
Run a quick scan on all submitted text. Flag anything above 65% for a second-pass review. Do not reject automatically based on the score alone.
Second-pass review
For flagged submissions, check per-signal breakdown. Multiple signals converging on a high score is stronger evidence than a single elevated signal. Look at which detectors fired.
Context check
Does the piece contain specific reporting: named sources, original quotes, first-person observations, verifiable details? AI cannot generate these. If they are absent in a reported piece, that is a red flag independent of the detection score.
Author conversation
For borderline or high-risk submissions, ask the author about their process. A writer who used AI heavily cannot explain their own reporting, sources, or argument development in detail.
Decision and documentation
Document the detection score, which signals fired, any contextual flags, and the outcome of any author conversation. This creates a defensible record if a decision is disputed.
Where detection fails and how to compensate
Heavily edited AI text
A writer who runs AI output through a heavy editing pass can reduce detection scores substantially. The text may still lack original reporting, specific sources, or genuine point of view. Detection tools give you the statistical layer; editorial judgment gives you the substance layer. Both are necessary.
Non-native English writers
Formal, structured writing from non-native English speakers can produce elevated detection scores because of writing patterns that overlap statistically with AI output. If a submission comes from a non-native English writer and scores in the 35-65% range, treat the score as ambiguous. Look at the actual content for reporting evidence, not just the number.
AI-assisted but human-written work
Many legitimate writers use AI for research organization, grammar checking, or idea generation without generating the prose itself. These submissions will typically score below 40%. The question is whether the prose, arguments, and reported content are the author's. Detection scores in the 30-40% range should not trigger rejection.
Domain-specific technical content
Technical documentation, medical writing, and legal content have formal structures that can overlap with AI patterns. Run baseline samples from established authors in the domain before applying detection to unsolicited submissions.
Writing an AI submission policy
Clear contributor guidelines are more effective than detection alone. When writers know what you expect, most will comply. The ambiguity around AI use in publishing has partly driven the problem: many publications did not define their position until AI submissions became unavoidable.
Example policy language
"We accept work generated with AI assistance for research, outlining, or editing support, provided the reporting, arguments, and prose are the contributor's own original work. We do not accept submissions where AI generated the primary prose, regardless of how it was subsequently edited. Submissions may be screened for AI content. Work found to rely primarily on AI-generated prose will be declined without further consideration, and the contributor may be flagged for future submissions."
The key elements: define what AI assistance is acceptable (research, editing), define what is not acceptable (AI-generated prose), state that screening may occur, and specify the consequence. This reduces ambiguity and gives you a documented standard to apply consistently.
Questions from editorial teams
Can we reject a submission based solely on an AI detection score?
Technically yes, but it is not advisable as a standalone policy. A high score is strong evidence, not proof. For low-stakes rejections (unpaid open submissions, slush pile queries), a high score plus a lack of reported content is reasonable grounds. For paid or commissioned work, or any context where the relationship with the contributor matters, follow the score with a conversation.
What score threshold should we use for flagging?
65% is a reasonable flag threshold for second-pass review. 80% and above warrants a direct conversation or rejection depending on your policy. Scores below 40% should generally not trigger concern. Scores between 40-65% warrant looking at the substance of the piece rather than the number.
How do we handle AI use that was disclosed?
Disclosure does not resolve the editorial question; it changes it. The question becomes whether the disclosed AI assistance crosses the line into AI-generated prose. A writer who says 'I used ChatGPT to outline and then wrote the piece' presents differently than one who says 'I used AI to write a first draft I then edited.' Your policy language should address this distinction.
Do we need a dedicated tool or does Airno work for this?
Airno is free, requires no account, handles text up to 100,000 characters, and returns a per-signal breakdown. For most editorial screening workflows, especially at publications that do not have a dedicated editorial technology budget, it is a practical starting point. For high-volume operations processing hundreds of submissions daily, an API-based automated pipeline is more efficient.
Know if it's real. Know if it's AI.
8 independent detectors. Per-signal breakdown. No character limit. Free, no account required.
Check a submission now