A literary agent's guide to AI-suspicious submissions.

Agents and acquiring editors did not sign up to be forensic analysts. But the slush pile changed, and "is any of this manuscript a machine?" is now a question that affects advances, contracts, and reputations. Here is a workflow that takes the question seriously without turning your inbox into a tribunal.

Start with the reading, not the tool

Before any software, trust the experience that got you this job. AI manuscripts tend to fail in a characteristic way: they are competent at the sentence and hollow at the scene. Plot beats arrive on schedule; nothing surprises. Characters have backstory but no interiority. The prose never embarrasses itself and never risks anything. If a submission makes you think "this is clean but I feel nothing," that instinct is data. The tool is for testing the hypothesis your reading already formed.

A three-tier triage

You cannot deep-analyze 2,000 submissions. You can triage them.

• Tier 1 — pass on craft. Most of the pile is declined for ordinary reasons. AI status is irrelevant to a manuscript you're rejecting anyway. Don't spend the cycles.
• Tier 2 — interested, run a check. For the small set you'd actually consider, a structural audit is cheap insurance. Treat an elevated result as "read the flagged chapters again, closely," not as a decision.
• Tier 3 — offer stage. Before anything contractual, a fuller pass plus a direct, normal conversation about process. By this point you're talking to the writer anyway.

Reading a result honestly

The most important habit is internalizing the false-positive cases. A spare, disciplined literary voice can score like a machine because the very features detectors punish — even cadence, low lexical surprise, sanitized dialogue — are also the features of mastered minimalism. A debut that reads like early Didion is exactly the submission a careless screen will libel. If you would be embarrassed to defend an accusation to the writer's face using only a number, you do not yet have enough to act on.

Look instead for convergence with the reading. A high score on a manuscript that also felt hollow, with the flags landing on the chapters that felt most generic, is a coherent picture. A high score on a manuscript that moved you is far more likely to be the tool punishing craft than the writer deceiving you.

How to raise it with a writer

If a project clears your interest bar but you have a real concern, the conversation is survivable if you frame it as process, not accusation. Useful, non-hostile prompts:

"Tell me about how you drafted this — what your revision process looked like, what changed between drafts, which parts were hardest."

A writer who lived inside the book can talk about it endlessly and specifically: the scene they cut, the character who started as someone else, the ending they rewrote five times. Vague, process-free answers are far more telling than any score. You are not running an interrogation; you are having the conversation you'd have anyway, listening for whether a human mind clearly inhabited this work.

What a policy should — and shouldn't — say

Agencies are writing AI clauses into submission guidelines. The ones that will age well share three traits: they distinguish assistive use (a model as a thesaurus or a brainstorming partner) from generative authorship; they ask for disclosure rather than threatening detection; and they never promise that a piece of software determines authorship. A guideline that says "we screen with AI detection and act on the results" is a lawsuit waiting for a plaintiff. A guideline that says "we expect substantially human authorship and ask you to disclose AI use" is enforceable and fair.

The one rule

Never decline a writer, and never make an allegation, on a number alone. Use structural analysis the way you use a second reader: to direct attention, sharpen questions, and check an instinct — then make the judgment yourself, the way you always have. The tools are good enough to be useful and not nearly good enough to be a verdict. Acting as if they were is the only way this technology actually hurts the writers you exist to represent.