Flagged by an AI detector? Read this before you panic

Start with the only number that matters in this situation: every AI detector has a false-positive rate, and it is never zero. Ours is 0.4 percent on genuine scientific papers at the default setting, and we consider that number the single most important thing we publish. Many tools do not publish theirs at all. If a detector's marketing says "99 percent accurate" without telling you how often it flags real human writing, you have learned nothing about your situation.

What a detector score actually is

A detector score is not a fact about authorship. It is a statistical statement: "text with these word choices, rhythms, and patterns appears more often in machine-generated training examples than human ones." That is evidence, in the way that a smoke detector is evidence of fire. Smoke detectors also go off when you make toast.

This matters because the consequences are asymmetric. A student or researcher wrongly accused of using AI faces real damage: academic integrity proceedings, a supervisor's doubt, a rejected submission. The tools producing these verdicts are screening instruments, and screening instruments are designed to be second-guessed.

Why genuinely human writing gets flagged

Several well-documented confounds push human text toward "machine" verdicts across the entire field, not just any one tool:

• Non-native English. Writers using English as a second language often use more regular, textbook-like constructions. Regularity is exactly what detectors key on. This is the most serious and best-documented bias in the field.
• Formulaic sections. Methods sections, abstracts, and literature reviews follow genre conventions so tightly that human and machine versions converge. "We conducted a randomised controlled trial with N participants" is how everyone writes it.
• Short passages. Under a few hundred words, every detector is guessing. Statistical fingerprints need material.
• Edited and polished text. Grammar checkers and heavy revision push prose toward the smooth median style that detectors associate with machines. Ironically, careful editing makes you look more artificial.
• Formatting artifacts. Markup, hard line breaks, and copy-paste damage change the statistics before a single word of your style is measured. We wrote about the LaTeX version of this problem separately.

What to do, step by step

1. Get the passage-level view, not just the number. A whole-document percentage tells you almost nothing actionable. Run the text through a tool that shows which passages drive the verdict (unslop shades every sentence). If the flagged material is your methods boilerplate and not your argumentation, that is worth knowing and worth saying.
2. Check the operating point. Any honest tool lets you see or set the sensitivity. At an aggressive threshold, everyone gets flagged sometimes. Ask what threshold was used and what its false-positive rate is. If the answer is "we don't know", that is your answer.
3. Assemble provenance, not protests. Version history is the strongest evidence a writer has: Overleaf and Google Docs histories, git commits, dated drafts, notes, and reference exports. A writing process leaves a paper trail no generator produces.
4. Ask for the specific tool and version. Detectors disagree with each other constantly, and the same tool changes verdicts between versions. A decision that cites "an AI detector said so" without naming the tool, version, and threshold is not a measurement, it is an anecdote.
5. Know the field's own position. Every serious detector vendor, including us, states that scores are not proof and should never be the sole basis for a decision about a person. If a process is treating a score as dispositive, it is misusing the tool by the tool's own documentation.

If you are the one doing the checking

Flip everything above around. Use passage-level evidence rather than a single number. Prefer tools that publish false-positive rates. Treat a flag as a reason for a conversation, not a conclusion. And weigh the asymmetry: a missed AI text costs little; a false accusation can cost someone a degree.

We built unslop around that asymmetry: free and unlimited so anyone accused can check their own work at the same operating point as the accuser, passage-level shading so the conversation can be about specific sentences, and a published false-positive rate so everyone knows how often the smoke detector goes off over toast.