How to spot text that's been through a humanizer.

What cheap humanizers leave behind

One-pass paraphrase tools dodge blacklist words by substitution — and leave their own artifacts: slightly-off synonyms ('utilize' becomes 'make use of', not 'use'), broken idioms ('at the end of the day' → 'when the day concludes'), sentence rhythm that's varied but mechanically varied (short-long-short-long like a metronome), and — the reliable one — uniform specificity. Human writing has lumpy detail: dense where the author knows things, thin where they don't. Laundered text is evenly, blandly semi-specific everywhere.

The three-question test (better than any detector)

1. Where's the lump? Find the paragraph with the most specific detail. If no paragraph is noticeably denser than the rest, be suspicious. 2. Can the author extend it? Ask one follow-up about the most specific claim. Humans elaborate easily on things they wrote; launderers can't. 3. Does the voice match the person? Compare against anything they wrote before — the same email thread will do. Voice discontinuity is stronger evidence than any detector score.

What not to do

Don't treat detector scores as proof — false positives hit non-native speakers and formal writers hardest, and false accusations do more damage than missed AI use. Detector output is a reason to look closer, never a verdict. The honest standard for graded or bylined work is process evidence: drafts, history, and the author's ability to discuss their own text.

The artifact zoo: laundering signatures by tool type

Thesaurus-substitution tools leave semantic near-misses: "significant" becomes "notable" where the sentence needed "large"; idioms get literalized ("hit the ground running" → "start with immediate momentum"). Sentence-restructuring tools produce grammatically flawless but rhythmically alien text — clauses reordered in ways no drafting human would choose, referents ("this", "it") drifting slightly off their targets. Prompt-based rewrites (the competent method) leave the fewest artifacts but keep two: uniform paragraph lengths and a suspicious absence of the writer's normal tics. If someone's emails always contain a dangling "anyway —" and their report contains zero, the polish itself is the anomaly.

A fair process for classrooms and newsrooms

Written policies beat detector roulette. What works in practice: declare the rules per assignment (allowed with disclosure / allowed for editing only / not allowed) instead of a blanket ban nobody believes; collect process artifacts routinely — outline, one draft, version history — so evidence exists before any dispute; make the conversation the instrument — ten minutes of discussing their own text settles authorship more reliably than any score; and never sanction on a detector score alone — vendors themselves publish false-positive rates that make solo reliance indefensible. The goal is a process a wrongly-accused student would call fair, because eventually one will be.

What a false accusation costs

The asymmetry deserves naming: a missed AI text costs a grade's integrity; a false accusation costs a student's record, a writer's reputation, sometimes a job — and documented cases exist of both non-native speakers and perfectly ordinary formal writers being flagged repeatedly by detectors. Any institution using detection owes its people the arithmetic: at scale, even a 2% false-positive rate accuses dozens of innocent writers per thousand texts. That's why every serious workflow treats scores as triage and humans as judgment — the standard this whole site argues for, from both sides of the desk.

The 60-second desk check (printable version)

When a text lands on your desk and something feels off, run these in order — each takes seconds. 1. Lump test: is any paragraph noticeably denser with specifics than the rest, or is detail spread evenly and thinly? 2. Blacklist scan: paste into the checker — over ~12 tells per 1,000 words is machine register, under 5 proves nothing either way. 3. Voice diff: put one paragraph next to any earlier writing by the same person; vocabulary level and rhythm should rhyme. 4. Idiom integrity: hunt for literalized idioms and almost-right synonyms — laundering artifacts, not writing mistakes. 5. Extension question: ask the author to expand the most specific claim, live or by reply. Confident elaboration in their own register closes the question; deflection to generalities opens it wider. No single check convicts; three pointing the same way justify the conversation, and the conversation — not the checklist — is the verdict.

When the text is fine and the process still matters

Increasingly the realistic case isn't fraud but undisclosed assistance: the report is accurate, the author understands it, an AI drafted half of it, and your organization has no rule either way. Treat that as a policy gap, not an offense. Teams that handle this well publish a simple norm — assistance allowed, author owns every claim, disclosure required where the reader would reasonably want it (bylined journalism, graded work, expert testimony) and not where they wouldn't (status updates, internal summaries). The uncomfortable alternative is selective enforcement based on whose style happens to trip a detector, which is how organizations end up defending exactly the false accusations this page warns about.