Honest about what it can and can't do.
A score is a reading, not a ruling. We show the bands plainly, and we tell you where we're wrong — because false accusations are the real harm.
Bands, not verdicts: the middle of this scale is genuinely uncertain territory, and we say so. No score here is proof about a person — it's a signal about the text.
The honest failure modes.
Formal & non-native writing
Formal, academic, or non-native English writing uses fewer contractions and more even sentences, so it can read high. We ease off those signals by register, but the risk never fully disappears.
Edited or newer AI
A light paraphrase or edit removes most surface tells, and the newest human-tuned models can slip past almost entirely. A low score means the text reads human — not that it was.
Short text
Under ~300 words most signals abstain or wobble. Give it a few solid paragraphs of writing for the most reliable read.
Accuracy questions, answered straight.
How accurate are AI detectors, really?
Nobody can honestly give you one number. Published accuracy claims come from each vendor's own test set, and independent studies show every detector — including the biggest names — missing rewritten AI and flagging some human writing. Accuracy varies hugely by text type, length and how much a person edited. That's why this page documents failure modes instead of advertising a percentage.
Why doesn't ScanForAI publish an accuracy percentage?
Because the number would be true only for our test set, and you'd read it as true for your text. A '99% accurate' claim measured on unedited model output says nothing about a lightly-edited essay from a non-native writer — the case that actually matters. We show the evidence per scan instead, so each result argues for itself.
What causes false positives in AI detection?
Formal register, very even sentence lengths, careful grammar, few contractions, template-like structure — the habits of disciplined writing overlap the habits of machine writing. Non-native English writers get hit hardest, on every detector. It's the #1 reason a score should start a conversation, never end one.
What does a mid-range score mean?
Genuine uncertainty — and we label it that way instead of rounding to a verdict. Mid-range text usually mixes signals: some machine-leaning patterns, plenty of human ones. Look at which lines are marked and judge those, not the number.
Which AI detector is the most accurate?
There's no honest king of that hill — rankings flip depending on the test corpus. The strongest signal available to you is agreement: when two detectors with different methods mark the same passages, that's worth far more than either one's score alone. We're built to be a good second opinion: free, and we show our reasons.