Run one essay through five free AI checkers and you'll get five different verdicts — sometimes spanning the whole range from "fully human" to "mostly AI." Students do exactly this the night before a deadline, watch the numbers swing wildly, and end up more anxious than when they started. The disagreement isn't a bug in one tool. It's structural, and understanding why tells you how much weight any of these numbers deserves.
There is no shared definition of "AI-written"
An AI detection score is the output of one specific model, trained on one specific dataset, with one specific decision threshold. Change any of the three and the number changes. One tool was trained mostly on ChatGPT essays, another on a broader mix; one calls a passage AI at 60% internal confidence, another at 90%; one scores sentence by sentence, another in big chunks. They aren't measuring the same thing and then disagreeing — they're measuring different things that happen to share a name.
Detection is also threshold-sensitive by nature. These models produce probabilities, and a passage sitting near the line can flip from 8% to 35% with a single paragraph edited — or with the same text pasted into a different tool. That's why your score can swing between checkers, and even between runs.
Free tools have incentives, and they aren't yours
Most free checkers are the top of a sales funnel: the free scan exists to convert you to a paid "humanizer" or subscription. Think about which error serves that funnel. A false "all clear" earns nothing; a scary number converts. That doesn't mean every free tool inflates deliberately — but when a tool profits from your alarm, its threshold choices deserve skepticism. Add the quieter costs (many free tools retain the text you paste; some train on it), and "free" is doing a lot of work in that sentence.
Turnitin is a different model with a different job
Turnitin's detector is its own classifier, tuned conservatively because its output lands in front of instructors making misconduct decisions — it even masks scores under ~20% with an asterisk rather than state an uncertain number. It counts only qualifying prose, excludes lists and references, and needs a minimum length to score at all. A free checker that happily scores your 150-word paragraph is running a completely different pipeline from the one your university runs.
So the correlation between a free checker's number and Turnitin's number is weak in both directions: a 0% on a free tool doesn't clear you, and a 70% doesn't condemn you. The only number that predicts a Turnitin score is a Turnitin score.
How to use checkers sanely
- Treat free checkers as a rough smoke alarm, not a forecast. Multiple tools all screaming is a signal worth acting on; one tool disagreeing with another is noise.
- Never paste sensitive or unpublished work into a tool that doesn't state what it does with your text.
- If the stakes are real — a flagged-adjacent writing style, a final-year paper, a resubmission after an accusation — test against the engine your university actually uses. A real Turnitin check returns both reports in about 10 minutes, in no-repository mode so nothing is stored.
The five conflicting numbers were never the problem. The problem is that only one number counts, and it's the one you couldn't see — until you could.