Why different detectors give different scores

Quick answer: Detectors use different models and different thresholds. They often disagree on the same text. No single detector is the "right" answer.

A real detector result for partially-humanized text: the gauge shows 99% AI overall, the per-detector grid has all 8 detectors marked, but the segment legend at the bottom splits the same text into 51% Human-written, 19% Mixed, 30% AI-generated — different layers of the analysis disagreeing on the same input.

Why they disagree

AI detectors are not interchangeable. They differ in three ways:

1. Training data

Each detector trained on a different mix of human and AI text. A model trained on news articles will read fiction differently than a model trained on essays.

2. Threshold

"AI" vs. "human" is a yes/no decision built on top of a probability. GPTZero might call 0.55 "AI." Originality.ai might use 0.40. Same text, different verdicts.

3. Update cadence

LLMs evolve. Detectors update at different speeds. A detector that is three months out of date scores GPT-4 output differently than one that updated last week.

Which detector should I trust?

It depends on what you need the text for:

Academic submission? Trust the detector your institution actually uses (often Turnitin).
Client deliverable? Trust the one your client cares about.
General-purpose check? Look at the overall picture across all 8. If 6+ say human, you are in good shape.

Note: No detector is authoritative. Even Turnitin issues guidance saying its AI detection should be one signal among several, not a sole judgment.

What does "GPTinf's detector" mean, then?

GPTinf does not have a single proprietary detector trying to compete with the others. It runs your text through 8 third-party detectors and shows you each result. That is the point — you see disagreement when it happens, instead of relying on one black box.

Why your text might score differently on the same detector twice

Some detectors are non-deterministic — they sample. If you run the same text twice and get 0.60 then 0.58, that is not a bug.

FAQ

Is one detector better than the others? Not in general. Some are stricter (lower threshold). Some update more often. "Better" depends on what your reader will use.

Can I argue with a teacher who got a "high AI" score? That is between you and your teacher. The honest message is: detectors are imperfect, and humans should treat them as one signal — not the final word. See GPTinf's stance on AI ethics.

Does running text through GPTinf affect future scans? No. GPTinf does not share your text with the detectors as training data.