Is ZeroGPT accurate? What our own tests showed
We ran real human writing, ChatGPT text and rewritten text through ZeroGPT. Here's what it got right, where it struggles, and how to read its score.
ZeroGPT is one of the most-used free AI detectors. Because it is free and needs no sign-up, many teachers, editors and clients use it as a quick check. We tested it while building TrulyWrite, so here is what we saw, with the limits of a small test spelled out.
What we tested
- Real human writing published online before ChatGPT existed: news pages, blogs and forum posts, 45–90 words each.
- Short and medium texts in typical ChatGPT style: emails, LinkedIn posts, product copy, cover letters and essay paragraphs.
- The same ChatGPT texts after rewriting by several tools and methods.
What we found
- Human writing: all six real human samples we checked scored 0% AI. On real, specific human text it did well in our test.
- Raw ChatGPT text: almost always flagged as 100% AI.
- Rewritten generic text: short, generic paragraphs (for example a few sentences about “the benefits of remote work”) stayed flagged at 80–100% even after heavy rewriting. ZeroGPT seems to react to how predictable the content is, not only to stock phrases.
- Text with real details: rewritten texts that kept specific names, numbers and events often scored low, even when a few formal phrases remained.
How to read a ZeroGPT score
Treat it as a signal, not a verdict. Independent reviews report much lower accuracy than the 98% ZeroGPT advertises, with false positives on human writing, especially formal and academic prose.
If a score surprises you, look at which sentences are highlighted. Generic, evenly paced sentences are the usual trigger. Adding a real example, a number or your own experience usually helps more than swapping synonyms.
TrulyWrite's AI Detector scores every sentence and explains the overall result, so you can see exactly what to fix.
The limits of our test
This was a small, informal test run in November 2026, and detectors change often. Run your own text through more than one detector before relying on any single score.