I Tested 5 AI Detectors on 50 Samples. Here's Why They're Not Reliable in 2026

A PhD student in my network wrote her thesis with zero AI.

Turnitin flagged it: 67% AI-generated.

She spent 2 weeks dumbing down her writing to beat the detector. The paper got worse.

So I tested 5 detectors myself on 50 samples: Turnitin, GPTZero, Copyleaks, ZeroGPT, and Originality.ai.

The Test

• 10 Human papers
• 10 AI-generated
• 10 AI + human edit
• 10 AI + humanized
• 10 ESL human papers

The Results
No tool broke 84.4% accuracy. Worst was 69.4%.

1 in 4 verdicts was wrong.

• Originality.ai: 84.4% best overall. Still missed 50% of humanized text.
• Turnitin: 72%. Flagged 40% of ESL papers. One hit 52%.
• GPTZero: Caught all AI but had 12% false positive rate.
• ZeroGPT: Gave different results 30% of the time on the same text.

Why They Fail

1. Hybrid text: AI + human edit drops accuracy to 54-71%
2. ESL Bias: 28-61% false positives on non-native writers
3. Academic prose: Formal writing looks "AI" to detectors Dev Takeaway Don't use detector scores as proof. If you're building with AI, keep git history, prompts, and drafts. Demand human review.

Full data + charts: [https://worldcutruygdski.blogspot.com/2026/07/ai-detector-accuracy-2026-test-results.html]

What has your experience been with false positives?

Comments

[Luis Cruz]

I found the results of your test on AI detectors to be quite eye-opening, especially the high false positive rate for ESL papers. I've had similar experiences with Turnitin in the past, where it flagged some of my colleagues' papers as potentially AI-generated simply because of their non-native writing style. I'm wondering if the developers of these detectors have considered incorporating more diverse training data to mitigate this ESL bias. Do you think it's possible to develop a more accurate detector that can account for the nuances of human writing, or are we stuck with these limitations for the foreseeable future?

[Rehab]

Thanks Luis! You hit the nail on the head. The ESL bias is the biggest flaw right now.
I don't think current detectors are trained on enough diverse, non-native writing samples.
Until that changes, we’re stuck with high false positives.

[Rehab]

Thanks so much Luis for this thoughtful comment! 🙏

You’re absolutely right about the ESL bias. It’s a huge problem that a lot of these detectors flag non-native writing patterns as “AI” just because they’re different from the training data.

I agree that detection shouldn’t rely only on the final text. Looking at context like drafts, version history, and the writing process would make the results way more fair and accurate.

Appreciate you sharing the Teams community link too. I’d love to join the discussion and learn from others working on responsible AI tooling.