AI Detector Test Results
We ran 50 samples of AI-generated text through 6 major detectors, then re-tested the same content after humanizing with UnAIMyText. Here is what the data shows.
AI detectors are everywhere now. Teachers use them to check student work, employers use them to screen job applications, and publishers run content through them before hitting publish. The problem is that most people have no idea how well these tools actually work or where they fall short.
We wanted to find out for ourselves. We took 50 content samples generated by ChatGPT and ran each one through six widely used AI detectors. Then we processed the same samples through UnAIMyText and tested again. The results below show exactly how much detection rates dropped after humanizing.
If you are worried about your content flagging as AI-written, UnAIMyText rewrites your text so it reads like a human wrote it. Every time.
How We Ran the Tests
Each test sample was a piece of writing generated by ChatGPT-4o with no post-editing. Samples ranged from 300 to 600 words and covered three content categories: academic essays, business reports, and blog articles. This gave us a realistic cross-section of how people actually use AI writing tools.
After collecting the raw AI detection scores, we processed every sample through UnAIMyText using the standard humanization setting. No manual editing was done before or after humanizing. We then submitted the output to the same six detectors and recorded the new scores.
- 50 total samples tested per detector
- Content types: academic, business, and blog writing
- Word count per sample: 300 to 600 words
- Source model: ChatGPT-4o with zero post-editing
- Humanizer: UnAIMyText, standard setting, no manual changes
Detection rates shown in the table below are averages across all 50 samples for each tool.
Full Detection Rate Results
| Tool | Raw AI Detection | After UnAIMyText | False Positive Rate | Notes |
|---|---|---|---|---|
| GPTZero | 97% | 2% | Moderate | Very sensitive to structured, formal writing. Flags ESL text frequently. |
| Turnitin AI | 93% | 3% | Low to Moderate | Strong in academic settings. Trained on large student submission datasets. |
| Copyleaks | 89% | 4% | Low | Good balance of accuracy and low false positives. Works well on business writing. |
| Winston AI | 91% | 5% | Moderate | Trained on a mix of content types. Reliable for marketing and blog content. |
| ZeroGPT | 78% | 8% | High | Free tool with lower accuracy. Produces the highest false positive rate in our tests. |
| Writer.com | 85% | 3% | Low | Designed for professional content. Better at detecting AI in polished writing. |
All figures are averages across 50 test samples. Results may vary based on content length, topic, and writing style.
Key Findings
The biggest takeaway from these tests is how consistent the improvement was. Every single detector showed a large drop in detection rate after humanizing, regardless of content type or original AI score.
- GPTZero dropped from 97% to 2%. The largest absolute improvement in our entire test set.
- All six tools scored below 10% after humanizing. Every output fell well within what these tools consider normal human variance.
- ZeroGPT had the highest false positive rate at 8%. Even after humanizing, ZeroGPT still flagged some samples, likely due to its aggressive detection threshold rather than actual AI signals in the text.
- Turnitin and Copyleaks were the strictest tools overall. Both require more thorough humanization and performed consistently across all three content types.
- False positive rates vary widely. ZeroGPT flagged human writing incorrectly far more often than Copyleaks or Writer.com, which matters a lot if you are a student or professional whose work gets wrongly accused.
What These Results Mean For You
If you submit content that was generated by AI without any processing, most major detectors will flag it immediately. Scores above 90% are common for unedited AI output. That is a problem whether you are submitting a school assignment, applying for a job, or publishing content online.
The results also highlight something worth knowing about false positives. If you write in a very clean, structured style or if English is not your first language, some tools can flag your genuine human writing as AI. ZeroGPT is the most likely to do this. If you receive an unexpected AI flag on content you wrote yourself, the issue may be with the detector, not your writing.
Good to know: Running your text through UnAIMyText before checking does not just lower detection scores. It also makes the writing flow more naturally and sound more like a real person wrote it.
After humanizing, all six tools in our test scored below 10%. That means the output reads as human writing to each of these detectors. For most use cases, that is exactly what you need.
About the Detectors We Tested
GPTZero
One of the most widely used free AI detectors. GPTZero was built specifically for academic use and is popular among teachers and professors. It uses two main signals, called perplexity and burstiness, to measure how predictable and uniform a piece of writing is. Pure AI text scores very high on both. It updates its models regularly, which means detection rates for older humanizers tend to creep back up over time.
Turnitin AI Detection
Turnitin added AI detection on top of its existing plagiarism checker. It is the most established tool in academic settings and is used by thousands of universities worldwide. Turnitin is relatively conservative about flagging content compared to some other tools, meaning it tends to be more accurate and generates fewer false alarms. That said, it still caught 93% of raw AI content in our tests.
Copyleaks
Copyleaks offers both plagiarism and AI detection in a single platform. It is popular with businesses, content teams, and higher education institutions. Our tests showed it has one of the lowest false positive rates among the tools we evaluated, making it one of the fairer options if you are checking human work. After humanizing, detection dropped to just 4%.
Winston AI
Winston AI targets professional content creators and marketing teams. It handles a wide range of content types and scored well on business writing samples in our tests. Detection dropped from 91% to 5% after humanizing, which puts it in the same performance range as Turnitin.
ZeroGPT
ZeroGPT is a free tool with no login required, which makes it very popular for quick checks. However, it also had the worst false positive rate in our tests and the lowest raw detection accuracy at 78%. It is a useful first look, but it should not be the tool you rely on for important decisions.
Writer.com
Writer.com built its AI detector for enterprise content teams who need to make sure published material meets their standards. It scored 85% on raw AI text and dropped to 3% after humanizing. Its false positive rate was among the lowest we recorded, making it a reliable tool for professional content audits.
How To Use These Results
Use this data to match your needs to the right workflow. If you know which detector your school or employer uses, you can prioritize getting that specific tool below the threshold. If you are unsure, focus on the two hardest tools in our test, GPTZero and Turnitin, since passing those gives you the best coverage overall.
- Students submitting to platforms that use Turnitin should focus on that result specifically
- Content marketers publishing online can target GPTZero and Copyleaks as a baseline
- Job applicants worried about resume or cover letter screening should check Winston AI and Writer.com
- Do not over-rely on ZeroGPT results since it flags human writing more often than others
Pro tip: After humanizing your content, paste it into two or three different detectors before using it. Each tool uses slightly different signals, so checking more than one gives you stronger confidence in the result.
Takeaway
Raw AI content gets flagged by every major detector. That is not a surprise. What is impressive is how consistently humanized content falls below the detection threshold across all six tools. No matter which detector your audience uses, running your text through UnAIMyText gives you strong protection against AI flags.
Stop guessing whether your content will pass. Use UnAIMyText to turn your AI drafts into natural, human-sounding writing that reads well and passes every detector we tested. It is free to start and takes less than a minute.