[self-promotion] 24 Recorded AI-detector Scores For Six Unchanged AI-written Texts (JSON)

I run AI Undetectable, an AI humanizer. I’m sharing the recorded detector scores from a small repeatability check, not a product ranking.

Six unchanged AI-written texts were checked on July 3 and July 11, 2026, using ZeroGPT and aidetector.com. That’s 24 readings: six texts, two detectors, two dates.

The JSON’s historicalControls section has a topic ID and four scores per row. ZeroGPT’s scores stayed unchanged for all six; aidetector.com’s changed in both directions. One went from 28% to 73%, another from 19% to 3%.

Direct JSON, no signup: https://aiundetectable.com/research/ai-humanizer-benchmark-2026.json

Methods and limitations: https://aiundetectable.com/ai-humanizer-benchmark

This is only six texts, with no human-written control group. The file contains scores, not the original texts. It can’t establish false-positive rates or explain why scores changed. The separate holdout results in the same file are aggregate summaries, not 32 individual records.

AI helped prepare this post.