Experiments
AI humanizer experiments
Focused, reproducible tests that go beyond the core benchmark — edge cases, specific detectors, prompt styles and long-tail questions we wanted to answer with data.
Source-model experiment
A comparison of seven AI humanizers across ChatGPT, Claude and Gemini source text shows how detector performance and factual preservation vary by source model.
Read the experiment ExperimentFactual-preservation test
A controlled test of seven AI humanizers measures whether they preserve names, dates, numbers, monetary values and negations.
Read the experiment ExperimentRepeated-humanization test
We humanized the same texts up to five times to measure whether additional passes improve detector results or simply damage text quality.
Read the experiment ExperimentText-length test
We tested seven AI humanizers across six input lengths to see where detector performance peaks and which tools are most sensitive to very short or long text.
Read the experiment