Stealthwriter AI Detector Test: Does “Humanized” Writing Actually Look Human?

AI humanizers promise something students instantly understand: take a robotic draft, rewrite it, and make it sound natural. But there is a more important question hiding underneath that promise. If a tool can raise a detector score, does that mean the writing is actually better, safer, or more believable?

To answer that, 100 Stealthwriter AI rewrites were tested across five AI detectors: Copyleaks, GPTZero.me, Originality AI, Quillbot AI, and ZeroGPT. The detector results were converted into human scores, so a higher number means the text looked more human to the detector. For example, an 80% human score means the detector leaned toward “human-written,” while a 20% human score means it leaned toward “AI-written.”

The results were not simple. Stealthwriter did well against some detectors, struggled badly against others, and sometimes produced rewrites that looked human to software but still sounded awkward to a real reader.

Also Read: Can Stealthwriter AI Beat QuillBot's AI Detector? We Tested 100 Rewrites

The Scoreboard: Which Detectors Were Easier to Beat?

The first chart shows the average human score given by each detector. This is the simplest way to see where Stealthwriter looked strongest.

Also Read: [STUDY] StealthWriter vs Sapling AI: Can 100 Humanized Rewrites Slip Through?

Average human score by detector. Higher scores mean the rewritten text looked more human to that detector.

Quillbot AI was the most forgiving detector in this dataset, giving the rewrites an average score of 81.4% human. Originality AI was also relatively high at 74.3%, followed by Copyleaks at 67.7%. ZeroGPT was more cautious at 58.5%, while GPTZero.me was the hardest for Stealthwriter to pass, averaging just 47.2%.

For students, the lesson is straightforward: there is no single “AI detector result.” The same rewritten paragraph can look human to one detector and suspicious to another. That makes detector-based confidence much less stable than it may seem.

Also Read: [STUDY] Can Stealthwriter Really Bypass Copyleaks? What 100 Samples Show

Passing Scores: The 50% and 80% Question

Averages are useful, but students usually care about a more practical question: “How often did the rewrite pass?” To keep this simple, this analysis uses two lines. A 50%+ human score means the detector leaned human. An 80%+ human score means the detector gave a much stronger human-looking result.

Pass rates at two cutoffs: 50%+ human and 80%+ human.

At the 80%+ human level, Stealthwriter passed 79% of samples on Quillbot AI, 66% on Originality AI, and 61% on Copyleaks. But that success dropped to 40% on GPTZero.me and only 26% on ZeroGPT.

This matters because a humanizer is not just trying to win against the easiest detector. In real academic or publishing situations, the text may be checked with different tools, updated models, or more than one detector. A rewrite that passes one scanner is not automatically protected everywhere else.

Also Read: [STUDY] Can Stealthwriter Outsmart GPTZero? A 100-Sample Test

The Bigger Problem: Detectors Disagreed A Lot

The most surprising part of the test was not that some detectors were stricter. It was how strongly they disagreed. In many samples, one detector gave a near-perfect human score while another gave almost zero.

For each rewrite, this counts how many of the five detectors scored it 80%+ human.

Only 4 samples cleared the 80%+ human mark on all five detectors. Another 4 samples failed to reach 80% on any detector. Most samples landed in the messy middle, where two, three, or four detectors treated the rewrite as human-looking while the others did not.

The score spread chart below shows the same issue from another angle. A “spread” simply means how far apart the scores were. When the spread is wide, the detectors are not agreeing.

Also Read: Stealthwriter vs ZeroGPT: I Tested 100 Rewrites, and the Results Were Complicated!

Score spread across the 100 samples. Wider variation means less predictable detector behavior.

Across the dataset, the average gap between the highest and lowest detector score for the same rewrite was 80.6 percentage points. The median gap was even larger at 96.5 points. In plain English: detector results were often wildly different for the exact same rewritten text.

Human Score Does Not Mean Human Quality

Detector scores are only one part of the story. The rewritten text also needs to make sense. When the CSVs were reviewed for rewrite quality, several problems appeared that a detector score alone would not catch.

Quality red flags found while comparing original text and rewritten text.

These issues are important because they show the danger of judging a rewrite only by detector scores. A student might get a high human score and still submit a paragraph with a factual error, strange phrasing, or broken formatting. In an essay, those problems can be more obvious to a teacher than an AI detector result.

What Students Should Take Away

• Stealthwriter can improve detector scores, but not reliably across every detector.
• Quillbot AI, Originality AI, and Copyleaks were easier for these rewrites to pass than GPTZero.me and ZeroGPT.
• A single high score should not be treated as proof that the writing is safe, accurate, or genuinely human.
• Manual editing is still necessary. Check facts, names, headings, lists, and sentence meaning before using any rewritten text.