[STUDY] Can Phrasly AI Humanizer Bypass ZeroGPT?

As we all know Quillbot gets easily detected by ZeroGPT. However, is it the same for Phrasly, which actually claims to bypass AI detectors? The short answer is YES, it still gets flagged. The longer answer is the devil truly lies in the data. Keep reading to know more about it.

Why Phrasly might get detected by ZeroGPT?

The simple answer is that even though Phrasly is actually made to bypass AI detectors, it doesn’t seem to accomplish this task very reliably. We tested 100 paraphrases generated by Phrasly and fed them to ZeroGPT, which is an AI detector that assigns a score ranging from 0 to 100. A 0 generally means very human-like and 100 means very likely to be AI-generated.

Let’s quickly define these things for clarity:

• ZeroGPT score range: 0 (human-like) to 100 (AI-like)
• “Bypass” means the paraphrase is under a certain threshold like <50 or <70. If the score is under that threshold, you can say it bypassed the detector.

Dataset and Setup

We used a sample size of 100 paraphrases. For each one, we recorded:

• The original text (not that relevant here)
• The Phrasly-generated paraphrase
• The ZeroGPT score assigned to that paraphrase

We wanted to see if the giant claims of Phrasly “evading AI detectors” hold true for ZeroGPT’s detection algorithm.

Also Read: Can Phrasly AI Humanizer Bypass GPTZero?

Overall Score Summary (n=100 paraphrases)

threshold_score_lt	count_bypass	pct_bypass	count_flagged	pct_flagged
20	5	5.0	95	95.0
50	7	7.0	93	93.0
70	8	8.0	92	92.0
90	13	13.0	87	87.0

This is where it gets more interesting. We found that:

• Median score is a solid 100. That means at least half of the paraphrases have a perfect 100, which is basically screaming “AI!”
• Mean score is 91.96, so overall the content is flagged almost near the max.
• The 75th percentile is exactly 100. That means that 75% of the paraphrases scored 100 or below.
• 87% of paraphrases have a score ≥ 90.
• 7% of the samples are below 50, which is a common “probably human” threshold among many detection tools.
• About 5% of them scored 0, which shows that sometimes AI detectors can drastically drop the score or random short content might appear super “human.”

Score Distribution

If you put these 100 samples on a histogram focusing on their ZeroGPT scores (0 to 100), you’ll see a heavy concentration near 100 with a tiny tail that dips into those lower scores.

Empirical CDF (Cumulative Distribution Function)

We plotted the fraction of paraphrases below a given score. For instance, if you move your finger to the 50‐score on that CDF, you would spot that only about 7% of paraphrases are below that point (which means a 7% bypass rate if your threshold is 50).

Does Paraphrase Length Matter?

A lot of people think that maybe if you produce shorter paraphrases or slightly longer ones, you might trick an AI detector. So we tested the relationship between paraphrase length (word count) and the ZeroGPT score. The scatterplot basically shows that the paraphrases are pinned high (scores near 100) across all sorts of word counts.

Quantitative Findings on Length vs. Score

We even computed correlations:

• Linear correlation is about 0, meaning no direct relationship.
• Spearman correlation (which is rank-based) is a bit negative but too small to matter.

In simpler terms, short or long, Phrasly paraphrases are largely flagged as AI.

Bypass Rates by Threshold

If you pick different thresholds, you’ll see how many paraphrases pass as “human.”

• Score < 50 is about 7% bypass
• Score < 70 is slightly higher (though not by a lot)
• Score < 90 is ~13% bypass, because 87% are at or above 90

If your instructor or your detection tool uses a stricter threshold, the bypass rate is even lower. So that’s not good news for folks hoping for a consistent AI-evading strategy.

Main Conclusion

So, does Phrasly actually bypass ZeroGPT? The data says… mostly no. Yes, there might be a few outliers that get a lovely 0 or a low score, but that’s only 7% or so for the <50 threshold. That is simply not consistent enough to rely on if your entire point is to fool ZeroGPT into thinking it’s human content.

My own take on this

I think you should not rely on Phrasly if your primary objective is bypassing AI detectors. If you plan to use Phrasly, just know that your paraphrase is very likely to be flagged by ZeroGPT.

Suggested Extensions

It might be worthwhile to have a “Methodology” section in a future post that describes the code snippet used for analyzing these scores, plus a confidence band approach (like “likely human,” “uncertain,” “likely AI”) to interpret results in a more user-friendly manner. For simpler explanation, a confidence band approach just lumps the scores into categories. If your piece has a ZeroGPT score below 30, maybe it’s “likely human,” between 30 and 70 maybe “uncertain,” and above 70 “likely AI.”

Frequently Asked Questions

Q1. Does ZeroGPT detect Phrasly?

Yes, it mostly does. Around 87% of the tested paraphrases had extremely high scores (≥90), which means it is extremely likely to get flagged.

Q2. Why is Phrasly not bypassing ZeroGPT even though they claim to do it?

Even though Phrasly tries to paraphrase in a style that attempts to evade AI detectors, ZeroGPT probably uses advanced stylometric and statistical analysis. The rewriting alone is not enough.

Q3. Is it possible to pass with Phrasly?

Yes, but only about 7% of paraphrases scored <50, and 5% scored 0. That’s very low. You can’t rely on it consistently to pass.

Q4. Does the length of text matter?

No. Our correlation analysis says there is no direct effect from length alone.

Q5. Is ZeroGPT better than Turnitin?

They are just different. ZeroGPT focuses on a specific method of detection, while Turnitin has its own approach.

The Bottom Line

Phrasly is made to bypass AI detectors but the results show that it fails with ZeroGPT for the most part. A handful of paraphrases do slip by with a low or zero score, but it’s not enough to claim any reliability.

If you’re a student or a content creator worried about ZeroGPT, you either need to do your paraphrasing manually or look for a dedicated tool that can more consistently subvert AI detection.