Why Light AI-Text Paraphrasing Can Actually Increase Your Detection Score

Quick answer

Running AI text through a light paraphrase pass, the kind a "humanizer" performs, does not reliably lower a detector's score. A March 2026 benchmark study out of IIT (BHU) Varanasi and IIT Guwahati found something counterintuitive: light rewriting by a small instruction-tuned model, Qwen2.5-1.5B, left detection scores flat or higher than the unmodified AI text across five neural detector architectures and two separate text domains. Only heavy, aggressive rewriting moved scores down, and even then no detector dropped below a strong detection rate. Here's why. A light paraphrase doesn't erase the first model's statistical fingerprint. It stacks a second model's fingerprint on top of it, and detectors pick up both.

The study: a benchmark built to stress-test the whole detection stack, not just one tool

The paper is "Detecting the Machine: A Comprehensive Benchmark of AI-Generated Text Detectors Across Architectures, Domains, and Adversarial Conditions," by Madhav S. Baidya, S. S. Baidya, and Chirag Chawla, posted to arXiv in March 2026. It's a preprint, not yet peer-reviewed, so the specific numbers should count as early-stage evidence rather than settled fact, but the scale of the test is real: five fine-tuned transformer detectors (BERT, RoBERTa, ELECTRA, DistilBERT, DeBERTa-v3), trained on the widely used HC3 human-vs-ChatGPT corpus, evaluated across two text domains, HC3's mixed-topic Q&A and the more conversational ELI5 dataset.

To test paraphrase evasion, the researchers took 200 ChatGPT-generated samples from each domain and put them through three treatment levels: the original AI text left untouched, a light-humanization pass that varied sentence length and dropped formal transitions and numbered lists while preserving meaning, and a heavy pass that repeated that same process a second time with more aggressive pattern removal and looser grammar. Every rewrite, at both humanization levels, came from Qwen2.5-1.5B-Instruct, a small open-source model roughly the size and type many free "humanizer" web tools run under the hood.

What actually happened to the scores

Across the five detectors, light humanization didn't push detection scores down. In most conditions it pushed them up instead. RoBERTa's detection AUROC on HC3 text climbed from 0.9896 to 0.9911. BERT's score on the ELI5 domain jumped from 0.9530 to 0.9945. ELECTRA moved from 0.9540 to 0.9854 on the same domain. The paper's own summary puts it plainly: light paraphrasing "superimposes additional model-specific patterns, rendering the composite text more detectable," not less.

That's the counterintuitive part worth sitting with. A light paraphrase isn't a subtraction, it's an addition. The original model already wrote in a statistically distinct way: low burstiness, predictable word choices, a certain rhythm (the same signals most detectors are built to measure). Send that text through a second model and the first signature doesn't disappear. It gets a second layer stacked on top of it, from a different generator. Detectors built to spot "this doesn't read like a human wrote it" now have two AI signatures to find instead of one, not zero.

Heavy humanization did lower scores, but nowhere near what the "AI detection is dead" narrative claims, and detection never disappeared. RoBERTa barely budged, dropping from 0.9896 to 0.9621 even at the heaviest rewrite level tested. DistilBERT moved the most, falling from 0.9900 to 0.8567. Still comfortably above chance. Still flagging most texts correctly. No detector, in any domain tested, fell below an AUROC of 0.857. An AUROC of 0.5 means a coin flip; 0.857 means the detector is still calling it right far more often than not, even against the roughest rewrite the study threw at it.

Why this matters more than the usual "humanizers work" headline

Most coverage of AI-text humanizers frames the question as binary: does it beat the detector or not. This study's more useful contribution is showing that the answer depends heavily on how much you're willing to degrade the text, and that the lightest, least destructive version of "humanizing" is close to useless as an evasion strategy. That tracks with what we found looking at dedicated humanizer tools directly: the versions of paraphrasing that measurably move a score also measurably damage the writing, introducing the grammar slips and broken clauses that make humanized text read worse than the AI draft it started from. This study adds a mechanism for why the light version doesn't even get you that trade. You're not swapping detectability for readability. You're often getting worse readability with no detectability benefit at all.

It's also a reminder that "humanizer" is doing a lot of marketing work as a term. The study didn't test a commercial humanizer product. It tested what a small, generic instruction-tuned language model does when asked to rewrite text informally, which is functionally close to what many free humanizer tools run under a friendlier interface. If the underlying operation backfires at the light setting, wrapping that same operation in a nicer UI doesn't change the underlying math.

The part that doesn't change: cross-domain and cross-model behavior stayed uneven

The same benchmark separately found that all five detectors lost significant accuracy when tested on text from a different domain than they were trained on, a distinct finding from the humanization results but relevant context for reading them. DeBERTa-v3, in particular, collapsed to near-random performance, AUROC as low as 0.499, when trained on HC3 and tested on the ELI5 domain, regardless of any humanization applied. That's a domain-generalization weakness, not a paraphrase-evasion one, but it's a useful reminder that a detector's real-world reliability depends on more than whether it survives a rewrite. Scores from any single detector, on any single tool, are one data point, not a verdict, which is the same reason understanding what a detection score actually means matters more than chasing the number down. It's also why our own free detector builds a sentence-by-sentence breakdown instead of a single flat percentage: a document-level score can hide exactly the kind of layered pattern this study describes.

FAQ

Does using a free "humanizer" tool actually help avoid AI detection? Not reliably, based on this study's light-rewrite condition, and it sometimes pushes detection scores up instead of down. Heavier rewriting reduced scores somewhat but never wiped out detection on any tested detector.

Why would rewriting AI text make it easier to detect, not harder? A light paraphrase stacks a second model's statistical fingerprint on top of the first model's instead of erasing it, so detectors trained to spot AI-typical patterns get two sets to find.

Is this the same finding as "paraphrasing beats AI detectors" studies from a few years back? No. Those studies, the 2023 DIPPER paper especially, tested a large, purpose-built 11-billion-parameter paraphrasing model built specifically to dodge detection, and it did cut scores substantially against the perplexity-based detectors of that era. This study used a small, general-purpose 1.5-billion-parameter model doing ordinary light rewriting against newer fine-tuned transformer detectors. Different tool, different detector generation, different result. Read more about what the earlier paraphrasing research actually found, including its limits.

Should you run your own human-written text through a humanizer just to be safe? No. Nothing here suggests that helps. Humanizer tools alter patterns specific to AI-generated text, so running genuinely human writing through one adds the same grammar risk this and other studies have documented, with no detection upside, since your original text wasn't tripping any AI-pattern signal to begin with.

Check a specific piece of writing yourself

If you're evaluating a piece of writing and want more than a single number, run it through our free AI detector and look at where the flagged sections actually are, not just the headline score.