AI Humanizers for Student Writing: A 29-Text Comparison

StealthGPT ranked first among 11 AI humanizers on 29 academic texts, with an 83.9% mean strict-bypass rate across six external detector models. Undetectable.ai scored 77.6% and WriteHuman scored 71.3%. StealthGPT also cleared every detector on 12 of the 29 texts; Undetectable.ai did so on one, and the other nine humanizers on none.

This is a new student-writing analysis of HumanizerEval’s October 22, 2026 review, calculated on November 7, 2026. We selected academic genres from the published source list, then recomputed their results from the original score CSV. The 29-text competitor tests took place in October; that table reports the dated evidence. A separate November head-to-head below tests new academic passages against WriteHuman and Undetectable.ai.

A fresh November head-to-head adds complete outputs: Super met all three stated detector thresholds on 2 of 3 academic passages, compared with 0 of 3 for WriteHuman Standard API and 0 of 3 for Undetectable.ai v11sr. The original October comparison and this new three-passage demonstration are presented separately, with their own data downloads.

For an essay, research discussion or reflective assignment, the useful question is how a humanizer handles the kind of writing you actually work with. This study isolates those genres so you can compare the tools on academic text. StealthGPT’s tested configuration was Super in quality mode through the API. You can use Super in the AI Humanizer workspace.

The 29-text student-writing comparison

The same 29 source texts were used for every humanizer. Each resulting output was evaluated by the same six detector models. A strict bypass means that the published AI percentage was below 50 for that detector. The mean in this table is the average of six detector-specific pass rates. The final column counts texts that cleared all six detectors together.

The untouched source texts scored 0.0% mean strict bypass in this subset, and none cleared all six detectors. StealthGPT’s lead over Undetectable.ai was 6.3 percentage points on the mean; its lead over WriteHuman was 12.6 points. Those are descriptive differences within these 29 texts.

Twelve texts cleared all six detectors

An average pass rate and an all-detector result answer different questions. A tool can pass several easier detectors on every text while repeatedly missing a harder one. The mean captures its performance across the panel; the all-six count asks how often an individual output cleared the whole panel.

StealthGPT produced 12 outputs that cleared the entire six-detector panel. Undetectable.ai produced one. None of the other nine humanizers produced an all-six result in the selected academic texts. That is the strongest practical distinction in this analysis: StealthGPT’s result was supported across the full panel on more of these passages.

StealthGPT, Undetectable.ai and WriteHuman by detector

These are the three highest-scoring humanizers in the academic subset. Each cell shows how many of the same 29 outputs cleared that detector. Keeping the models separate makes the strengths and differences visible.

StealthGPT’s Pangram result was 28 of 29, while Undetectable.ai passed 17 and WriteHuman passed 10. On Originality.ai 5.0, StealthGPT passed 12 texts, compared with two for Undetectable.ai and one for WriteHuman. Undetectable.ai and WriteHuman performed strongly on several other detectors, as the table shows.

These results help explain why a useful humanizer comparison needs a panel. A single detector would give an incomplete view of the same tools and the same academic material. The full HumanizerEval review includes all 50 original texts and every product’s results.

What kinds of student writing were included?

The subset covers 29 English texts from five source models, ranging from 364 to 994 words. The original texts were AI-generated benchmark material. They were not collected from real students, course submissions or customer papers.

• Argumentative, compare-and-contrast and reflective essays.

• History, psychology, science, sociology, philosophy and other subject essays.

• Literary and film analyses, plus a literature review.

• A biology lab-report discussion and a healthcare ethics case analysis.

• Class discussion posts, a scholarship essay and a student reflection.

You can inspect the complete source list and open each sample to see the original text, prompt and source model. The download below lists every selected sample ID and genre, so the membership can be checked against the published source CSV.

How we selected and recalculated the academic subset

Before calculating subset scores, we saved one inclusion rule: include every published genre containing “essay”, plus the exact labels literary analysis, lab report discussion, healthcare ethics case analysis, literature review, discussion board post, film analysis and reflection. Every source row that matched was included. We did not choose samples by which product won, by detector score or by source model.

For each product and detector, we counted the selected rows whose published AI percentage was below 50. We divided by 29 to obtain the detector’s strict-bypass rate, then took the arithmetic mean of the six rates. To calculate the all-six count, we counted rows that were below 50 on all six detector columns. Raw detector percentages were kept separate.

This is an exploratory genre analysis of an existing benchmark. The original experiment was run by the StealthGPT team using external detector APIs. The public archive supplies the source texts, per-sample scores, recorded settings and methodology. Humanized output texts are withheld in that archive, so this subset analysis measures the published detector outcomes rather than making new claims about the academic accuracy of each competitor’s rewrite.

For the full experimental design, detector versions and model settings, see HumanizerEval’s methodology. The study data are published under CC BY 4.0. This article adds genre selection, calculations and interpretation; it does not change the original test records.

Download the academic-text results

Download the academic results CSV for all 29 selected texts across 11 humanizers plus the original source baseline. The reproducibility pack includes the selection rule, sample membership, source metadata, original public data and verification script. It contains no private account credentials or customer submissions.

Attribution: HumanizerEval, “October 2026 review”, October 22, 2026. Academic-genre subset and recalculation by StealthGPT, November 7, 2026. The original source-text CSV has SHA-256 ae28a51cec25560fb8520aeb444e0f78cc69ec8311de21725dbca8df58e90349; the score CSV has SHA-256 fa08bce7ac679b6c067c2cbc39948e82940e6834a3c396562aae6d480037debc.

Fresh November test: Super vs WriteHuman and Undetectable.ai

We also ran a new, separate head-to-head on November 7: an argumentative essay, a laboratory discussion and a research-paper limitations passage. Every service received the same original 349–357-word texts. We saved the first returned outputs before scoring and kept the original baselines. These passages were original AI-assisted editing fixtures; the numerical datasets were explicitly fictional teaching material.

Super met all three stated detector thresholds on two of the three passages; neither WriteHuman nor Undetectable.ai did so on any passage. This small matched demonstration publishes all outputs, including the results that missed a threshold. The counts describe these three texts; they are separate from the 29-text October benchmark above.

StealthGPT used Super, quality mode and plain-text output without protected spans. WriteHuman used its Standard API model, academic tone and English, with Enhanced disabled on the tested plan. Undetectable.ai used v11sr, University readability, Essay purpose and More Human strength. We retained the first returned output from each service, without rerolls. The observed detector versions were GPTZero 2026-11-04-base, Originality.ai model-5.0 and Pangram 4.1. Each detector received exactly the saved output, verified by its text hash.

Detector results and factual fidelity are separate checks. In the research example, Super broadened “missed” a class to “late to or missed”; WriteHuman changed the source’s 35% into 35 respondents and reported 57%. Those changes need correction against the original. The evidence pack contains the full outputs and specific editorial findings, so the detector comparison can be inspected alongside meaning.

Download the complete fresh head-to-head evidence pack for the three originals, all Super, WriteHuman and Undetectable.ai outputs, exact settings and provider evidence. The 36-row detector CSV includes the originals and every service’s rewrites. The original/source and Super scans were reused from the same-day model demonstration; nine new scans evaluated WriteHuman and nine evaluated Undetectable.ai.

These are operational threshold counts, kept separate from each service’s native verdict labels; raw scores are not averaged. For example, the Undetectable.ai research output had a Pangram human-text fraction of 76.7%, meeting the table’s threshold, while Pangram’s headline still read “AI Detected”. Both the fraction and headline are retained in the evidence. Originality served model-5.0 for every compared text, including the earlier Super/source requests that asked for Turbo. Turnitin was not part of this head-to-head.

Which StealthGPT model should students choose?

Start with Super when your priority is the strongest humanization demonstrated by this benchmark. That is the StealthGPT configuration actually tested in the October comparison. Standard and Lite were not part of that competitor panel.

For a direct look at the current models, read our separate Super vs Standard vs Lite comparison. It uses three new fixed academic passages, publishes the actual outputs and records the latest detector results alongside editorial checks. Its November test is kept separate from the October competitor data here.

After choosing a model, compare the rewrite with your draft. In academic writing, the argument, qualifications, cited sources and numerical meaning all matter. The ChatGPT essay guide walks through that revision process, and our research-paper guide covers section-level editing and source checks.

Frequently asked questions

Which AI humanizer ranked first on these academic texts?

StealthGPT ranked first of 11 humanizers on this 29-text academic subset, with an 83.9% mean strict-bypass rate across six detector models. Its tested setting was Super in quality mode through the API.

Does 83.9% mean every text passed every detector?

No. It is the mean of six detector-specific pass rates. Twelve of the 29 StealthGPT outputs passed all six detectors on the same text. The study reports those two measures separately.

Was Turnitin tested?

Turnitin was not tested in this panel. The measured detectors were Pangram 4.1, GPTZero 2026-10-07-base, Originality.ai Turbo, Originality.ai 5.0, ZeroGPT and Winston AI 5.0.

Were competitors tested again in November?

The 29-text table recalculates the original October 22 tests. We also ran a separate November 7 head-to-head: StealthGPT Super, WriteHuman Standard API and Undetectable.ai v11sr rewrote the same three new academic passages. Complete outputs and external detector evidence are published above.

Explore StealthGPT’s AI Humanizer, open the Humanizer workspace, or compare the current plans to choose the workflow that fits your writing.