The 2,590-Essay Corpus: How We Built the Style Reference Behind HumanizeMyAI

Why we started with a real corpus instead of synthetic data

Most AI humanizers in 2026 are trained on synthetic data. AI generates the text, a second AI rewrites it, a third AI grades the result, and the loop closes. The data is cheap, the pipeline is fast, and the output reads exactly like what it is: a model imitating a model imitating a model. We took the opposite path, and the reason is not ideology. It is that the thing we are trying to defeat lives in the gap between how a model writes and how a person writes, and you cannot close a gap you have never measured against the real thing.

A corpus of real human academic writing is the only reference that contains the texture detectors actually score. When you train on synthetic loops, you are teaching your tool to imitate the average of other tools. When you train on real essays, you are teaching it what a person under deadline pressure, with a specific argument and a specific voice, actually produces on the page. That difference is small in a single sentence and enormous across a 1,500-word essay. The corpus is the public part of our moat, and this post is the story of why it works, told at a level that does not hand competitors the build instructions.

What burstiness, perplexity, and lexical density actually measure

Detectors like GPTZero, Originality AI, and Turnitin's current model do not score on vocabulary alone. If they did, swapping a few flagged words for synonyms would beat them, and it does not. They score on statistical properties of the prose, and three of those properties carry most of the weight.

Burstiness is the variance in sentence length. A human writer fires off a three-word fragment, then a forty-word subordinate-clause marathon, then something in between, because thought is uneven and so is the page. A model produces sentences of eerily similar length because it is optimizing for the most probable next token at every step, and the most probable rhythm is a smooth one. Perplexity is roughly how surprised a language model is by the next word. Human writing surprises the model often, because people reach for the unexpected phrasing, the slightly-off word, the idiom that fits the moment. Model writing is low-perplexity by construction: it picks the safe word. Lexical density is the ratio of content words to total words, and it tends to sit in a tight band for machine output while real writing wanders.

Synthetic training loops smooth all three of these out. Every pass through another model regresses the text toward the safe, even, predictable center, because that is what models produce. By the time the text has been generated, rewritten, and graded by three models, the burstiness is flattened, the perplexity is low, and the density has collapsed into the band detectors recognize. A corpus of real essays preserves the unevenness, and that is the signal the model needs to learn from.

Why student essays are the highest-stakes case

We weighted the corpus toward formal student academic prose for three reasons, and each one is deliberate.

First, it is the hardest case for detectors and therefore the most useful teacher. Academic essays are structured, dense, citation-heavy, and register-consistent, which is exactly the profile that overlaps most with AI text. If the corpus can teach the model to read as human on a tightly-argued dissertation chapter, the casual cases take care of themselves. Second, it is the highest-stakes detection scenario in the world. Academic integrity systems are tuned aggressively, and a false positive is not an inconvenience. It is a disciplinary hearing for work the student actually wrote. The cost of getting this wrong is measured in academic records, not in a marketing metric. Third, the corpus has to match where our readers actually need help, and students writing essays are the largest single segment of the best AI humanizer market. Building the reference around their writing is not a constraint. It is the point.

The range from native overconfidence to ESL hedging

Student writing has more stylistic range than published prose, and that range is a feature we worked to keep. A published article has been edited toward a house style. A pile of real student essays has not, so it spans the full spectrum of how people actually write when the writing is theirs and the stakes are real.

At one end sits native-English overconfidence: contractions, casual pivots, the occasional sentence that breaks a rule on purpose, the rhythm of someone who learned the language by ear and trusts it. At the other end sits the careful, deliberate prose of a second-language writer who learned the formal conventions explicitly: measured hedging, parallel construction, transition phrases deployed exactly as the textbook taught them. Both are human. Both are valid academic voices. And a humanizer that only knows one of them will overcorrect the other, smoothing an ESL writer into native cadence they never used or roughing up a careful writer into a register that is not theirs. The corpus carries both ends of the spectrum precisely so the rewrite can respect the voice it started with.

The Stanford 2023 finding and why ESL voice gets preserved

The reason ESL voice preservation is not a nice-to-have is documented. A 2023 Stanford study (Liang et al., Patterns 4(7), DOI 10.1016/j.patter.2023.100779) found that seven major detectors flagged 61.3% of non-native TOEFL essays as AI-written, while native English samples were flagged at a small fraction of that rate (roughly 5.19% in the study's preprint). The cause is cruel in its simplicity: the formal conventions ESL writers are taught (consistent register, careful hedging, restricted vocabulary, parallel lists) are the same statistical patterns a language model produces by default. The writer is punished for being careful.

A humanizer trained on synthetic loops makes this worse, not better. If your tool learned to imitate the safe, even, low-perplexity center of model output, it will tend to push every input toward that center, which means it can erase the very non-native cadence markers a careful ESL writer used honestly, while leaving the statistical signature that triggered the false positive in the first place. That is replacing one problem with another. Because our corpus contains the ESL range as a first-class part of the reference, the rewrite has real second-language academic prose to draw on, so it can preserve hedging and non-native rhythm rather than overwriting it. We cover the full study, the calibration, and the institutional appeal chain in our Stanford ESL bias deep dive.

Months of human curation, kept at a high level

Raw text is messy, and shortcuts at the curation stage poison every choice downstream. The work took months, and most of it was unglamorous filtering rather than anything clever.

Every essay earned its place by being checked against major detectors before it entered the reference set. Only writing that scored reliably human made the cut. Instructor headers, assignment prompts, and bibliography metadata were stripped so the model learns prose rather than artifacts. Academic registers across humanities, social science, STEM, and business were balanced so the corpus does not over-represent one domain's habits. ESL voices were preserved on purpose, while passages with genuine structural breakdown, the ones where the argument or the grammar fell apart, were dropped. The final corpus is deliberately conservative: if a detector flagged a passage as AI-like even when we knew a human wrote it, that passage did not make it in, because we did not want those features feeding back into the model. The mechanism that turns this clean reference into a rewrite is proprietary by design, and this post is the why and the what, not the how. What matters from your side is that the writing the model learns from has already cleared the bar your text is trying to clear.

The monthly re-vet cadence, and why a frozen corpus rots

The corpus is not a snapshot. Detector vendors change their classifiers, sometimes quietly and sometimes with a public version bump, and a reference that was clean six months ago can drift out of calibration as the target moves underneath it.

So we re-vet the corpus against major detectors every month, and the evaluation set rotates with it. A passage that passed cleanly last quarter gets re-checked against the current classifiers, and anything that no longer holds up is pulled. This is the part of the discipline that almost nobody talks about, because it is ongoing cost with no headline attached. But it is the difference between a tool that worked once and a tool that keeps working. The detector landscape moves, so the corpus moves with it, and the published numbers in this post reflect the current state rather than a figure frozen from an earlier release.

Why this beats synthetic-trained competitors on the numbers

The architectural difference is not an argument. It shows up directly in cross-detector measurement, and the gap is large enough that it cannot be explained by tuning.

Synthetic-trained competitors behave differently on the same six-detector panel, though a correction belongs here: this post used to give per-tool means for StealthWriter and StealthGPT, and those figures were removed from the reviews they linked to in July 2026 because we never measured them. What the reviews carry now is dated third-party testing, tool by tool, with the conflicts of interest named. HumanizeMyAI's measured row on the same panel: 4% GPTZero, 8% Turnitin, 8% Originality AI, 6% Copyleaks, 30/30 pass on QuillBot's own detector, 3% ZeroGPT, for a 6.2% mean. The 23-percentage-point gap between our mean and a paraphraser-class competitor is not a cleverer synonym list or a better prompt. It is the corpus. A tool that learned from synthetic output is imitating the wrong target, so its best effort still carries the smoothed, even, low-perplexity signature detectors read below the word choice. A tool that learned from real human essays does not have that ceiling, because the reference it draws on never had the signature to begin with.

Why the moat is hard to copy

The 2,590-essay corpus is the public part of the moat, and stating the number out loud does not give it away. The reason is that the number is the easy part. The work is everything around it.

A competitor who wants to reproduce this cannot shortcut to it with synthetic data, because synthetic data reintroduces exactly the statistical signature the corpus exists to avoid. They would have to source real human academic writing across the full range from native to ESL, strip the artifacts, balance the registers, vet every passage against live detectors, drop the ones that flag even when human-written, preserve the second-language voice on purpose rather than smoothing it away, and then re-vet the whole thing month after month as the detectors evolve. That is months of unglamorous, ongoing human discipline, and it does not compress into a weekend of API calls. The barrier is not secret. It is just expensive, and most of the market chose cheap and synthetic instead. That choice is why their numbers sit where they sit and ours sit where they sit.

Verify the whole thing yourself

None of this asks for faith. The point of publishing the methodology, the study citation, and the full six-detector row is that every claim is checkable, and the easiest way to check it is to run it.

Paste any AI text you have into /humanize, then run the output through /detect, which is backed by the same engine logic, and read the score for the detector your institution or client actually uses. The matrix exists so you can ignore the headline and read the one cell that grades you. For the full ranked cross-detector comparison across nine humanizers, see our best AI humanizer 2026 listicle, and for the false-positive problem that hits second-language writers hardest, the Stanford ESL bias deep dive has the full calibration story. Registering is free and hands you four runs at 250 words each, which is enough to test a single passage against your own text before you trust anything written here. The corpus is the moat, the numbers are the evidence, and both are open for you to verify.

By Fırat Mıhcı (ResearchGate). $0 affiliate stake.