Methodology
How We Evaluate Rewriting Quality
What our engine changes, how we judge whether a rewrite is good, and why we don't publish pass-rate claims against third-party AI detectors.
Last updated: August 5, 2026
The short version
AI2Human is a rewriting engine. It changes sentence structure, clause order, vocabulary register and phrase repetition while preserving your meaning, citations and argument. We evaluate it on whether the output reads naturally and stays faithful to the input — not on scores from third-party detectors, which we neither control nor can reliably predict.
Why we don't publish detector pass rates
It would be easy to put a large percentage at the top of this page. We don't, for four reasons:
- Detector models change without notice. A figure measured in one month can be wrong the next, with no announcement from the vendor.
- Scoring is proprietary and unpublished.No major detector documents its thresholds, so "pass" is not a stable or externally verifiable definition.
- Results vary enormously by input. Subject matter, length, register and source model all move the result. A single headline number hides that variance rather than describing it.
- It sets the wrong expectation.A pass-rate promise invites people to treat the tool as a guarantee about someone else's system. It isn't one, and we would rather say so plainly.
If you see a humanizer advertising a precise, stable pass rate against a named detector, treat that number sceptically — including when a competitor publishes it about us.
What the engine actually changes
Sentence length variation
Machine drafts tend toward uniform sentence length. The engine deliberately varies it, mixing short declarative sentences with longer subordinated ones.
Clause restructuring
Rather than swapping synonyms, it reorders clauses and changes which idea carries the main verb — the structural change that actually alters how a passage reads.
Phrase de-duplication
Repeated connectives such as “it’s important to note”, “delve into” and “in conclusion” are replaced or removed.
Register adjustment
Vocabulary is shifted toward the register you selected, so an academic draft doesn’t come back sounding like marketing copy.
What we hold constant
A rewrite that changes your meaning is a failed rewrite, however natural it reads. Our internal quality checks reject output that:
- alters a factual claim, figure or conclusion in the source text
- damages or drops an inline citation, reference or quotation
- introduces grammatical errors in an attempt to look less uniform
- drifts away from the register the user selected
Limitations we know about
- Quality is highest in English. The engine supports 40+ languages, but results vary by language and by how much training signal exists for that register.
- Very short inputs give the engine little structure to work with, so changes are minimal below roughly 50 words.
- Highly technical or formulaic text — legal boilerplate, method sections, code comments — has less structural room to vary without distorting meaning.
- The tool improves how text reads. It cannot supply analysis, evidence or judgement, and it does not remove the need to cite sources.
Acceptable use
AI2Human is a writing-quality tool for text you are entitled to submit. It is not intended to misrepresent authorship or to evade academic integrity rules, and using it that way breaches our Terms of Service. Your institution's or employer's policy on AI assistance always takes precedence over any tool, including this one.