For non-native English writers who keep getting flagged.

Most detectors score text on how predictable it is. The technical term is perplexity: if each word is close to the word a language model would have guessed next, the passage scores as machine-like. Variation in sentence length, sometimes called burstiness, gets weighed the same way.

That measurement punishes one kind of writer in particular. Someone working in a second language tends to use a narrower and safer vocabulary, reach for constructions they are confident are correct, and keep sentence structure regular. Those are the habits of a person being careful. To a perplexity model they are indistinguishable from generated text.

The size of the gap is the part worth sitting with. A 2023 study from Stanford found AI detection tools flag 61% of writing by non-native English speakers as AI-generated. For native speakers the false positive rate drops to around 3%. Same tools, same task, and the error rate multiplies roughly twentyfold depending on who wrote the essay.

Ask which tool produced the result and what number it returned. “Our system flagged it” is not a finding. If nobody can name the tool and the score, there is nothing to discuss yet, and saying so calmly is reasonable.

Then ask two follow-up questions. What false positive rate does that vendor publish, and what threshold does the institution treat as actionable? Most detector vendors document an error rate somewhere. Putting that number next to your score reframes the conversation from your honesty to the tool's reliability.

Gather what you already have before you write anything defensive. Version history in Google Docs or Word, dated file copies, outlines, research notes, library or browser history, messages where you talked the piece over with someone. None of it is conclusive on its own. Together it describes a process, and a process is much harder to fake after the fact than a finished document is.

One thing not to do: do not rewrite the piece so it scores better. Editing text specifically to defeat a classifier looks exactly like tampering, and it destroys the version history that was working in your favour.

A suggestion that circulates in student forums is to deliberately introduce small errors, break up your sentence rhythm, or push the draft through a paraphrasing tool so the detector reads it as human. Every version of this costs you something real. Errors cost marks. Paraphrasers leave their own statistical fingerprint, and several of them get flagged as machine output themselves.

The deeper problem is what the advice concedes. The clarity you spent years building is the exact property being penalised, and the fix on offer is to give it up as a tax on a tool's limitations. That is the worst trade available.

It is also a treadmill. Detectors get retrained, thresholds move, and whatever you learned to do last term stops working. Nothing you do to the surface of the text ends the argument, because the argument is about origin and the surface of the text does not contain that information.

Humanums does not read your writing and guess. It records how the writing happened: keystroke cadence, where you stopped and thought, how much you revised, how much text arrived by paste, how the work was distributed across sessions. The characters you type are never stored, only the timing and shape of the behaviour.

Certify a finished draft and you get a signed certificate, a public verification page, and a badge you can attach to the work. Whoever is questioning the piece clicks a link and reads the evidence instead of weighing your word against a percentage.

This approach is indifferent to how your English reads. A precisely constructed sentence and a rough one produce the same behavioural record if a person typed them, because the measurement never touches the prose. That indifference is the entire reason it does not inherit the bias that detectors have.

Overclaiming here would just make Humanums a different kind of unreliable tool, so the boundary matters. A certificate shows that the certified text was typed by a human, at human pace, with human pauses and revisions, inside a monitored session, and that the text has not changed since. It does not establish which human sat at the keyboard. It does not prove that no AI was consulted while you were thinking. It does not speak to whether the ideas are original.

It also cannot be applied backwards. If the essay already exists as a finished Word file, there is no session left to certify. The evidence has to be collected while the writing happens, which is a real constraint and worth knowing before you count on it.

And it does not overrule anyone's policy. If your school or employer runs a formal integrity process, a Humanums certificate is supporting evidence inside that process rather than a way around it. What it changes is the shape of the conversation: you arrive with something to show instead of only something to deny.