Editorial standards

How we choose what to write about

Every article starts from a question we have actually been asked, or one we had to answer for ourselves while building the tools on this site. That means a lot of our writing covers narrow, practical ground — why a particular phrase reads as machine-written, what a detector score actually measures, why grain has to follow luminance to look real. We would rather publish one article that answers a question completely than five that circle it.

We do not publish articles because a keyword looked promising. If we can't say something specific and useful about a topic, we skip it.

How articles are researched

Claims about how a tool behaves come from using the tool. When we describe what a detector does with a given passage, or how an image changes after processing, we ran that passage or that image ourselves. Where an article describes a technical mechanism — perplexity, burstiness, sensor noise, demosaicing — we describe the mechanism plainly and note where the explanation is simplified.

We try to be explicit about uncertainty. Detector accuracy in particular is frequently overstated by the companies selling detectors, and we say so rather than repeating vendor numbers as fact.

Our use of AI in our own writing

It would be strange for a site about AI-assisted content to be cagey about this. We use AI tools the way we suggest readers use them: for drafting scaffolding, for rephrasing a paragraph that isn't landing, and for catching repetition. Every published article is then read, restructured, fact-checked, and edited by a person, who is responsible for what it says. We do not publish unedited generated text, and we do not publish articles nobody has read end to end.

Corrections

If we get something wrong, we fix the article and note the change. For substantive corrections — a claim that was wrong, not a typo — we add a dated note at the foot of the article rather than quietly editing it. Articles revised after publication carry an "Updated" date in the byline.

Spotted an error? Email hello@humanetext.com and we will look at it.

Updates and review

Anything we write about detection tools or model behaviour has a short shelf life. We revisit articles in those categories when the underlying tools change materially, and we date every revision. An article with no update date has not needed one.

Disclosure: advertising

This site carries advertising served through Google AdSense, which is what pays for the tools to stay free. Advertisers have no input into what we publish, no advance sight of articles, and no ability to request changes. We do not accept paid guest posts, paid links, or sponsored articles. If that ever changes, it will be disclosed on this page and labelled on the article itself.

Disclosure: our own tools

We build and give away the Text Humanizer and Photo Humanizer, and we link to them from articles where they are genuinely relevant. That is an obvious conflict of interest, so we try to handle it honestly: when a manual technique works better than our tool, we say so, and when our tool can't do something, we say that too. Our comparison articles include our own tools' weaknesses alongside competitors'.

What we will not do

We won't promise that any tool defeats any specific detector. That claim is unverifiable, changes weekly, and encourages exactly the use cases we think people should avoid. Our position on academic and professional integrity is set out in this article and in our terms of service.