Why do you not publish bypass rates?

Because no reproducible public benchmark exists. Detectors change without notice, disagree with each other on identical text, and any figure would be produced by us about our own product. We would have to either mislead you or define the claim until it meant nothing.

Where do competitor figures come from?

Each vendor's own pricing or product page, read on a stated date that appears on every comparison page. Nothing comes from affiliate roundups. Where a vendor blocks automated access we say so on the page rather than reusing an old number as though it were current.

What happens when you get something wrong?

We correct it on the page, in visible text, rather than editing quietly. Two competitor pages currently carry corrections stating what we previously published and why it was wrong — including one where the correction makes our own product look worse.

Do you take money from the companies you compare?

No. There are no affiliate links, no sponsored placements and no paid inclusions anywhere on this site. We compete with every tool we write about, which is a conflict no disclosure fully cures — the mitigation is that every figure is checkable against the vendor's own page.

How will the public benchmark work?

A fixed corpus published in full, named detector versions with test dates, identical settings across tools, and raw results released alongside the summary. Until all four are in place we publish nothing, because a benchmark missing any of them is the thing we criticise others for.

How we test and write

Every competitor figure comes from the vendor's own page on a stated date. We publish no bypass rates, because none can be measured honestly. When we get something wrong we correct it in visible text on the page rather than editing it away.

Last reviewed 4 August 2026

Sourcing

Competitor prices, word allowances and free-tier limits are read from the vendor's own pricing or product page. Every comparison page states the date it was read and links the source, so you can check both the figure and how stale it is. There are currently 12 of these.

Nothing is taken from affiliate roundups or review aggregators. Those recycle each other's numbers, and a price that has been copied four times is not a source.

Several vendors block automated requests. Where that happens we say so on the page, state that figures came from their own domain via search rather than a direct read, and mark the specific values we could not establish rather than filling them with plausible ones. Two pages currently carry that notice.

Research claims get the same treatment. Where a study is cited, it is named, linked and attributed — and where a study was funded by a party with an interest in the result, the page says so. That applies to research favourable to detection as much as research against it.

Why there are no bypass rates here

The obvious number for a company like ours to publish is how often our output gets past detectors. We do not publish it, and the reasoning is worth setting out because it costs us traffic.

A bypass rate would need a fixed corpus, a date, the version of each detector tested, a threshold that counts as passing, and a method someone else could repeat. Vendor figures in this category ship with none of those. They are produced by the company selling the tool, on text it chose, against detector versions that no longer exist by the time you read the number.

Detectors also change without changelogs and disagree with each other on identical text — the same paragraph can read 12% on one and 89% on another. A percentage published against that is a claim about systems the claimant does not control and cannot observe.

So: HumanFlow does not promise to beat any detector, because nobody honestly can. We compare on price, allowances, caps and published features, all of which you can verify without trusting us.

The benchmark we intend to publish

The honest version of the missing number is possible; it is just expensive and slow. When we publish one it will meet four conditions, and if it cannot meet all four we will not publish it.

A published corpus. The full test set released alongside the results, so anyone can rerun it. A benchmark on text nobody else can see is a press release.

Named detector versions and test dates. Not "tested against Turnitin" but which build, checked on which day, with the understanding that the result expires.

Identical settings across tools. Same input, same tier, same mode. Most vendor comparisons quietly run their own product on its best setting and rivals on their defaults.

Raw results, not only a summary. Per-sample outputs published so someone can find the case that undercuts our conclusion.

Until then the placeholder on our ranked list of humanizers stays empty. An empty column is more honest than a filled one we cannot defend, and a competitor's filled-in version is not better evidence than our blank one.

Ranking, and our thumb on the scale

We rank tools we compete with, which is a conflict of interest that cannot be disclosed away. Two things reduce it.

The ranking rule is published above the table rather than below it, and it is a single stated criterion with a tiebreak — free words per request, then entry price. We deliberately do not compute a weighted composite score, because choosing the weights is exactly where a vendor ranking itself would put its thumb.

And the rule is applied to us on the same terms as everyone else. Where it puts us first, the page says what first place means — that we accept the most free words in one request, not that we are the best tool. Where competitors beat us, the pages say so: several are cheaper, one has a far more generous free tier, and several ship an API, a browser extension and mobile apps where ours are in private beta.

Corrections

When we publish something wrong, the correction goes on the page in visible text, saying what we previously claimed and why it was wrong. We do not silently edit and we do not remove the page.

This is not hypothetical. Our Grammarly comparison opens by correcting two claims we published about them — that they had no AI humanizer, and that their AI detection required a paid plan. Both were wrong, both flattered us, and the page says so. Our WriteHuman page carries a correction to their entry price and to a comparison we had backwards.

If you find something wrong, tell us at support@humanflow.pro. Corrections to competitor figures are the most useful thing you can send, because pricing moves faster than we recheck it.

Who writes this

Pages currently carry a team byline rather than a named author. We would rather show an organisational byline than invent a person with credentials, and named authorship is something we intend to add properly rather than fabricate.

Our editorial policy covers the commercial relationships; the changelog records what has shipped.

Common questions

Why do you not publish bypass rates?: Because no reproducible public benchmark exists. Detectors change without notice, disagree with each other on identical text, and any figure would be produced by us about our own product. We would have to either mislead you or define the claim until it meant nothing.
Where do competitor figures come from?: Each vendor's own pricing or product page, read on a stated date that appears on every comparison page. Nothing comes from affiliate roundups. Where a vendor blocks automated access we say so on the page rather than reusing an old number as though it were current.
What happens when you get something wrong?: We correct it on the page, in visible text, rather than editing quietly. Two competitor pages currently carry corrections stating what we previously published and why it was wrong — including one where the correction makes our own product look worse.
Do you take money from the companies you compare?: No. There are no affiliate links, no sponsored placements and no paid inclusions anywhere on this site. We compete with every tool we write about, which is a conflict no disclosure fully cures — the mitigation is that every figure is checkable against the vendor's own page.
How will the public benchmark work?: A fixed corpus published in full, named detector versions with test dates, identical settings across tools, and raw results released alongside the summary. Until all four are in place we publish nothing, because a benchmark missing any of them is the thing we criticise others for.