humanflow

Corrections

Every claim we have published and got wrong, with the date it shipped, the date it was fixed, what we originally said and what is actually true. 8 so far. We do not edit quietly and we do not remove the page that was wrong.

Last updated 2026-09-26

A correction that only appears on the page it corrects is findable by everyone except the person who read that page last week. This log exists so the record can be read in one place, including by people checking whether we do what our methodology says we do, and by anyone deciding whether this site is worth citing — how to cite us points at this log rather than at itself for that. Who we are is on the about page.

Most entries below made HumanFlow look better than it was. That is the ordinary direction of a vendor's mistakes, and it is the reason this page is worth more than a promise to be careful. Where an error went the other way, it says so.

Published 4 August 2026 · corrected 26 September 2026

Word limits explained

What we published
That the Team plan had “no per-scan cap at all”. The Choosing a plan help article said the same thing in other words: Team “removes the per-scan cap entirely”.
What is true
Team has a per-scan cap, 30,000 words, and both help articles now read it from lib/pricing.ts like the other plans' caps.

This error made HumanFlow look better.

It was never true. While Team was being published as unlimited, the API held every plan to the same 10,000-character ceiling; on 11 August the pricing data gave Team the 30,000-word limit it has now. The pricing page changed with it; these two help articles kept the unlimited wording for another seven weeks.

Published 17 August 2026 · corrected 26 September 2026

UGC rules on AI and plagiarism, explained

What we published
That the government's ShodhShuddhi programme is “a plagiarism detection service”, so an AI score quoted at a student in an Indian university “may not be the one the government supplied”.
What is true
INFLIBNET's FAQ says the DrillBit licence it supplies through ShodhShuddhi produces a separate AI report for all submissions without requiring any settings. DrillBit's own user guide excludes regional-language uploads.

This error flattered nobody — it was just wrong.

We have not established when the AI report joined the national licence, so the sentence may have been wrong from the day it was published. It surfaced while building our DrillBit page, which reads the same FAQ.

Published 18 August 2026 · corrected 26 September 2026

Does AI detection work on PDFs?

What we published
That Turnitin's AI writing detection supports “English, Spanish and Japanese”, so a PDF in any other language gets no score.
What is true
It supports English, Spanish, Japanese and Arabic. Turnitin's release notes date Arabic support to 18 August 2026, the day this page shipped.

This error flattered nobody — it was just wrong.

The same three-language list sat on other pages, including the Turnitin pillar, the accuracy page, the section hub and the detector comparisons. There it was right when written and went out of date on 18 August. Those pages were corrected on 26 September too. It took 39 days because nothing re-reads a vendor's release notes on a schedule.

Published 9 September 2026 · corrected 26 September 2026

Can you trust a 98% accuracy claim?

What we published
“Three languages for detection”, supported by a quotation from Turnitin's FAQ naming long-form English, Spanish and Japanese.
What is true
Turnitin added Arabic on 18 August 2026, three weeks before the post went up. Its FAQ, as worded on 26 September 2026, names long-form English, Spanish, Japanese and Modern Arabic, and the post now quotes that wording.

This error flattered nobody — it was just wrong.

We cannot tell whether the FAQ itself still named three languages on 9 September. The release notes already listed Arabic, and they were the source to check.

Published 4 August 2026 · corrected 15 August 2026

HumanFlow vs Walter Writes

What we published
The per-request cell of the Walter Writes comparison table read “[VERIFY] — not stated in the material we could retrieve”. An editor's note to himself, published as though it were a figure.
What is true
The cell now reads “Not stated in the material we could retrieve”, which is what we actually know: their site refused our reads, so the figure was never established.

This error flattered nobody — it was just wrong.

It was live for eleven days, and not only on the Walter Writes page — the same row appears in the comparison table on eleven /alternatives pages, so the marker rendered on twelve URLs in total. Our content gate did not catch it because it only ever read the blog, and this copy lives in TypeScript. The gate now reads the source as well.

Published 4 August 2026 · corrected 15 August 2026

The best AI humanizer, ranked on what you can verify

What we published
The page said “eleven tools” in five places — the meta description, the opening line, the byline, a section heading and an FAQ answer.
What is true
The table had thirteen rows the whole time, and its structured data said so. The count is now computed from the list instead of typed by hand.

This error flattered nobody — it was just wrong.

The count went stale the day the table grew and nothing recalculated the word. On a page whose argument is that its figures can be checked, a number contradicting the table beneath it is the cheapest possible way to lose that argument.

Published 4 August 2026 · corrected 4 August 2026

HumanFlow vs Grammarly

What we published
That Grammarly did not offer AI humanizing, and that its AI detection sat behind a paid Pro plan.
What is true
Grammarly publishes a free AI humanizer, plus a Humanizer agent with four voices and a custom voice profile built from a 200-word writing sample. Its standalone AI detector is free.

This error made HumanFlow look better.

Both halves were wrong in the same direction, which is the direction worth being suspicious of. The correction stands at the top of that page rather than in a footnote.

Published 4 August 2026 · corrected 4 August 2026

HumanFlow vs WriteHuman

What we published
That WriteHuman's entry plan was $18 a month, and that they were cheaper than us on annual billing.
What is true
Their Basic plan is $12 a month on annual billing against our $10.80, so we are the cheaper entry — and $18 is their Pro tier, not their entry.

This error made a competitor look better.

The rarer direction: this one had us understating our own position, by comparing their mid tier against our entry.

Found something wrong?

Tell us at support@humanflow.pro. If it changes a conclusion it gets an entry here, whichever way it cuts. How we source and review what we publish is set out in our editorial policy.