Full count of DEV posts finds 1 in 25 carries an AI-disclosure label

When DEV introduced three optional AI-disclosure labels — Hand Written, AI-Assisted and Fully Autonomous — on 26 August 2026, it published no figures on how often they were being used. A study posted on dev.to by MV2 of Munim has now filled that gap by counting the whole platform rather than sampling it, using DEV's public API, which returns each post's label. The snapshot spans 12 August to 14 October: 414,144 posts in total, 304,071 of them published after the labels appeared.

About 1 post in 25 carries a label

Across posts published since launch, 95.9% carry no label at all. Hand Written accounts for 881 posts (0.3%), AI-Assisted for 7,886 (2.6%) and Fully Autonomous for 3,647 (1.2%) — roughly one labelled post in every 25. Authors showed almost no interest in going back: only 76 of the 110,073 posts that predate the labels now carry one.

One burst inflated the AI-Assisted figure

The study's most striking finding is how fragile the headline number is. A tag table showed 1,912 posts tagged "massage training brussels", every one of them labelled AI-Assisted, sharing the same three tags and posted minutes apart — 664 on 24 September, another 1,245 by 27 September and 3 more on 4 October. Not one has a reaction or a comment.

That burst alone is 24.2% of all AI-Assisted posts since launch. Strip it out and the labelled share drops from 4.1% to 3.5%, with AI-Assisted at 2.0% rather than 2.6%. It also manufactured an apparent trend: AI-Assisted posts looked like they quadrupled in the week of 23 September (5.6%), but without the burst the figure was 1.8%. The study draws a general lesson: any per-post count on an open platform can be moved by a single account posting on a timer.

Disclosure is rising anyway

Even with the burst removed, the underlying trend is upward. Labelled posts per week grew from roughly 1,160 in the week of 16 September to about 2,630 in the week of 7 October, and Fully Autonomous posts went from around 350 to about 1,180 a week. On 14 October, the final day of the snapshot, 10.4% of posts carried a label.

The author cautions that part of this rise in share is a shrinking denominator: total listed posts fell from about 50,000 to about 30,000 a week over the same period, for reasons the data cannot explain. Posts deleted before the snapshot are invisible to it, so moderation may play a part.

Challenge entries disclose far more

Hacktoberfest entries tagged hf26challenge (3,766 posts) are labelled at 18.0% — 14.0% AI-Assisted, 3.2% Fully Autonomous and 0.8% Hand Written. Across all devchallenge posts the rate is 19.3%, nearly five times the platform-wide figure. The data cannot say why.

Who labels, and whether it pays

Label use is rare and concentrated: 4,937 of the 50,138 authors who posted since launch used a label at least once, and 109 authors account for half of all Fully Autonomous posts (165 for AI-Assisted, versus 1,422 for unlabelled posts). Meanwhile 1,636 authors have both labelled and unlabelled posts, so the label is chosen post by post rather than set once per account.

Labelled posts also rank higher in same-day reaction percentiles — Hand Written averages 0.66 against 0.50 for unlabelled posts, and no-reaction rates are lower across every labelled group. The study warns against reading causation into this: labelled posts far more often include cover images and complete metadata, and the AI-Assisted figure still contains the 1,912 zero-reaction burst posts.

Method and limits

The count is a single pass over DEV's /articles/latest endpoint, 1,000 posts per request, 418 requests in 43.5 minutes. Code is Apache-2.0 licensed and the per-post data CC BY 4.0, with usernames and author IDs never written out. Notably, the study discloses that it was itself produced by AI agents working for MV2 of Munim. Its stated limits matter: a label is an author's claim, not a measurement, and "no label" does not mean "no AI".

Why it matters

This is one of the few full-census measurements of voluntary AI disclosure on a large writing platform, and the early picture is double-edged: most authors say nothing, yet disclosure is growing quickly and Fully Autonomous posting is scaling fastest of all. It also demonstrates how easily open-platform statistics can be distorted by a single automated burst — and how strong a moderation signal that burst is, with nearly 2,000 posts attracting zero engagement. If DEV or any platform wants disclosure labels to mean something, this kind of continuous, reproducible counting with built-in burst detection is a workable template.