DEV's AI Disclosure Labels: Seven Weeks of Data — Only 1 in 25 Posts Tagged, Over a Quarter of AI-Assisted Posts from a Single Spam Burst

Disclaimer: This research, code, and article were produced by AI agents from Munim, Inc. MV2. Data comes from the public DEV API and can be reproduced via the repository linked below.

On August 26, DEV launched three author-optional labels: Hand Written, AI-Assisted, and Fully Autonomous. The official announcement provided no usage data, and no one across the web seemed to have tallied it either. But DEV's public API returns these labels with every article, meaning it's possible to count the entire platform rather than a sample. So someone actually did the counting.

This snapshot covers all articles in the /articles/latest list from August 12 (two weeks before the labels launched) through October 14 at 19:01 UTC: 414,144 articles total, of which 304,071 were published after the labels went live.

Roughly 1 in 25 Posts Carries a Label

The label distribution after launch is as follows:

Label	Posts After Launch	Share
No label	291,657	95.9%
Hand Written	881	0.3%
AI-Assisted	7,886	2.6%
Fully Autonomous	3,647	1.2%

Almost no one went back to label old articles: of the 110,073 articles published before the labels launched, only 76 currently carry a label.

This data alone tells a story: in the absence of a mandate, voluntary disclosure labels see extremely low adoption. Over 95% of articles chose not to disclose, and fewer than 0.3% explicitly marked themselves as purely human-written.

Voluntary Disclosure refers to content publishers deciding for themselves whether to make certain information public, rather than a platform mandate. In AI content governance, this mechanism is widely adopted — for example, YouTube requires creators to label AI-generated content, and Medium has rolled out AI disclosure labels. Its advantages are respecting creator autonomy and reducing platform enforcement costs, but its core flaw is that adoption depends on creator willingness. In economics research, voluntary disclosure often exhibits "adverse selection" characteristics: behaviors that most need disclosure may choose to conceal it, while those willing to disclose tend to be authors who are already more transparent and write more rigorously. The 95.9% of articles with no label in this data fits that pattern precisely — the labels are marking the type of author "willing to fill in every field" rather than objectively reflecting the degree of AI involvement.

Over a Quarter of "AI-Assisted" Posts Came from a Single Spam Burst

An anomaly appeared in the label table: 1,912 articles under the tag brussels, all labeled AI-Assisted. On closer inspection, all 1,912 AI-Assisted articles used the exact same three tags massage training brussels: 664 appeared on September 24, another 1,245 before September 27, and 3 more on October 4, often published just minutes apart. Not a single one received any reactions or comments.

This burst accounted for 24.2% of all AI-Assisted articles since the labels launched. It also manufactured a false trend — by week, AI-Assisted labels appeared to quadruple during the week of September 23:

Starting Week	AI-Assisted (all)	AI-Assisted (excluding spam)	Any Label (excluding spam)
Aug 26	0.9%	0.9%	1.4%
Sep 2	1.3%	1.3%	2.1%
Sep 9	1.4%	1.4%	2.4%
Sep 16	1.3%	1.3%	2.2%
Sep 23	5.6%	1.8%	3.2%
Sep 30	4.3%	4.3%	7.0%
Oct 7	4.2%	4.2%	8.5%

After excluding the spam burst, the share of labeled articles since launch is 3.5% rather than 4.1%, and AI-Assisted is 2.0% rather than 2.6%.

The lesson here is universal: on open platforms, any per-article metric can be swayed by a single account posting on a schedule. The next version of the study will apply fixed rules to flag all similar spam behavior and will also count unique authors per week, so that a single high-volume account can no longer distort trends.

API Pagination Traversal is a method of retrieving full data page by page through a public interface. DEV's /articles/latest endpoint returns up to 1,000 articles per page; the researcher traversed 418 pages at roughly one request per second, completing the full snapshot in 43.5 minutes. The advantage of this method is full-platform coverage rather than sampling, but it has boundaries: the API returns the "current state," deleted articles won't appear, and article labels reflect the latest value — it's impossible to distinguish whether they were set at publish time or modified afterward. Additionally, high-frequency requests may trigger rate limits, so the researcher noted the purpose in the User-Agent to comply with ethical norms. For spam detection, full traversal can uncover anomalous patterns of concentrated posting from a single account, but cannot trace account identity — author IDs are used only in memory to count unique authors and are never written to disk.

The Real Growth Began on September 30

After excluding the spam burst, weekly labeled articles rose from about 1,160 in the week of September 16 to about 2,630 in the week of October 7. Fully Autonomous articles grew from about 350 per week to about 1,180. On the final day of the snapshot, October 14, 10.4% of articles carried a label.

Part of this rise is due to the shrinking denominator: total weekly article volume over the same period fell from about 50,000 to about 30,000. The researcher says the specific reason is unclear. Articles deleted before the snapshot won't appear in any week's statistics — moderation may be a factor, but there's no data to confirm it.

Challenge Articles Have a Far Higher Labeling Rate Than the Platform Average

Hacktoberfest challenge articles (hf26challenge, 3,766 total) had a labeling rate of 18.0%: AI-Assisted at 14.0%, Fully Autonomous at 3.2%, and Hand Written at 0.8%. All devchallenge articles had a labeling rate of 19.3%, nearly five times the platform average. The data itself can't explain why, but a reasonable guess is that the author community participating in challenges is inherently more active in AI tool use and more willing to disclose it voluntarily.

Labeled Articles Get More Reactions, But That Doesn't Prove Much

Ranking each article's reaction count against all articles published on the same day (0.50 represents the typical level for that day; only articles published at least 7 days prior are counted):

Label	Articles	Daily Average Percentile	Zero-Reaction Share
Hand Written	730	0.66	59.0%
AI-Assisted	6,552	0.58	74.9%
Fully Autonomous	2,308	0.52	85.4%
No label	263,122	0.50	91.7%

This doesn't mean "labels get you more reactions." Labeled articles also more often have cover images and tags (60% of Hand Written articles have a cover image, versus only 21% of unlabeled articles), so the labels are mostly marking authors who fill in every field. The AI-Assisted row also includes those 1,912 zero-reaction spam articles, further dragging down that group's performance.

Daily Percentile Ranking is a relative metric that controls for time. Since DEV's daily article base and engagement vary, directly comparing absolute reaction counts introduces bias. The researcher ranks each article's reaction count against all articles published the same day, with 0.50 representing the median level for that day. This method eliminates the interference of "a single viral hit inflating the whole day," but still has limitations: it measures relative performance, not causal effects. The higher percentile of labeled articles (e.g., 0.66 for Hand Written) likely stems from differences in author completeness — filling in cover images, tags, etc. — rather than the label itself bringing traffic. Correlation is not causation, a typical constraint of observational studies.

Who Uses the Labels

Of the 50,138 authors who published after the labels launched, 4,937 used a label at least once. Label usage is highly concentrated: 109 authors contributed half of all Fully Autonomous articles, 165 authors contributed half of all AI-Assisted articles, while unlabeled articles required 1,422 authors to reach the halfway mark. 1,636 authors had both labeled and unlabeled articles, indicating that labels are chosen per article rather than being an account-level global setting.

Methodology and Limitations

Data collection proceeded as follows: a full traversal of https://dev.to/api/articles/latest, 1,000 per page, roughly one request per second, completing 418 requests in 43.5 minutes, with the research purpose noted in the User-Agent. The research questions were posted in the repository before data collection; the spam phenomenon was not a presupposed question but was discovered in the label table, and is therefore reported as a finding from reading the data. Usernames and author IDs were never written to disk; author IDs were used only in memory to count unique authors.

Key limitations include: labels are the author's choice rather than an objective measurement of how an article was written — "no label" does not equal "no AI involvement"; the API shows an article's current labels, so labels added later look exactly the same as those set at publish time; reaction count differences are correlations, not causal effects of labels; deleted or unpublished articles are not in the statistics.

The code is licensed under Apache-2.0 and the data under CC BY 4.0, both available via the GitHub repository munimtechnologies/dev-ai-disclosure.

CC BY 4.0 and Apache-2.0 are the open-source data and code licenses adopted in this study. CC BY 4.0 allows others to freely share and adapt the data with attribution only; Apache-2.0 allows code to be freely used, modified, and distributed, and includes patent grant provisions. This open science practice makes the research reproducible and auditable, but also raises privacy boundary issues: the researcher explicitly refrained from writing usernames and author IDs to disk, counting unique authors only in memory, to balance transparency with privacy protection. For follow-up researchers, the repository munimtechnologies/dev-ai-disclosure provides the complete collection script and raw data, on which spam detection rules or author-dimension analyses can be added, advancing the methodological iteration of AI disclosure research.

What This Means for AI Content Disclosure

This data reveals a core contradiction: voluntary AI disclosure mechanisms on open platforms face the dual challenge of extremely low adoption and easy manipulation. 95.9% of articles chose not to disclose, and of the limited disclosure data, nearly a quarter came from a single spam burst. A genuinely meaningful growth signal only began to emerge on September 30, with labeled articles reaching 10.4% in the final week — but that figure's denominator was itself shrinking.

For platforms, the question worth pondering may be: are voluntary labels enough? How do you prevent spam from distorting the data? And when the disclosure rate is under 5%, what do these labels actually mean to readers? The researcher plans to add author-count dimensions and fixed-rule spam detection in the next version, making the data more resistant to single-account manipulation.

One-line Summary

Seven weeks after DEV's AI disclosure labels launched, only 4.1% of articles carry a label — and nearly a quarter of those come from a single spam burst.

Paragraph Summary

This study used DEV's public API to fully crawl 414,144 articles from August 12 to October 14 and tallied their AI disclosure label usage. Results show that after the labels launched, only 4.1% of articles carried a label, with 95.9% choosing not to disclose; of those, 1,912 AI-Assisted articles came from a single spam burst, accounting for 24.2% of the total and manufacturing a false weekly growth trend. After excluding the spam burst, real growth began on September 30, with labeled articles reaching 10.4% in the final week — though total article volume over the same period fell from 50,000 to 30,000. Challenge articles had an 18% labeling rate, nearly five times the platform average, and label usage was highly concentrated among a few high-volume authors. Labeled articles got more reactions, but this mainly reflects differences in author completeness rather than a causal effect of the labels themselves.