Updated November 2026

Free datasets from Clipwipe

Clipwipe publishes 8 small datasets on AI watermarks, AI labelling rules and the tools around them. Each one names its sources, records when its rows were checked, and is free to download as CSV or JSON and reuse under CC BY 4.0.

All datasets at a glance

What each dataset covers

AI generator watermarks

Which AI generators add a visible watermark, by plan and region: Google Gemini Apps, Flow and Gemini Notebook, OpenAI Sora, Microsoft Designer, Pika, Luma, Higgsfield, Adobe Firefly and Runway. 18 rows, each with a link to the maker's own page, the date it was checked and its caveats.

Download: CSV (11 KB) · JSON (16 KB). 18 rows, 15 columns. Rows last checked 2 Nov 2026 to 7 Nov 2026; rows last changed 7 Nov 2026.

Every row, what each column means and how it was checked: Which AI generators add a visible watermark?

AI labelling laws by jurisdiction

AI labelling and deepfake rules in 9 jurisdictions: European Union, United States (federal), United States (California), India, China, South Korea, Italy, United Kingdom and Vietnam. Each row gives the law, who it binds, what it requires, what it says about visible watermarks, its dates, its primary source and how and when it was last checked. General information, not legal advice.

Download: CSV (11 KB) · JSON (14 KB). 9 rows, 14 columns. Rows last checked 2 Nov 2026 to 13 Nov 2026; rows last changed 14 Nov 2026.

Every row, what each column means and how it was checked: Is removing an AI watermark legal?

Stock site rules on AI-generated content

AI-generated content rules for contributors at 6 stock sites (Shutterstock, Storyblocks, MotionElements, 123RF, Envato and Unsplash): whether AI is accepted, why, labelling, proof of rights, what happens to violators, the sentences quoted from each site's own page and that page's date. General information, not legal advice.

Download: CSV (14 KB) · JSON (17 KB). 6 rows, 15 columns. Rows last checked 13 Nov 2026; rows last changed 14 Nov 2026.

Every row, what each column means and how it was checked: Is removing an AI watermark legal?

AI content label rules by platform

AI content labelling rules on 4 platforms (YouTube, LinkedIn, Snapchat and Pinterest): whether you must label, what triggers it, how to disclose, what happens if you don't, the label's name, the sentences quoted from each platform's own page and that page's date. General information, not legal advice.

Download: CSV (6.3 KB) · JSON (8.0 KB). 4 rows, 15 columns. Rows last checked 13 Nov 2026; rows last changed 13 Nov 2026.

Every row, what each column means and how it was checked: AI label rules by platform

AI video clip length limits by model and plan

Clip length limits of AI video generators by vendor, model, surface and plan (Google, OpenAI, Runway, Kling, ByteDance and MiniMax): the longest single generation or the longest after extensions, the exact lengths allowed, conditions, status, the vendor page each value is read from and the date it was recorded.

Download: CSV (8.4 KB) · JSON (15 KB). 18 rows, 20 columns. Rows last checked 12 Nov 2026; rows last changed 14 Nov 2026.

Every row, what each column means and how it was checked: AI video clip length limits, by model and plan

What watermark removers say about uploads, retention, sign-up and AI training

What 12 watermark removers (RaveTools, MinoRemove, Markless, Wink (Meitu), Vmake, Media.io, Pixelbin, Unwatermark.ai, Dewatermark, WatermarkRemover.io, HitPaw and Magic Hour) say on their own pages about uploading your file, how long they keep it, sign-up and AI training, each with the vendor's quote, the page and its last-updated date. What vendors say, not a test.

Download: CSV (8.2 KB) · JSON (14 KB). 12 rows, 23 columns. Rows last checked 13 Nov 2026; rows last changed 13 Nov 2026.

Every row, what each column means and how it was checked: On-device vs upload: which watermark remover is safe?

Registration dates of sites ranking for AI watermark remover searches

Registration dates of the 165 sites seen in the Google and Bing top 10 from 1 Nov 2026 to 4 Nov 2026, for the searches Clipwipe checks every day (mostly AI watermark remover searches about Gemini and Veo, plus a few for its other tools and its own name): when each domain was registered according to its registry's RDAP record, when we first saw it rank, its best position and the engines. 161 have a date.

Download: CSV (26 KB) · JSON (71 KB). 165 rows, 14 columns. Rows last checked 6 Nov 2026 to 13 Nov 2026; rows last changed 13 Nov 2026.

Every row, what each column means and how it was checked: On-device vs upload: which watermark remover is safe?

Clipwipe's measured tool tests

Every run of Clipwipe's test bench against its own live tool pages from 9 Nov 2026 to 11 Nov 2026: 9 of the Gemini watermark remover, 6 of the Veo watermark remover and 2 of the Video to MP3 that finished, plus 1 run with no output, kept out of the figures. Each row gives the clip size, run time, output size, the bytes the browser sent after the file was chosen, a one-frame picture match where measured, the date and the browser. Headless Chromium (Playwright), service workers blocked, on one machine.

Download: CSV (5.5 KB) · JSON (12 KB). 18 rows, 18 columns. Rows last checked 9 Nov 2026 to 11 Nov 2026; rows last changed 11 Nov 2026.

Every row, what each column means and how it was checked: Clipwipe’s measured tool tests

Clipwipe’s measured tool tests

We run our own tools the way you would: open the live page, choose a video, wait for the file. 17 runs finished between 9 Nov 2026 and 11 Nov 2026: 9 of the Gemini watermark remover, 6 of the Veo watermark remover and 2 of the Video to MP3. Every one sent 5.0–9.1 KB over the network after the file was chosen, for clips of 2.7–53.9 MB, so the video itself was not uploaded.

One Video to GIF run (11 Nov 2026) is in the downloads but in none of the figures. Its note: “Stopped at the 300 s limit with no output: the bench did not press 'Make GIF'. Not a measurement of the page.”

What these runs don’t show:

• Picture quality is one frame. The match score (SSIM, 1 is identical) compares a single frame of the output with the input, not the whole clip. 1 run has no score.
• MP3 sound is unchecked. The Video to MP3 runs record the output size, not whether the file plays with sound.
• One browser, one machine. Headless Chromium (Playwright), service workers blocked. Your browser, device and other open tabs will change the times.
• Times vary run to run. On the same clip, the slowest run took up to 48% longer than the fastest (Video to MP3, clip A: 1.1–1.6 s); on the clips over 10 MB, up to 22%. Read a single run as a rough guide.
• The bytes are a total. They count everything the browser sent after the file was chosen (request headers, the page’s analytics, its own code), not where each byte went.

Download all 18 runs, 18 columns: CSV · JSON. New runs are added as rows, and this page’s date moves when they are. Free to reuse under CC BY 4.0.

How to read the dates

Rows last checked is the date range recorded in the rows themselves: when each row was last compared with its source, or, for the tool tests, the day each run was made. A monthly check shows as a month. Rows last changed is the dataset’s own date, which moves only when a row does, so a check that finds nothing new leaves it alone. Each dataset’s page explains how its rows are checked and what they leave out.

Using this data

You can reuse any of these datasets, including commercially, under CC BY 4.0. Credit Clipwipe with the dataset’s name and a link to its page. The licence covers our compilation and wording; anything quoted from a source stays its owner’s.

Spotted a row that is wrong or out of date? Email nocodetalks@gmail.com with the dataset and the row.

← Clipwipe home