Nine routes are on record, from Common Crawl's Opt-Out Ledger, renamed on 2025-09-17, to an erasure request under Article 17 of Regulation (EU) 2016/679.
Three kinds of holder take requests: dataset publishers and hosts, the AI developers that train, and controllers under data protection law. This page reads each route from the side of the holder of the content, the person or company deciding between removal and licensing, within the wider field of AI training data licensing.
Which removal route reaches which copy
Nine routes are on record, and they differ less in how a request is sent than in which copy it reaches: a crawl index, a hosted file, a developer's future training or a controller's processing.
The table compares the routes by operator, sender, reach and limit, as of 2026-11-17.
| Route | Operator | How a request is sent | Who may use it | What it reaches | What it does not reach | Source |
|---|---|---|---|---|---|---|
| Opt-Out Ledger | Common Crawl Foundation | email info@commoncrawl.org; requests listed in the Opt-Out Ledger | anyone sending a legal request | subsequent crawls and public indexes | WARC files already published (immutable) | commoncrawl.org |
| Takedown form | LAION e.V. | form named on its GDPR page | anyone with a verifiable GDPR request | repositories LAION controls, current data, future releases | past releases circulating via torrents; the image on its host | laion.ai/faq |
| DMCA takedown notice | Hugging Face, Inc. | dmca@huggingface.co; counter-notification for uploaders | copyright holders | the hosted repository | earlier downloads and copies elsewhere | huggingface.co/content-policy |
| Report and privacy complaint | Hugging Face, Inc. | Report option on the content; safety@huggingface.co | anyone | content violating the policy, including third-party privacy | earlier downloads | huggingface.co/content-policy |
| PII removal form | Hugging Face (FineWeb card) | form linked on the dataset card | people who find their own PII | the FineWeb dataset | content that is not personal data | huggingface.co |
| Dolma Personal Data Removal Request | Ai2 | Google Form linked on the card and paper | the person or someone acting for them | the Dolma dataset | "all information" in OLMo outputs (not guaranteed) | forms.gle |
| Version opt-out cutoff | BigCode (The Stack v2) | opt-out request (mechanism not described here) | code authors | versions after the cutoff (v2.2.0: opt-outs before 2026-07-29) | earlier versions | huggingface.co |
| EU objection form | Meta | link in in-app and email notice | EU users of Meta products | future training on that user's public content and AI interactions | models already trained (Meta: "won't change in that moment") | about.fb.com; facebook.com/privacy |
| GDPR Article 17 erasure request | any controller of personal data | request to the controller; answer within one month | the data subject | the controller's processing; other controllers informed where the data was made public | processing within the Article 17(3) exceptions | eur-lex.europa.eu |
Three copies sit outside every route: copies downloaded earlier, datasets derived from them and the weights of models already trained. That is why the record of who copied what and when matters, and it is kept under training data provenance.
Takedown routes run by dataset publishers and hosts
Dataset publishers and the platforms that host their files each run their own route, and none of them recalls a release once it has been copied.
Common Crawl: email requests and the Opt-Out Ledger
The Common Crawl Foundation takes removal requests by email at info@commoncrawl.org and publishes every legal request it has received in its Opt-Out Ledger, a list it renamed from Opt-Out Registry on 2025-09-17 "to reflect the immutable nature of it".
In a statement of 2025-11-04, Common Crawl says that its WARC files are immutable and that affected URLs are removed or filtered from subsequent crawls and public indexes. It asks publishers to use robots.txt for future crawling. The Opt-Out Ledger page states no entry count, so none is given here.
LAION: the takedown form and future releases
LAION points people who want an entry removed to a takedown form on its GDPR page, and states that it will "remove the entry from all data repositories we control", including current data and future releases of the datasets (LAION FAQ, read 2026-11-17).
The limit is in LAION's own words: "We cannot act on data that are not under our control, for example, past releases that circulate via torrents".
LAION's sets hold links and captions, not images, so for removing an image from the web itself the FAQ refers the person to the original hosting provider. LAION took LAION-5B down on 2023-12-19 and released Re-LAION-5B on 2024-08-30, both covered under LAION-5B and Re-LAION-5B.
Hugging Face: report button, DMCA notices and the public notice folder
Hugging Face, which hosts many open training datasets, takes copyright takedown notices at dmca@huggingface.co and other reports through a Report option on the content or at safety@huggingface.co, under a content policy effective 2025-04-10. The uploader may challenge a takedown with a counter-notification, and privacy complaints fall under the policy's general rule on content that violates a third party's privacy.
Hugging Face publishes the notices it receives in the huggingface-legal/takedown-notices repository, which held 29 files in the 2025 folder and 30 in the 2026 folder up to 2026-10-12, recounted on 2026-11-17. One example is the University of Denver's notice of 2026-09-28 over AffectNet redistribution, which named 3 repositories.
A takedown removes the hosted repository, and nothing in the policy reaches forks or downloads made earlier.
Personal-data removal forms of single corpora
Some corpora built from the web crawl offer their own form for personal data: the FineWeb card says "If you find your own PII in FineWeb and would like it removed, please fill out our PII removal form", and Ai2 offers a "Dolma Personal Data Removal Request".
Ai2's form states that it "does not guarantee that all information about you will be removed from OLMo outputs". Both forms cover personal data only, and the dataset cards describe no route for a copyright holder or site owner. The Stack v2 records opt-out removals by version cutoff instead, described in the next section.
Registries that record a refusal before the next release
A registry or opt-out list records a refusal so that a publisher or developer can leave the content out of its next release or crawl, and it acts on nothing already built.
Signals sent from a website itself, such as robots.txt rules and machine-readable reservations, are a separate route compared under robots.txt rules for AI crawlers.
The Stack v2, a code dataset from BigCode, shows how a cutoff works: its card records that version 2.2.0 removed opt-outs received before 2026-07-29, so each version carries its own cutoff and an earlier version keeps what it held.
Spawning's do-not-train registry cannot be described from its own pages. Spawning's site and haveibeentrained.com showed only an "Under Maintenance" page when checked on 2026-11-17, so the registry's scope and the developers that honor it are not stated.
Objection and removal forms of AI developers
AI developers run forms for objecting to training or for removing personal information, and each form acts on that developer's own processing, not on public datasets. Account and platform settings that stop future training are compared under opting out of AI training in platform settings.
OpenAI's help center and privacy policy returned HTTP 403 when checked on 2026-11-17, so no OpenAI form is described here.
Google's Gemini Apps Privacy Hub (updated 2026-11-16, checked 2026-11-17) states that under laws including the GDPR a user may object to processing of personal data, or ask for inaccurate personal data in Gemini responses to be corrected, through a request in its Help Center or "Report legal issue" on a response.
Turning off Gemini history stops future chats being used to train generative AI models unless the user gives feedback, and reviewed chats are kept up to three years. The page names no form for public web data already used in training.
Meta's post of 2025-04-14 (updated 2026-03-27) says EU users receive notifications with a link to a form to object "at any time", and that Meta will "honor all objection forms we have already received, as well as newly submitted ones".
The privacy group noyb states that the effect is limited to objections made before training starts, and both claims stand as claims.
Meta's Privacy Center says that when a user deletes information a model learned from, "the model won't change in that moment", while the deleted information is not used for future training. The rulings on this form sit in the case record on Meta AI training in the EU.
Erasure requests under GDPR Article 17 against a dataset or a model
Where a dataset holds personal data, Article 17 of Regulation (EU) 2016/679, applicable since 2018-05-25, gives the person a right to erasure that any controller of that data must handle, publisher or developer alike.
What Article 17 grants and where it stops
Article 17(1) requires erasure without undue delay on six grounds, among them that the data are no longer necessary, that consent is withdrawn, that the person objects under Article 21(1), or that the data were unlawfully processed. The others are a legal obligation to erase and data collected from a child for information society services.
Under Article 17(2), a controller that made the data public takes reasonable steps, including technical measures, to inform other controllers processing it, which is the clause closest to a public dataset. Article 12(3) requires information on the action taken within one month of receipt.
Article 17(3) lists exceptions where processing is necessary, including freedom of expression and information (point a), archiving, scientific or historical research or statistical purposes (point d) and legal claims (point e). Research corpora may invoke these. The wider rules for training data are covered under GDPR rules for AI training data.
What erasure means for a model already trained
The CNIL's recommendations (page dated 2026-01-05) state that exercising rights over a model that is not anonymous "by default" requires retraining while the training data is still held, and that the right is "not absolute".
Retraining may be done periodically for several requests at once and "may take up to three months", and the person is told the timeline. Where retraining is disproportionate, the CNIL points to filters around the system, and a controller need not collect more data just to answer. The CNIL's page is guidance, not binding law.
The EDPB lists erasure of the dataset or of the model itself as a corrective measure (Opinion 28/2024, adopted 2024-12-17, para. 114), and FTC staff wrote on 2024-01-09 that prior orders required deleting models built with unlawfully obtained data.
These are regulator remedies, not a route a person triggers. Whether a model is outside the GDPR altogether is tested under when an AI model is anonymous under Opinion 28/2024.
Benefits, risks, downsides and alternatives of requesting removal, including licensing instead
A removal request takes content out of future versions at the cost of the holder's time, and it neither pays the holder nor reaches what was already copied.
Benefits:
- Content leaves the next release or crawl of the dataset that receives the request.
- An opt-out list or notice folder creates a public, dated record of the request.
- For personal data, Article 17 gives a one-month answer deadline and a legal duty on the controller.
Risks:
- A request names the content and the requester publicly where notices are published.
- A personal-data form does not guarantee removal from model outputs, as Ai2's form states.
- A route can be unavailable when needed, as Spawning's pages and OpenAI's pages were on 2026-11-17.
Downsides:
- Each publisher and developer needs its own request.
- Nothing reaches downloaded copies or trained weights.
- Research exceptions in Article 17(3) can defeat an erasure request.
Alternatives:
- Reserve rights in machine-readable form for future crawls.
- License the content on stated terms, so that use is agreed and recorded.
- Leave the content where it is.
- Take no action and do not license at all.
Where removal meets opt-out settings, crawler rules and the law
Removal is the after-the-fact route; the pages below cover the signals that act before collection, the datasets themselves and the contract terms that govern deletion in a license.
- Provenance records, standards and disclosure on the parent hub.
- Common Crawl: archive, terms and opt-out.
- FineWeb's filters and PII form.
- Dolma's personal data removal request.
- The Stack and its opt-out cutoffs.
- Major AI training datasets and their removal records.
- The shared AI training opt-out signal.
- The EU text and data mining opt-out.
- Deletion and machine unlearning clauses in a license.
- Visual artists: licensing, opt-out and compensation.
- Alternatives to selling data, including not selling at all.
- Chain of rights for training data.
- Training data certifications and labels.
Frequently asked questions
Can content be taken out of an AI training dataset?
Yes, from future versions of the dataset that holds it, by that publisher's own route. The route is an opt-out list, a form or a notice, and it never reaches copies already made.
Does removal from a dataset remove it from a trained model?
No, a model changes only by retraining or by filters. The CNIL describes both for personal data, and Meta states that its model "won't change in that moment" when information is deleted.
How long does a controller have to answer an erasure request?
One month from receipt under Article 12(3) of the GDPR. The CNIL describes retraining cycles of up to three months for models, and the person is told the timeline.
How do I ask LAION to remove an image link?
Use the takedown form that LAION names on its GDPR page. LAION removes the entry from repositories it controls and from future releases, not from torrented past releases.
How this page is sourced and dated
The page rests on the publishers' and developers' own pages (Common Crawl, LAION, Hugging Face, Ai2, BigCode, Google, Meta), the EUR-Lex text of Regulation (EU) 2016/679, CNIL and EDPB documents and an FTC staff post, with status as of 2026-11-17. This page is information, not legal advice.