Researchers Opting Out of AI Training Data: What to Know
Researchers can sometimes opt out of having their work used to train AI models, but there is no universal opt-out that covers every publisher, repository, dataset, or AI provider. The available route depends on who controls the material, what rights and permissions apply, and whether a provider offers a specific exclusion process.
Why research may enter AI training data
AI training datasets can include text, images, code, and other material collected from publicly accessible websites, licensed sources, or datasets assembled for research. Academic articles and related materials may be available through publisher platforms, institutional repositories, preprint servers, personal websites, and scholarly databases. Their presence online does not by itself establish that every use is authorized, prohibited, or subject to an opt-out.
“Research” also covers different kinds of material. A paper’s full text, abstract, figures, supplementary files, a researcher’s code, and a dataset may have different owners and terms. Copyright may belong to an author, publisher, or another rights holder; a repository may host a copy without controlling all reuse rights. Personal information or confidential research can raise separate concerns beyond copyright.
It helps to distinguish training from other AI uses. A system may index or retrieve documents to answer a question without using them to update a model’s learned parameters. Providers may call these activities different things, and policies change. An opt-out aimed at model training may not prevent indexing, search, or retrieval-augmented responses.
How opt-outs generally work
There is no single switch that researchers can set to remove all of their work from all present and future training collections. Opt-outs are usually specific to a provider or collection, and may apply only to material the provider can identify or control. Depending on the service, a request may concern a URL, an account, an entire domain, or particular content.
Some providers describe technical signals, such as instructions in a website’s robots.txt file or metadata, that automated crawlers may recognize. These signals are not a universal legal standard or a guaranteed deletion mechanism: compliance depends on the crawler and provider. They may also affect future collection rather than copies already gathered. A publisher or repository may manage its own crawler settings, but an author may not have permission to change them.
Other routes include a provider’s content-removal or privacy request form, a publisher’s rights-management process, or a dataset’s documented exclusion procedure. The request should identify the material precisely and state the requested action, such as preventing future collection for training. Read the provider’s scope and terms carefully; a form described as a removal request may address public display or search results rather than model training.
Who should researchers contact?
Start with the party that controls the relevant copy or rights. For a publisher-hosted article, consult the publisher’s permissions or author-rights guidance. For a university repository, contact its library or repository team. For a preprint server or data repository, review its submission agreement and contact the service operator. If the work is on a personal or laboratory website, check who administers the site and whether its settings can be changed.
- Authors: Check publication agreements, licenses, and any rights retained by coauthors before requesting restrictions.
- Principal investigators and lab managers: Review project agreements and data-management commitments, especially for funded, collaborative, or human-subject research.
- Libraries and research offices: Ask about repository policies, publisher terms, institutional guidance, and the correct contact for a formal request.
- Dataset creators: Document the source, permissions, and any exclusion request; notify downstream users if a dataset’s terms or contents change.
A researcher who did not create or control a work may still report a concern, but cannot necessarily authorize removal on behalf of its rights holder. For a paper with several authors, a publisher may require the corresponding author or an authorized representative to act.
What to check before making a request
First identify the exact content and its location. Record stable URLs, title, DOI if available, repository record, and relevant files. Then establish who owns or administers each copy and review the applicable publishing agreement, open license, repository terms, and provider policy. A Creative Commons license, for example, can permit specified reuse subject to its conditions; it does not mean every use is automatically covered, nor does it guarantee that a particular provider will accept an opt-out.
Next, ask what the request actually changes. Does it stop future crawling, exclude a URL from a training dataset, remove a copy already held, or address only a particular product? Does it cover derivatives, cached versions, or material supplied by a third party? Keep a dated copy of the request and any response. If a response is unclear, ask the provider to distinguish model training from indexing, retrieval, and public display.
Do not publish sensitive or confidential information merely to make an opt-out request. If a concern involves participant data, a data-use agreement, unpublished results, or a possible confidentiality breach, consult the institution’s research compliance or privacy office. An opt-out is not a substitute for securing data appropriately or correcting a publication’s permissions.
Limitations and common misconceptions
An opt-out usually cannot guarantee that material will be erased from models already trained on it. Training data can be copied, combined, or transformed, and a provider may not be able to identify every source retrospectively. Some services may offer deletion or unlearning procedures in specific circumstances, but these are not equivalent to a general promise that a model has forgotten a work.
Another misconception is that making a paper available online automatically permits all AI training—or that adding a notice automatically blocks it everywhere. Accessibility, copyright permission, contract terms, privacy obligations, and a provider’s technical practices are distinct questions. A robots.txt instruction may communicate a preference to compliant crawlers, but it does not change the rights in a work or bind every collector.
Researchers should also consider the benefits and trade-offs of restricting access. Reuse can support discovery and analysis, while exclusion may matter for licensing, confidentiality, attribution, or institutional policy. The choice should follow the actual rights and risks of the material, not a blanket assumption that every AI use has the same consequences.
Practical steps for an institution or research group
For a durable approach, create a simple inventory of important publications, datasets, code, and hosting locations. Record rights holders, licenses, repository contacts, and whether the material includes restrictions. Establish who may submit requests, how approvals are documented, and where responses are stored. Revisit the inventory when a paper is republished, a repository changes, or provider policies are updated.
When publishing new work, ask the publisher or repository about reuse and automated collection before signing or depositing. Use clear licensing and access terms where appropriate, and avoid placing confidential or restricted data in public files. If an institution wants to express a preference about crawling, its web or library team can assess whether site-level instructions are available and what they do—and do not—cover.
Frequently Asked Questions
Can I stop AI companies from training on my published paper?
Sometimes a provider or publisher offers a process to request exclusion, but there is no universal control. Check the rights agreement, the hosting service, and each relevant provider’s current policy.
Does robots.txt prevent AI training?
It can signal crawler preferences to services that recognize and follow the instruction. It is not a guarantee for all crawlers and may not remove data already collected.
Can an opt-out remove my work from a model that is already trained?
Usually, an opt-out is more straightforward for future collection than for changing an existing model. Ask the provider whether it offers a specific removal or unlearning process and what that process covers.
Can I opt out if my publisher owns the article?
You may need to contact the publisher or another rights holder. Check your publication agreement and retained rights; authors should not assume they can direct use of a copy they do not control.
Does an open-access license prevent AI training?
Not necessarily. The license terms determine which reuse permissions apply, and the provider’s conduct may involve additional questions. Read the specific license and seek institutional guidance if the intended use is disputed.
Related
Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.