Can you remove yourself from an AI training dataset?

LAION, the non-profit behind some of the largest public image datasets, describes its datasets as indexes to the internet: lists of URLs to the original images with the alt text found alongside them. LAION says the datasets contain no original images, and that anyone using them has to download the images from those links.

So removing a dataset entry does not take the image off the web, and removing the image from its host means the link no longer leads to it for anyone who downloads the dataset afterwards. Copies already downloaded are not affected by either step.

1. Remove the image at its source

Delete the image yourself if it is on your own account or site. If someone else hosts it, ask the site owner or platform to take it down. LAION gives the same advice: to remove content from the web entirely, contact the original hosting provider.

2. Ask for the dataset entry to be removed

LAION's FAQ says that if a URL or picture in its dataset contains your image, you may request a takedown of the entry through a form it provides under the GDPR. If the request is verifiable, LAION says it removes the entry from all data repositories it controls, including future releases. It also says plainly that it cannot act on data outside its control, such as past releases that circulate via torrents.

3. Object to future training

In the EU, the GDPR gives you a right to ask an organisation to erase your personal data and a right to object to processing based on legitimate interest, a legal basis the European Data Protection Board has examined for AI training. The organisation must respond within one month, extendable by two months for complex requests. The UK's Information Commissioner's Office describes the same right to object and the same one-month limit.

Some AI providers publish their own routes. OpenAI's privacy portal lets account holders ask that the content they submit not be used for training, and lets anyone ask for their personal data to be removed from ChatGPT responses. OpenAI describes that second request as working to prevent the information from appearing in responses going forward, which changes outputs, not the trained model.

A trained model does not hold a folder of its training images. Training adjusts the model's numeric weights, and the effect of any one image is spread across them. Google researchers put it directly: fully erasing the influence of data means more than deleting it from the databases where it is stored, because the influence also has to be erased from trained models. Retraining a large model without the data is computationally expensive, and methods that avoid retraining, known as machine unlearning, are still being researched.

In its December 2024 opinion, the European Data Protection Board said that whether a model trained on personal data counts as anonymous must be assessed case by case, and that unlawful processing during development could affect the lawfulness of the model's deployment. That is a matter between regulators and companies. It does not give an individual a way to be removed from a model.

Be cautious of any service that says it can delete you from an AI model. What can be shown with evidence is narrower: a source image is gone, a dataset entry was removed, or a company acknowledged an objection.

For new images, the most reliable control is where and how publicly you post them. Glaze, an image-altering tool from the University of Chicago, is built to protect artists against style mimicry, not to hide a person's face, and its authors state that it is not a permanent solution because future techniques may overcome it.