Skip to content
Dataset By Humans
Licensing & compliance

Photo takedown requests for AI datasets: how they work

What to do if you find yourself or your property in an AI training dataset, what a good takedown process looks like, and what happens to models already trained.

By Dataset By HumansPublished 2 min read

In this article

Short answer: if you find a photo of yourself, your home or your property in an AI training dataset, contact the dataset provider with the dataset name, the image ID or link, and why it identifies you or infringes your rights. A responsible provider confirms receipt within days, removes the image from the dataset and future deliveries, publishes an updated version and tells its customers to switch to it. Models already trained on the image are usually not retrained, but the image stops being used from then on.

How to find what to report

Most dataset providers publish sample images with an ID or a page per image. Note:

  • the dataset name and version
  • the image ID, file name or page URL
  • a screenshot, if the page might change

What to include in your request

  1. Where you saw the image (dataset name, image ID or link)
  2. Why it concerns you: you can be identified, it shows your private property, or you own the photo
  3. How the provider can contact you

You should not need to send identity documents for most requests. A provider may ask for limited information to confirm the request is yours, but no more than necessary.

What a good takedown process looks like

StepWhat to expect
AcknowledgementWithin a few business days
ReviewThe provider checks the image and your reason
RemovalThe image leaves the catalog and all future deliveries
New versionAn updated dataset version is published
Customer noticeLicensees are told to replace the old version
ConfirmationYou are told when it is done, typically within a month

In the EU and UK, the GDPR gives you the right to request erasure of personal data, and you can complain to your data protection authority if a provider does not respond.

What happens to models already trained?

Removing one image from a trained model is not practical today; research on "machine unlearning" is still early. What matters is that the image is not used again: good licenses require customers to stop using the removed image and train future models on the updated version.

How to tell if a dataset provider takes this seriously

  • A public takedown page with a named contact address
  • Faces and license plates blurred before images are published
  • GPS location removed from photo metadata
  • A license that obliges customers to apply takedowns
  • A datasheet that documents consent and privacy processing

Our process

We blur faces and plates, remove GPS data and review every image by hand before it is published. If you still recognise yourself or your property, follow the steps on our photo takedown page; we acknowledge requests within 5 business days and confirm removal within one month. You can read more about how we collect and clean every photo.

photo takedownimage removal requestprivacyai datasets