Every Photo You Post Is Training Someone's AI — Here's How

Let me ask you something.

When was the last time you posted a photo on Instagram, Facebook, or X without reading the privacy policy first?

Never. Nobody does.

That is exactly what AI companies are counting on.

Right now, billions of photos posted publicly on the internet are being scraped, processed, and used to train AI image generation models — without the knowledge or consent of the people in them. Your face. Your kids. Your home. Your location. All of it is potentially being used to teach a machine what humans look like.

Here is what is actually happening, what it means for you, and what you can do about it.

Where Does AI Training Data Actually Come From?

To understand the problem, you need to understand one thing: AI image models need enormous amounts of data to learn from.

Midjourney, Stable Diffusion, DALL-E — all of these tools were trained on hundreds of millions or billions of images scraped from the public internet.

One of the most widely used training datasets is LAION-5B — a collection of over 5 billion image-text pairs assembled from web scrapes conducted between 2014 and 2022. As reported by MIT Technology Review, researchers found that LAION-5B contains millions of examples of personally identifiable information.

Here is the part that should give you pause.

Most of those images were scraped before ChatGPT even existed. People who posted photos publicly in 2015 or 2018 could not have consented to having those photos used to train AI models — because those models did not exist yet.

What Happens When Your Photo Gets Scraped?

This is not just about your image appearing somewhere. The process goes much deeper than that.

When AI systems process your photo, they extract what researchers call a faceprint — a mathematical representation of your facial geometry expressed as 128 to 512 floating-point numbers. As detailed by FaceSift, this faceprint is more durable than a photograph.

It cannot be unseen. It survives cropping, filtering, and low resolution. And critically — it can be matched against new photos of you taken years later.

Deleting your original photo does not delete the faceprint. The mathematical embedding was computed and stored before you pressed delete. In many cases it has already been incorporated into model weights that are impossible to reverse.

This is the technical reality behind the phrase "once it's out there, it's out there."

The Clearview AI Case — And What It Shows

The most documented example of AI training on scraped photos is Clearview AI.

Clearview scraped over 30 billion images from the public web and built a facial recognition database sold to law enforcement agencies. As noted by Proton, Clearview has since faced lawsuits, fines, and legal warnings for violating privacy across multiple countries.

But Clearview is not unique. It is just the most visible.

LAION-5B — the dataset that trained Stable Diffusion — included images of children scraped from personal blogs and social media without consent. Human Rights Watch found over 170 images of Brazilian children in that dataset used without permission.

Facebook's DeepFace facial recognition system was trained largely on user-generated photo tags. Every time you tagged a friend in a photo, you were contributing a labelled training sample to one of the most powerful facial recognition systems in the world.

What Happens When You Tag Someone?

Here is something most people do not know.

When you tag someone in a photo on a major platform, you are not just identifying them to your friends. You are contributing a labelled training sample — this face belongs to this named person — to systems that use exactly that kind of data to build and improve facial recognition.

As noted by FaceSift, Facebook's DeepFace system achieves over 97 percent accuracy in identifying faces — higher than the FBI's technology at 85 percent. It was built largely on publicly available photos and user-generated tags.

You helped build it. Everyone who has ever tagged a photo on Facebook did.

The Opt-Out Problem — Why "You Can Turn It Off" Is Not Good Enough

Most major platforms now offer some form of opt-out for AI training. The problem is how those opt-outs work.

The default is always opt-in.

As documented by The Slow AI, every major platform follows the same pattern: the setting that lets them train on your data is switched on when you arrive, and switched off only if you go find it somewhere in the settings. Almost nobody changes defaults — the companies know this — and so "we offer an opt-out" becomes a mechanism for taking your data while appearing to ask for consent.

Real consent is a yes. What most users have given is the absence of a no.

TikTok offers no clean opt-out toggle at all. A 2026 study ranked it the hardest major platform to leave, and noted it analyzes even unpublished drafts.

And opt-outs are not retroactive. If your photos were scraped into a model's training data last year, turning off the setting today does not remove them from the model that has already been trained.

What About the Legal Landscape?

The law is catching up — but slowly.

The EU AI Act, now partially in force, places AI systems that use biometric data into the high-risk category, requiring transparency, conformity assessments, and explicit consent in many contexts. As analyzed by Termly, companies training AI on publicly available personal data under GDPR must have a lawful basis, notify data subjects that their data has been used, and honor deletion requests — requirements that most AI companies have not consistently met.

In the United States, regulation has moved more slowly. Several states have passed biometric privacy laws — Illinois's BIPA is the strongest — but there is no federal standard.

As noted by Medium's analysis, a string of 2025 and 2026 settlements and court decisions has started to push the industry away from the earlier "scrape now, deal with consequences later" approach. The legal landscape is shifting, but the direction of travel matters more than any single ruling, and the existing training data is not coming back.

What Can You Actually Do?

Here are five things that make a real difference.

Check if your photos are already in AI training datasets. Use HaveIBeenTrained.com to search for your images in datasets like LAION-5B. If found, you can request removal — but as noted by Proton, this does not remove your photos from models that have already been trained on them.

Turn off AI training settings on every platform you use. On ChatGPT: Settings, Data Controls, turn off "Improve the model for everyone." On Claude: Settings, Privacy, at claude.ai/settings/data-privacy-controls. On X: untick the Grok training box under Privacy and Safety. These settings do not undo past data collection but prevent future use.

Stop tagging people's faces in photos. Every tag you add is a labelled data point. This applies to your own photos and to anyone else's.

Use Glaze or Nightshade before posting creative work. Developed by researchers at the University of Chicago, these tools subtly alter images at the pixel level to confuse AI training systems, while the images look unchanged to human eyes. They are specifically designed for artists and creators but can be used by anyone.

Think before you post location-specific or child photos publicly. As noted by Proton, AI trained on your photos can potentially recognize what you look like, how you dress, and where you live. Location tags and identifiable backgrounds add data that compounds the privacy exposure.

The Uncomfortable Truth

Here is what nobody in the tech industry wants to say directly.

The photos that are already in training datasets are not coming back. There is no verified way to un-train a model on specific images. Once a photo has shaped model weights, those weights encode something about your face that persists regardless of what happens to the original image.

The legal frameworks being developed now — the GDPR requirements, the EU AI Act, the state biometric laws — are important. But they govern what happens from here, not what has already happened.

According to Wikipedia on data privacy, the right to informational privacy encompasses the ability to control information about oneself — including how it is collected, processed, and used. The gap between that right as a principle and the current reality of how AI systems have been trained is one of the most significant unresolved tensions in technology policy.

The question "is my photo training AI?" is no longer hypothetical for millions of people. The more useful question — the one that actually gives you some control — is what you post next, and under what terms.

Those terms are worth reading, even once.