That "Made by AI" Label? It's Hiding Something You Can't See

When you see a photo online that says "made by A.I.," you probably picture a little label in the corner. But the real marker might be hidden inside the pixels themselves — invisible to you, readable only by a machine. And right now, most A.I. tools aren't even adding it.

A study from twenty twenty-five found that only

A study from twenty twenty-five found that only about a third of A.I. image generators are marking their content properly. That means nearly two out of three tools pumping out fake images have no reliable way to prove where those images came from. If you've ever wondered whether a picture in your feed was real or generated, this is exactly the problem the world is trying to solve. And there's a deadline. A big European law kicks in on August second, twenty twenty-six, and it's going to force a scramble. So why is something that sounds this simple turning into an engineering nightmare?

Let's start with what the law actually asks for. It's called Article Fifty of the E.U. A.I. Act. The rule says A.I.-generated content must carry a signal that a machine can read — one that's reliable, hard to fake, and works across different systems. Sounds reasonable, right? The catch is buried in the details. This article is part of a series — start with Blocked By A Bot Europe Just Gave You The Right To Demand An.

The tension is this. To make a watermark tough to remove, you have to change the content more aggressively. But the more you change it, the more likely a human will notice the distortion. So engineers are stuck. A watermark strong enough to survive gets ugly. A watermark that looks clean gets stripped away easily. For the rest of us, that means the "invisible proof" on your favorite app might vanish the moment you screenshot it.

That's exactly what happens

And that's exactly what happens. Early tools leaned on something called metadata — basically a hidden tag describing the file. But metadata is fragile. Screenshot the image, upload it to social media, or convert the file, and that tag disappears. One layer just isn't enough.

So the E.U. now recommends stacking three defenses together. First, provenance data using a standard called C2PA, which records the file's origin. Second, an imperceptible watermark baked into the image itself. Third, a central log that keeps a record of what was made. Think of it like the difference between a visible logo stamped on a photo and a barcode woven into the pixels. You can crop out a logo. But that hidden barcode survives cropping, compressing, even re-uploading — and only a scanner can find it. Previously in this series: Eu Ai Act Watermarking Requirements Generative Ai Explained.

Here's where most people get it wrong. When you hear "watermark," you picture a stamp or a copyright notice. That's fair — early A.I. tools trained us to think that way. But a visible label and a machine-readable watermark are two different things. The label tells you a human warning. The watermark tells a computer the truth. The E.U. wants both. Most companies have built only one. That's why so many vendors aren't ready — it's not laziness, it's genuinely hard.