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

Here's a number that should stop you mid-scroll: only 38% of AI image generators have watermarking that actually meets the upcoming legal standard. That means roughly six out of ten AI tools pumping out synthetic photos, audio clips, and videos right now are doing it with no reliable way for anyone — human or machine — to later prove what they're looking at is fake. And on August 2, 2026, that becomes illegal in the EU, with fines up to €15 million per violation.

But here's the twist: the watermark you're imagining right now — the semi-transparent logo in the corner of a Getty image — is almost nothing like what regulators are actually demanding. The gap between what most people picture and what the law requires is enormous. And closing that gap is, it turns out, one of the hardest engineering problems in AI right now.

First, Forget Everything You Think "Watermark" Means

When most people hear "watermark," they picture exactly one thing: a visible stamp. A logo. A little translucent text floating over a photo that says "DRAFT" or shows a company name. You can see it. You can crop it out. You can screenshot around it.

That's not what we're talking about. Not even close.

The watermarks the EU AI Act requires are imperceptible to human eyes. They live inside the actual pixel data of an image — or inside the waveform of an audio file, or woven through the frame data of a video. You cannot see them. You cannot hear them. A normal screenshot doesn't remove them. Uploading to social media doesn't strip them. Even resaving the file often leaves them intact.

Think of it this way. A visible watermark is like writing your name on a piece of paper. An imperceptible watermark is like encoding your DNA into the paper's fiber. One you can erase with a pen. The other survives being crumpled, photocopied, and rained on — and only a lab test reveals it's there at all.

This is why the "watermark = visible logo" mental model is so understandable, and so wrong. Visible watermarks dominated early AI tools. When ChatGPT's image features first launched, many outputs had visible "Created with AI" labels or logos. That's labeling — which is a separate thing entirely. Labeling gives a human reader a notice they can see. Watermarking creates a signal that a machine can detect later, even if the label has been removed. The EU's rules require both. Most vendors are still only building one. This article is part of a series — start with Blocked By A Bot Europe Just Gave You The Right To Demand An.

What the Law Actually Demands — and Why It's So Hard

Article 50(2) of the EU AI Act sets out four requirements for AI content watermarks. They must be: machine-readable, effective, interoperable (meaning different systems can read each other's signals — like how any DVD player can read any DVD), and robust — meaning the watermark must survive the kinds of everyday abuse content goes through online.

That last word, "robust," is where things get genuinely tricky.

Here's the engineering tension nobody explains: making a watermark harder to remove usually means making bigger, more disruptive changes to the file. Embed a stronger signal into an audio clip, and you might introduce a faint hiss. Encode a more durable pattern into an image's pixel structure, and sharp-eyed viewers might notice something slightly off. The more survivable the watermark, the more it risks degrading the content quality. There's no free lunch here — it's a constant tradeoff that engineers are still actively solving.

And survivability matters enormously in practice. Think about what happens to an image between creation and consumption. It gets compressed when uploaded to Instagram. It gets re-saved when someone downloads it. It gets screenshotted, cropped, filtered, and re-uploaded. A watermark that only survives one of those steps is nearly useless as evidence of anything.

"Watermarking and labeling serve different purposes: watermarking creates a machine-readable signal that systems can detect later, while labeling gives people a visible notice that content is AI-generated, so both controls may be needed in public workflows." — Resemble AI, Complete Guide to EU AI Act Watermarking Requirements

This is exactly why single-layer solutions keep failing. Early implementations tried one approach: embed metadata (structured data about the file's origin, stored alongside it) into the file header. Simple, clean, easy to read. Also easy to destroy — one screenshot, one format conversion, and it's gone. Metadata is the first thing that gets stripped when a file travels across platforms.

---

The Three-Layer Fix Regulators Now Require

The EU's approach — and the one now being adopted across the industry — is a defense-in-depth strategy. Three layers working together, because no single method survives everything.

Layer one: C2PA metadata. C2PA stands for Coalition for Content Provenance and Authenticity (basically, an industry-wide standard for tracking where a piece of content came from and how it was made). This is the structured data layer — think of it like a digital birth certificate attached to the file, recording which AI model created it, when, and how. It's readable by machines, searchable, and detailed. It's also the most fragile layer. A screenshot removes it instantly. Previously in this series: Nervous On A Bank Call An Ai Just Judged You And Its Probabl.

Layer two: imperceptible watermarking. This is the signal embedded directly into the pixel data of an image, the audio waveform of a sound file, or the frame structure of a video. It's invisible to humans. A 2025 industry audit found this type of watermarking present in only 8 out of the AI systems reviewed — which tells you exactly how far behind the industry is. This layer is harder to build and harder to remove.

Layer three: logging. A centralized record kept by the AI system's creator, documenting what was generated, when, and with what parameters. Even if the file itself is scrubbed clean of every other signal, a proper log means there's a paper trail somewhere — assuming the company that made the AI tool is cooperative and still exists.

No single layer is sufficient. The C2PA metadata survives unless someone screenshots the image. The imperceptible watermark survives screenshots but might not survive heavy compression. The log survives everything — but requires cooperation from the original platform. Together, they create overlapping coverage that's much harder to defeat than any one method alone.

Wire services figured this out before regulators caught up. The Associated Press, Reuters, AFP, and the New York Times now require signed Content Credentials — a form of C2PA-compliant provenance data — on all wire images of major news events. That's not a legal requirement for them yet. They did it because their credibility depends on it. The regulatory world is now catching up to what journalism already understood: provenance has to be built into the file, not just claimed verbally.

---

What This Actually Changes for the Rest of Us

Here's the practical shift — and it's a big one. For years, the advice around deepfakes and AI-generated content was essentially: look harder. Check the fingers (AI used to be terrible at hands). Look at the teeth. Watch for unnatural blinking. Squint at the background.

That advice isn't wrong. But it's losing. AI-generated content has gotten good enough that visual inspection is genuinely unreliable for most people — and even for most experts. A 2025 position paper on arXiv makes a pointed argument: "Watermarking Without Standards Is Not AI Governance." The warning embedded in that title is that without agreed-upon, interoperable standards for what a watermark looks like and how to verify it, the technology becomes theater — a signal nobody can reliably read.

What watermarking standards shift the question from is: "Can your eyes catch the fake?" The new question becomes: "Does this file carry verifiable proof of how it was made?" That's a question a machine can answer far more reliably than a human. Up next: Liveness Detection Selfie Id Verification Explained.

At CaraComp, where the work centers on comparing faces with precision — measuring distances between features, checking consistency across images — the move toward file-level provenance is a natural extension of the same problem. Visual comparison catches a lot. But verifying what the file itself says about its own origin? That's a different, complementary layer of certainty. The future of content trust probably needs both: the visual analysis and the embedded signal working together.

The U.S. National Institute of Standards and Technology (NIST) — the government body that sets measurement and technology standards — published guidance on exactly this: evaluating synthetic content requires checking both visual indicators and provenance data embedded in the file. Not one or the other. Both.

So what should you actually take away from all this? Not anxiety. Clarity.

"Just look carefully" was always a temporary answer. The real solution was always going to be systematic — baked into the files, verifiable by machines, standardized across platforms. The EU's deadline is forcing that reckoning into the open. The 62% of AI tools that haven't built this yet don't have the luxury of waiting much longer.

And here's the question worth sitting with: if a photo or video arrived with a verified "AI-generated" badge — not a logo you could crop out, but a cryptographic signal (basically a tamper-proof digital signature) that any platform could independently confirm — would you trust it more? Or would you still want someone to check the file itself?

Because that instinct — don't just take the label's word for it, check the underlying signal — is exactly the right one. And for the first time, the technology to actually do that is being standardized and required by law. The label is just the beginning. The proof is in the pixels.