AI Image Watermarks in 2026: Every Visible Badge and Invisible Signal, Explained

You generate an image you are proud of, post it, and within seconds a small grey label appears underneath: AI info. You never told anyone how it was made. Something in the file did.

That "something" is almost never what people assume. There is a widespread belief that AI images carry a single watermark, and that one clever trick makes it disappear. In reality there are three separate layers of marking, they behave nothing alike, and confusing them is why most advice on this subject is useless.

This is the pillar guide for the whole topic. By the end you will know exactly which signals your image carries, which of them you can genuinely control, and which ones no honest tool can promise to erase.

The three things people call "the AI watermark"

Everything else in this guide hangs off that table. A visible badge is a geometry problem. Metadata is a container problem. An invisible watermark is a signal-processing problem — and it is the only one of the three that is genuinely hard.

Which generators mark what

The landscape shifts constantly, and providers rarely document their marking in full. This is the practical shape of it as of 2026 — treat it as a map of what to look for, then verify against your own files.

Two patterns matter more than any individual row. First, metadata is nearly universal — it is the cheapest signal for a provider to add and the one most likely to be present in your file right now. Second, the invisible column is the honest unknown, because providers deliberately do not publish enough detail to make their watermarks easy to attack.

Why generators watermark at all

It is tempting to read watermarking as branding. It is mostly compliance.

• Voluntary industry commitments. Major providers signed up to marking synthetic media as detectable, well before any law required it.
• The C2PA standard. An industry coalition built a cryptographically signed manifest format for recording how a file was made. Adopting it is how a provider proves it is participating in provenance.
• Regulation. Transparency obligations under the EU AI Act and disclosure requirements in laws like California's AI transparency act expect synthetic content to be identifiable by machine. Watermarking and metadata are how providers satisfy that.

We cover the rules themselves in detail in the AI image labelling laws and agreements spoke, because the legal layer is what makes all of this permanent rather than a passing product fashion.

What actually survives an edit

This is the part most guides get wrong, and it is the single most useful thing to internalise.

Read the columns, not the rows. Metadata dies to almost anything. A visible badge dies only to geometry. An invisible watermark dies to almost nothing you would willingly do to your own image.

The visible vs invisible watermarks spoke goes deeper into why that durability gap exists, with the engineering reasoning behind it.

How platforms turn signals into labels

Here is the sequence that puts a label under your post. Platforms do not run one check — they run a cascade, and the first hit wins.

1. 1Is there a signed C2PA manifest saying this was AI-generated?
2. 2Is there an IPTC digital-source-type field declaring synthetic origin?
3. 3Are there generator tags in EXIF or XMP naming a model?
4. 4Does an invisible watermark detector fire?
5. 5Does a visual classifier think it looks generated?
6. 6Did the uploader self-declare it?

That cascade is why partial cleanup fails so often. You can strip every byte of metadata and still get labelled at step four or five. You can crop the badge and still get labelled at step one. The how Facebook knows your image is AI spoke walks the whole chain with what each step can and cannot see.

What you can actually control today

Two of the three layers are genuinely, completely in your hands — and both can be handled in your browser without uploading the file anywhere.

The metadata layer. Decoding an image to raw pixels and re-encoding it produces a new file with no container data at all. No C2PA manifest, no EXIF, no XMP, no IPTC, no PNG text chunks. This is not a best-effort scrub; the old container simply does not exist in the output.

The visible badge. A corner mark or edge strip is pixels in a known region. Cropping removes it outright; content-aware patching fills it from surrounding detail when you need to keep the full frame.

And the honest third: a robust invisible watermark cannot be guaranteed away. Our tools offer optional resampling and dithering that can degrade fragile schemes, and we tell you plainly that it is best-effort and costs image quality. If you see a competitor promising a guarantee here, they are describing something the underlying technology does not permit.

Check before you post, not after

The most avoidable outcome in this entire topic is discovering what your file contained *after* it is public and labelled. Inspecting it takes seconds: a before-and-after report lists every provenance signal found, what was removed, and an estimated detection likelihood for the result.

The checking your image for AI watermarks spoke walks through reading that report field by field.

Where to go next

This pillar is the map. The spokes below are the territory — each one takes a single generator, platform or rule and covers it properly. If you arrived here searching for a specific tool's watermark, start with its spoke; if you arrived wanting to understand the topic, read the visible-vs-invisible spoke next.

Read: how Facebook and Instagram decide your image is AI →