AI Watermarks Are Vanishing: How to Actually Tell if an Image or Video Is AI-Generated
Can you still tell if an image or video is AI-generated now that the labels are coming off? Yes — but usually not just by looking at it. On August 14, 2026, Google began letting users switch off the visible “sparkle” watermark on content made with its Nano Banana (images), Omni (video), and Lyria (music) models. The good news, and the part most of the panic misses: Google was clear that removing the visible mark does not remove the invisible SynthID watermark or the C2PA provenance metadata baked into every generation. In other words, the sticker on the outside is gone, but the fingerprint on the inside stays. Here is how ordinary people can actually check — the free tools first, the eyeball tricks last.
Two kinds of watermark, and only one of them left
The confusion is worth clearing up in one paragraph, because it changes what you should do. AI images from the big players carry two completely different marks. The first is a visible watermark — that little sparkle logo in the corner. It was always cosmetic, easy to crop, and now, on Google’s tools, easy to toggle off entirely. The second is an invisible watermark, woven into the pixels themselves (Google’s is called SynthID), plus a chunk of provenance metadata following an open industry standard called C2PA, or “Content Credentials.” Google, OpenAI and others embed these automatically, and there is no user-facing switch to turn them off. So the headline “you can now remove the AI watermark” is only half true — the meaningful one is still there.
Step 1: Check for Content Credentials (the fastest tell)
C2PA Content Credentials are like a nutrition label for a file: they can record what tool made or edited an image and whether AI was involved, all cryptographically signed so it is hard to fake. When they survive, they are the single clearest answer you can get.
To read them, you don’t need to be technical. Drag the image into a free verifier in your browser — the Content Authenticity Initiative’s Verify tool (verify.contentauthenticity.org) is the reference one, and there are others. If the file still has its credentials, you’ll see the origin spelled out. Google is also rolling C2PA verification directly into the Gemini app, with Search and Chrome to follow, and OpenAI offers its own provenance verification where you upload a file and it reports whether it detects Content Credentials, a SynthID watermark, or no supported signal at all.
The catch: metadata is fragile. Screenshotting an image, re-saving it, or running it through a social platform that strips metadata can wipe the Content Credentials clean. A blank result therefore doesn’t prove “real” — it just means the label didn’t survive the journey. Which is exactly why the next step exists.
Step 2: Test for the invisible SynthID watermark
SynthID is Google’s invisible watermark, and it’s tougher than metadata — it’s embedded in the image content itself, so it can survive cropping, resizing, screenshots, and moderate edits that would destroy a metadata tag. This is the layer that stays even after someone removes the visible sparkle.
For everyday use, the practical route is Google’s own tooling: the Gemini app can check an uploaded image for a SynthID watermark, and Google has a dedicated SynthID Detector portal (currently rolling out via a waitlist). Upload the media, and it tells you whether one of Google’s models likely made it. It isn’t a universal lie detector — SynthID only flags content from tools that use SynthID, so a “no watermark found” result doesn’t clear an image; it just means Google’s models probably weren’t involved. But when it does find one, that’s a strong signal. This invisible-signature approach is the same idea now spreading to AI text, too, which we broke down in our guide to Claude’s text watermarking.
Step 3: Reverse image search the origin
Before you trust any single detector, spend ten seconds on a reverse image search — Google Lens, the “About this image” feature, or TinEye. It won’t run a watermark analysis, but it answers a different and often more useful question: where has this picture been before, and who first posted it? If a “breaking news” photo actually turns up in a three-year-old stock library, or a “real” portrait traces back to an AI-art gallery, you have your answer without any special tooling. Origin and context catch a huge share of fakes that pixel analysis alone would miss.
Step 4: Trust your eyes — but only last
The old visual giveaways still work, and they’re getting rarer, so treat them as a tiebreaker rather than proof. Zoom in and scan for the classic slips: hands with too many or too few fingers, jewelry or teeth that don’t match side to side, background text that dissolves into gibberish, skin and lighting that look airbrushed to an impossible sheen, and edges where hair or fabric melts into the background. In video, watch for blinking that’s slightly off, lips that don’t quite track the audio, and physics that feel a half-beat wrong. The trouble is that top-tier models now clear most of these hurdles, so a clean-looking image proves nothing on its own — which is the whole reason the tool-based checks above come first. If you want a deeper field guide to the low-effort machine-made content flooding social feeds, we covered that in how to spot AI slop.
The realistic bottom line: stack your signals
Here’s the honest truth no detector vendor will lead with: there is no single button that reliably answers “is this AI?” every time. Metadata gets stripped. Invisible watermarks only cover the tools that use them. Visual tells are fading. Any one check can be fooled.
So don’t rely on one. The reliable habit is to stack signals: check for Content Credentials, test for SynthID, reverse-search the origin, and eyeball the details — then weigh the results together. Three quiet hits pointing the same way is far more trustworthy than one loud verdict. And calibrate to the stakes. Deciding whether to chuckle at a meme? A glance is fine. Deciding whether a photo is real evidence in a hiring, legal, insurance, or news context? Treat any single score as a lead to investigate, never a final ruling.
The removable-watermark news sounds like a step backward for honesty online, but it’s really a nudge to stop depending on a logo someone else controls. The durable proof — invisible watermarks and signed provenance — is still there, and the tools to read it are free and getting easier by the month. Being able to verify what you’re looking at used to be a specialist skill. In 2026 it’s just modern literacy: a two-minute habit that keeps you the one who decides what’s real, instead of letting the picture decide for you.
Sources & further reading:
- Google will now allow users to remove visible watermark from its AI generations — TechCrunch
- You can now turn off Google Gemini’s visible watermarks — The Verge
- Tools to understand how content was created and edited — Google blog
- Provenance signals (Content Credentials, SynthID) in OpenAI-generated content — OpenAI Help Center
Related Reading
- Your AI’s Words Now Carry an Invisible Signature: What Claude’s Text Watermarking Means for You
- AI Is Training on Your Data by Default — Here’s How to Turn It Off, Platform by Platform
- The Best AI Video Editing Tools in 2026: Create Professional Videos Without the Learning Curve