Where can I check if something is AI art?
Three real checks, in order of trustworthiness. Upload it to the Gemini app and ask if it's Google AI — SynthID watermarks are reliable when found. Check C2PA Content Credentials via Google's 'About this image' or contentcredentials.org/verify. Reverse image search for history. All three prove presence, never absence: no watermark ≠ not AI.
Why — the first-principles explanation
Every method here splits into two families, and confusing them is why people get bad answers.
Family one is provenance: the image tells you about itself. SynthID is a watermark baked into Google's AI-generated pixels; it survives cropping, filters, and compression. C2PA Content Credentials are a signed manifest attached to the file recording which tool made it and what edits followed. When either is present, it's strong evidence — cryptographically signed, hard to fake. But both are one-directional. Finding a watermark tells you a lot. Not finding one tells you nearly nothing, because SynthID only marks Google's own output, C2PA metadata dies to a screenshot, and roughly 80% of AI images through 2023 came from open-source Stable Diffusion, which anyone runs offline marking nothing. Absence of evidence isn't evidence here — it's the default state of most images on the internet.
Family two is guessing from the pixels — the "AI or Not" style detectors. These are classifiers trained to spot statistical fingerprints, and they're the ones people reach for first and should trust least. They give you a confident percentage with no confidence interval, they're trained on yesterday's generators while today's have moved on, and they have a nasty asymmetry: they tend to flag clean digital art, smooth airbrushed illustration, and stylized work as AI. Which means the people most often falsely accused are human artists whose style happens to look like the training data's idea of AI. A detector saying 92% AI is not a measurement. It's a vibe with a decimal point.
So the honest procedure inverts the usual instinct. Don't start with a detector. Start with history: where did this image first appear, does the artist have progress files, timelapses, sketches, a body of work with a consistent hand? Provenance-by-paper-trail is slow, unglamorous, and far more reliable than any one-click tool — because it's the one signal that's expensive to fake. Google's 'About this image' does exactly this: it shows when an image was first indexed, where it may have first appeared, and where else it's been seen.
Google's own framing is the right one: there is no silver bullet. Combine provenance, watermark checks, source history, reverse image search, and judgment. Any single tool giving you a yes/no is overselling.
An example that makes it click
Think about spotting counterfeit money. There's a real check — hold it to the light and look for the embedded strip. If the strip is there, it's genuine, and that's near-conclusive. But if you can't find a strip, you haven't proved it's fake. Maybe it's an older bill. Maybe it's a foreign note. Maybe it's a photocopy of a real one.
SynthID and Content Credentials are the security strip. Present, they tell you a lot. Absent, they tell you almost nothing — and most images simply never had one. Meanwhile a pixel-guessing detector is the guy at the counter squinting and saying "feels fake to me." Sometimes right. Never evidence. And he squints hardest at the crispest, cleanest bills — which is exactly why real artists with smooth digital styles keep getting accused.
How to do it
- Check for a SynthID watermark: open the Gemini app, upload the image, and ask whether it was generated or edited by Google AI. This is the option available to most people; Google's separate SynthID Detector portal is waitlisted for journalists and researchers.
- Check Content Credentials in Google: find the image in Google Images, click the three dots, and open 'About this image'. If the file carries a C2PA manifest signed by a trusted certificate, it can show whether AI tools created or edited it.
- Check Content Credentials directly: upload the file at contentcredentials.org/verify to inspect any C2PA manifest — the tool used, the creation date, and the edit history.
- Trace the history: use 'About this image' or a reverse image search to see when it was first indexed and where it first appeared. An image with no trace before 2023 and no original publisher is a meaningful hint, not proof.
- Look for the artist's paper trail — progress shots, layered source files, timelapse video, sketches, and a consistent body of work. This is slow and it's the most reliable signal available, because it's the hardest to fake.
- Use pixel-guessing detectors last, and only as a weak hint. Never treat a percentage as a verdict — they systematically misfire on clean, smooth, stylized human digital art.
- Remember the asymmetry before you accuse anyone: finding a watermark is strong evidence of AI. Finding nothing is the normal state of most images and proves nothing at all.
Key facts
- SynthID watermarks only mark content from Google's own AI products; they survive cropping, filters, and lossy compression, but Google publishes no false-positive or detection-failure rates.
- SynthID can be checked by uploading media to the Gemini app and asking whether it was made or edited by Google AI; the dedicated SynthID Detector portal is waitlisted for journalists and media professionals.
- Google joined the C2PA steering committee on September 17, 2024 and shows Content Credentials (C2PA 2.1) in 'About this image' across Google Images, Lens, and Circle to Search — only when the manifest is signed by a certificate on the C2PA Trust List.
- C2PA metadata is removed by screenshots, many re-uploads, and most editors, while the image looks unchanged.
- An estimated 80% of AI images through August 2023 — about 12.59 billion of 15 billion — came from open-source Stable Diffusion, which can run offline with no watermark or credentials (Everypixel Journal).
- 'About this image' launched in the U.S. in English in May 2023 and reports when an image was first indexed by Google and where else it has been seen online.
- Google states there is 'no silver bullet solution for all content online' for content provenance.
▶ The 60-second explainer (script)
Where can you check if something is AI art? Three real options, and one big trap. Best first: upload it to the Gemini app and ask if it was made or edited by Google AI. Google embeds a watermark called SynthID in its AI output, and it survives cropping, filters, and compression. Second: check Content Credentials — find the image in Google Images, click the three dots, About This Image. Or upload the file at contentcredentials dot org slash verify. That's a signed record of which tool made it and what edits followed. Third: reverse image search the history. Now the trap. Every one of those checks is one-directional. Finding a watermark tells you a lot. Not finding one tells you almost nothing. SynthID only marks Google's own AI. Content Credentials die to a screenshot. And roughly eighty percent of AI images came from open-source Stable Diffusion, which anyone runs offline, marking nothing at all. No watermark is the normal state of most images on the internet. So what about those AI detector sites that give you ninety-two percent AI? Trust those least. They're trained on yesterday's generators, they publish no confidence intervals, and they have a nasty habit: they flag clean, smooth, stylized digital art as AI. Meaning the people most often falsely accused are human artists whose style resembles the machine's. Ninety-two percent isn't a measurement. It's a vibe with a decimal point. The real check is boring. Where did the image first appear? Does the artist have sketches, layers, timelapses, a consistent hand across years? That's the signal that's expensive to fake. Google's own words: there's no silver bullet.
What authoritative sources say
People also ask
Are sites like 'AI or Not' accurate?
They're a weak hint, not evidence. They're trained on older generators, publish no error rates on current ones, and disproportionately misfire on clean, smooth human digital art — the exact case where a false accusation does the most damage.
If there's no SynthID watermark, is it human-made?
No. SynthID only marks Google's own AI output. Midjourney, Stable Diffusion, and everything else leave no SynthID at all, so absence is the normal state, not a clearance.
Can watermarks and credentials be removed?
C2PA metadata is trivially stripped — a screenshot does it. SynthID is embedded in the pixels and much tougher, surviving cropping and compression, though Google publishes no failure rates.
What are the old visual tells — hands, text, jewelry?
Increasingly unreliable. Current models handle hands and lettering far better than 2023 models, so the classic giveaways now mostly catch old or careless output while missing recent work.
What's the strongest evidence something is human-made?
A paper trail: layered source files, progress shots, timelapse recordings, and a consistent body of work over years. It's slow to check and expensive to fake, which is exactly why it's the best signal.