Is this image AI-generated?
No detector can tell you reliably, and counting fingers no longer works. The strongest evidence is provenance, not inspection: check the image's Content Credentials (C2PA), the open standard backed by Adobe, Google, Meta, Microsoft, OpenAI, the BBC, and Sony. Then reverse-image search for an original. Visual artifacts prove editing, not fabrication.
Why — the first-principles explanation
There are two ways to answer this question, and almost everyone reaches for the weaker one.
Detection means looking at the pixels and guessing. It's losing, structurally. Image generators are trained to produce images statistically indistinguishable from real photographs — that's literally the objective they're optimized against. Every detector that learns to spot today's generator becomes training feedback for tomorrow's. It's a treadmill where the detector must be retrained constantly and the generator only has to improve once. Worse, real-world images get compressed, screenshotted, and re-uploaded, and each pass strips the faint statistical signal detectors depend on. A screenshot of a real photo can look, to a detector, a lot like a generated one. Then there's the base rate problem, which is where confident people go wrong. Suppose a detector is 95% accurate and you scan a feed where 5% of images are AI. Out of 1,000 images: 50 are AI and it catches ~48; 950 are real and it wrongly flags ~48. So of ~96 images flagged, about half are innocent. A 95% accurate detector produced a coin flip. This is why a detector score is not evidence about your specific image — and why schools and platforms accusing people on detector output are making a math error, not just a policy choice.
Provenance flips the problem. Instead of guessing after the fact, you attach signed information at creation. That's what C2PA — the Coalition for Content Provenance and Authenticity — standardizes. Content Credentials work, in the coalition's own words, like a nutrition label for digital content, showing the content's history to anyone who wants to look. The steering committee is unusually broad: Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, and Truepic. Cameras can sign at capture; generators can mark their output. Guessing is replaced by checking a signature.
The catch is honest: absent credentials prove nothing, since metadata is easily stripped and most images never had it. Provenance is strong positive evidence and weak negative evidence. And give up on the finger-counting folklore. The Will Smith concert video is the cautionary tale — real crowds, at real festivals, ran through an image-to-video model and then YouTube's sharpening filter, and came out with melted faces and garbled signs. The internet counted the artifacts and concluded the audience was fake. Artifacts tell you a model touched the pixels. They don't tell you a model invented the subject.
An example that makes it click
Think about how you'd verify a $20 bill. You could squint at it — does the ink look right, is the paper too smooth? That's detection, and a good counterfeiter will beat your eyes every time, because beating your eyes is the entire job description.
Or you could tilt it and look for the security thread and watermark the Treasury embedded when it was printed. That's provenance. You're not judging quality; you're checking for a mark that's hard to fake and was applied at the source. Content Credentials are the security thread for digital images. And note the same limit applies: a bill without a visible thread might be counterfeit, or might just be old and worn. Missing the mark isn't proof of forgery.
How to do it
- Check Content Credentials first. Use the C2PA/Content Credentials inspector to see if the file carries signed provenance data naming the camera or the generator.
- Reverse-image search it. Finding an earlier, higher-quality original — with a photographer, a date, a location — beats any detector result.
- Find the source account and check whether it existed before the image and posts other verifiable material.
- Look for corroboration: if a real event happened, other people photographed it from other angles. A single-source image of a dramatic event is the actual red flag.
- Treat any AI-detector percentage as a weak hint, never as evidence — the base rate problem makes most positive flags unreliable on realistic feeds.
- Remember artifacts cut both ways: compression, upscaling, and filters create AI-looking distortion in genuine photos.
Key facts
- C2PA (Coalition for Content Provenance and Authenticity) provides an open technical standard for establishing the origin and edits of digital content; its specification is at version 2.3 as of 2026-07.
- The C2PA steering committee includes Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, and Truepic.
- C2PA describes Content Credentials as working "like a nutrition label for digital content," showing content history available to anyone, at any time.
- The base rate problem means a 95%-accurate detector scanning a feed that is 5% AI-generated produces roughly as many false positives as true positives — about half of all flags would be wrong.
- The August 2025 Will Smith concert video demonstrated that real crowds processed through image-to-video AI and YouTube's Shorts unblur/denoise experiment produced classic AI artifacts — melted faces, warped sign text — despite the people being genuinely real.
- Provenance metadata can be stripped from files, so absent Content Credentials is not evidence that an image is AI-generated.
▶ The 60-second explainer (script)
Is this image AI-generated? Here's the uncomfortable answer: no detector can reliably tell you, and counting fingers stopped working a while ago. Let me explain why, because the why is what protects you. Detection means looking at pixels and guessing. It's structurally losing. Image generators are trained to produce images statistically indistinguishable from real photographs — that's literally their objective. Every detector that learns to catch today's generator becomes training feedback for tomorrow's. The detector has to be retrained forever; the generator only has to win once. Now the part that trips up smart people: the base rate. Say a detector is ninety-five percent accurate, and you scan a feed where five percent of images are AI. Out of a thousand images, fifty are AI and it catches about forty-eight. But nine hundred fifty are real, and it falsely flags about forty-eight of those too. So of the ninety-six images flagged, half are innocent. A ninety-five percent accurate detector just gave you a coin flip. That's why accusing someone based on a detector score is a math error. So flip the problem. Instead of guessing afterward, check what was signed at creation. That's C2PA — Content Credentials. The coalition calls it a nutrition label for digital content, and the steering committee is Adobe, Amazon, the BBC, Google, Meta, Microsoft, OpenAI, Sony, and Truepic. Cameras can sign at capture. Generators can mark output. You check a signature instead of squinting. One honest limit: missing credentials prove nothing. Metadata gets stripped, and most images never had it. Provenance is strong positive evidence, weak negative evidence. And the folklore? Remember Will Smith's concert video. Real crowds, real festivals — pushed through an AI animator and YouTube's sharpening filter, and out came melted faces and garbled signs. Everyone counted artifacts and concluded the audience was fake. It wasn't. Artifacts tell you a model touched the pixels. Not that a model invented the subject.
What authoritative sources say
People also ask
Are AI image detectors accurate?
Not reliably. Generators are explicitly optimized to be statistically indistinguishable from real photos, and compression or screenshots strip the signals detectors use. Even a genuinely 95%-accurate detector produces about half false positives on a feed that's 5% AI.
Does counting fingers still work?
No. Current generators mostly render hands correctly, and — more importantly — real photos pushed through upscaling or filters can develop the same distortions. The Will Smith concert video is the classic case of real people looking generated.
What are Content Credentials?
Signed provenance data attached to an image at creation, standardized by C2PA. They act like a nutrition label showing the file's origin and edit history, and can be checked rather than guessed at.
If an image has no Content Credentials, is it fake?
No. Metadata is easily stripped by platforms and re-uploads, and most images never carried it. Credentials are strong evidence when present and prove nothing when absent.
What's the single best way to check an image?
Reverse-image search for an earlier, higher-quality original with a known photographer and date. Provenance beat detection in the Will Smith case, and it usually will in yours.
The same question, asked other ways
- How can you tell if a picture is AI-generated?1,300/mo
- How can you tell if an image is AI-generated?1,000/mo
- How can you tell if a photo is AI-generated?1,000/mo
- Is this photo AI generated?880/mo
- How to tell if an image is AI generated?720/mo