Somewhere around 34 million AI-generated images get produced every single day at this point, and the advice for spotting them that was floating around a couple of years ago count the fingers, check for garbled text, look for that weirdly waxy skin mostly doesn’t work anymore. The models fixed most of that. Which is a little unsettling if you’re still relying on it.
Here’s the good news, though: AI images still aren’t perfect, they’ve just gotten better at hiding the specific mistakes everyone learned to look for. The tells have shifted, not disappeared. And there’s a second, more reliable path that has nothing to do with squinting at pixels at all.
The Short Version
The most reliable way to spot an AI-generated image combines two things: checking for provenance (does the image carry Content Credentials or a watermark like SynthID showing how it was made?), and inspecting the specific visual tells that still hold up background text, hands under complex poses, mismatched eye reflections, and shadow direction.
If neither is conclusive, treat the image the way you’d treat any unverified claim: trace it back to a real, named, accountable source before trusting it.
Why the Old Advice Stopped Working?
If you’ve been mentally relying on “AI hands always have six fingers” or “the text always looks scrambled,” it’s worth updating that instinct. Foreground text on signs and product labels has improved dramatically, and current-generation models produce convincing hands in standard, relaxed poses.
This isn’t because detection got harder in some abstract sense it’s because the obvious rendering glitches from a couple of years ago were exactly the kind of thing model developers could see and specifically train against.
The mistakes that are left are baked in at a deeper level, tied to how these models actually generate images rather than surface-level polish, which makes them harder to fully eliminate.
The Visual Tells That Still Hold Up
Background and secondary text. This is genuinely the fastest tell that still works. While foreground signage has gotten much better, text sitting further back a distant storefront, a license plate, a book spine on a shelf, small print on packaging is still where models routinely fail. Diffusion models simply allocate less attention to background detail during generation, since it contributes less to how “good” the overall image looks. Zoom into anything small and secondary before trusting it.
Hands under stress or complex poses. Simple, relaxed hands look convincing now, but the moment fingers overlap, grip an object at an odd angle, or interact with jewellery, things tend to fall apart. Check whether a ring sits naturally on a knuckle, whether a grip makes physical sense, and whether finger count and joint angles hold up under a closer look.
Eyes and reflections. In a real photograph, the catchlight reflections in both eyes should match same light source, same shape, same position. AI-generated eyes often have mismatched or slightly invented reflections between the two eyes, and pupils that are unnaturally perfect and identically round.
Shadows and reflections in the scene. Check whether a shadow’s direction actually matches the apparent light source, and whether reflections in glass, water, or metal show a scene consistent with what’s actually there. Generative models are good at making individual elements look right in isolation, but full scene-wide physical consistency is still a weak spot.
Repeated patterns and background crowds. In images with a crowd or repeating elements a row of windows, a group of people look for details that repeat a little too perfectly, or background figures whose faces and poses feel oddly similar to each other.
No single one of these is proof on its own. The combination is what actually holds up: zoom into background text, check hands under any real pose complexity, look at eye reflections, and check shadow direction together, this catches the large majority of AI images still being passed off as real.
Checking Provenance: The More Reliable Path
Visual inspection has a ceiling, and it’s dropping every time a new model ships. Provenance checking doesn’t have that problem, because it’s based on how the image was made, not what it looks like.
C2PA Content Credentials is an open standard, backed by Adobe, Microsoft, Google, OpenAI, and camera makers including Sony and Nikon, that attaches a cryptographic record of how an image was created and edited. You can check any image directly at contentcredentials.org if credentials are present, that’s strong evidence of the image’s origin. The honest caveat: most images circulating online still don’t carry a credential yet, so a missing one isn’t proof an image is fake, just that it’s unverified through this specific method.
SynthID and similar watermarking work a little differently some generators, including several of Google’s tools, embed an invisible watermark that only their own detection tools can reliably read. Same rule applies here: presence is meaningful, absence tells you nothing either way.
This area is also moving fast at a regulatory level. Google announced at its May 2026 I/O event that Chrome and Search would begin surfacing SynthID and C2PA signals directly to users, and the EU’s AI Act now legally requires visible labeling of AI-generated content across Europe starting August 2, 2026. Expect provenance checking to become a lot more mainstream and built-in over the next year rather than something you have to seek out manually.
Should You Trust an Automated AI Detector Tool?
Be skeptical. A 2026 benchmark testing 23 different detection tools found they perform reasonably well around 75% accuracy against older, 2020–2021-era generators. Against current commercial models, that accuracy drops to somewhere between 18% and 30%, which is close to a coin flip and, in some cases, worse.
This matters practically: the generators are improving faster than the detectors chasing them, and a detector’s confidence score is not a verdict, no matter how official the percentage looks on screen. If you’re using one of these tools, treat the result as one weak input among several, not the final answer and definitely not grounds to publicly accuse someone of faking something, especially in contexts like academic work where a false accusation carries real consequences.
If You’re Trying to Verify a Specific Viral Image
This is where most of the visual-tells advice above becomes less important than a much older skill: source verification. If a dramatic, emotionally charged image is circulating and you’re trying to figure out if it’s real, ask the boring questions first, before you even zoom into anything:
Who posted this, and where did it appear first?
Is it being independently reported by an established source, or does it exist only on one account?
Does the image make you feel a strong, urgent emotion and push you to act or share immediately? That urgency is itself a manipulation pattern worth being suspicious of, regardless of how convincing the image looks.
A reverse image search is a quick, practical step here too — it can surface earlier versions of an image, reveal it’s been circulating for years attached to a different story, or show it alongside a factual debunking that’s already been done by someone else.
The single strongest signal, when it exists, is whether a real, named, accountable source stands behind the image. That’s not a trick that expires when the next model ships it works on next year’s fakes just as well as this year’s, which is more than can be said for any specific visual tell.
Where This Actually Shows Up in Daily Life
This isn’t a hypothetical, academic concern. AI-generated images are already showing up in places that affect real decisions: fake dating profiles, fabricated product listings, manufactured “news” events, and impersonation of real people.
If you’ve followed the pace of change in AI tools generally and it’s worth reading about the latest AI advancements if you want the bigger picture none of this should be surprising. Image generation has improved at roughly the same blistering pace as text generation, and the verification habits that used to be optional are quickly becoming necessary.
If you’re curious how a strange or unbelievable image spreads and gets picked apart online, it’s worth looking at how a viral image gets investigated and explained after the fact the pattern of “this looks fake, let’s find out” plays out constantly now, for images that turn out to be AI-generated and for ones that turn out to be perfectly real but just strange.
Common Mistakes to Avoid
Relying on outdated tells as your only check. Six-fingered hands and obviously garbled text were reliable two years ago. Treating their absence as proof an image is real is exactly the mistake current models are built to exploit.
Trusting a detector tool’s percentage as a verdict. A confidence score from an automated tool is a data point, not proof, especially against current-generation models where accuracy has dropped sharply.
Assuming a missing watermark or credential means an image is fake. Adoption of provenance standards is still partial. Most real, unedited photos also lack any credential simply because the standard isn’t universally applied yet.
Skipping the source check because the image “looks convincing.” Convincing is no longer a meaningful bar the volume of genuinely convincing AI images generated daily is now in the tens of millions. Source verification matters more, not less, as visual quality improves.
Frequently Asked Questions
What Is the Most Reliable Way to Spot an AI-Generated Image?
Combining provenance checks (like C2PA Content Credentials and SynthID watermarks) with visual inspection of background text, hands, eyes, and shadows is the most reliable current approach, since no single method is conclusive on its own.
Do AI Images Still Have Extra Fingers or Garbled Text?
Rarely in current-generation models. Foreground text and hands in relaxed poses have improved dramatically. The tells that still hold up have shifted to background text, hands under complex poses, and eye reflection consistency.
Can I Trust an Online AI Image Detector Tool?
Not as a final verdict. A 2026 benchmark found automated detectors are only 18–30% accurate against current commercial image generators, making them unreliable as standalone proof either way.
What Is C2PA and How Do I Check It?
C2PA (Content Credentials) is an open standard that records an image’s creation and editing history, backed by major tech and camera companies. You can check any image for these credentials directly at contentcredentials.org.
Does a Missing Watermark Mean an Image Is Fake?
No. Provenance standards like C2PA and SynthID are only partially adopted so far, so most images online real or AI-generated currently carry no credential at all. A missing credential simply means “unverified,” not “fake.”
What’s the Fastest Manual Check for a Suspicious Image?
Zoom into any small, secondary text a distant sign, a license plate, a book spine since this remains the area where AI generators most consistently fail, even as foreground detail has improved dramatically.
Conclusion
The old checklist for spotting AI images is mostly obsolete, and pretending otherwise just means getting fooled by anything reasonably current. What actually holds up is a combination: check for provenance signals when they exist, inspect the specific visual tells that still trip up current models, and more than either of that trace anything genuinely important back to a real, accountable source before you trust it. That last habit isn’t clever or technical, but it’s the one that keeps working no matter how good next year’s models get.
