How to Tell If a Photo Is AI: 7 Signs With Real Examples
CozyPhoto4u by SomaShort answer: In 2026 the best AI generators produce photos with realistic light, readable text and correct hands, so no single visual clue is proof anymore. Start with the file: ask Google Search, Lens or the Gemini app whether the image carries an AI watermark, check its Content Credentials and do a reverse image search. Then look for six visual signs: lighting that's too perfect, objects nobody asked for, a place that could be anywhere, a background that stops making sense, reflections that don't match, and a complete lack of imperfections. When several show up together, the image is very likely AI generated.
I sell photographs for a living, so I'm obviously not neutral on this one, and I didn't want to write yet another angry post about AI. What I actually wanted to know was simpler: if I put one of my own photos next to an AI image of more or less the same thing, could I explain the difference to someone who doesn't care about photography at all?
It's harder than it sounds. In a 2025 study by Microsoft's AI for Good Lab, more than 12,500 people looked at around 287,000 images and picked correctly only 62% of the time, and for landscapes just 59%. So I set up my own AI photo vs real photo test: I took three of my photos and found AI images that come as close as possible, same subject, similar angle, similar mood. All of them are from StockCake, where every image is AI generated, so you can click through and judge for yourself.
Check if the image is AI generated: watermarks and Content Credentials
Because AI pictures have become so convincing, I'd start with the file rather than your eyes. Google hides an invisible SynthID watermark in images made with its AI tools, and since May 2026 you can ask Google Search, Lens or the Gemini app "Is this made with AI?" to check for it. ChatGPT images are getting SynthID too, alongside Content Credentials, a kind of digital passport showing how a file was made, which you can check for free at contentcredentials.org.
A missing watermark proves nothing, though. Many AI tools don't add one, social media often strips metadata, and only a few newer cameras write Content Credentials. My Nikon D600 from 2012 doesn't. A reverse image search helps too, since it shows whether the picture appeared earlier and whether it was labelled as AI. File clues are worth ten seconds as well: neat square sizes like 1024 x 1024 or names like gemini_generated_image point to AI, while camera names like DSC_1234 point to a real photo.
Online AI detectors can give a second opinion, but treat their score as a clue, not a verdict.
I ran my photos through two free AI detectors
I wanted to know if the online detectors would get it right, so I uploaded my three night photos and the three AI images next to them to two free tools, Sightengine and Hive. No sign-up, just drag and drop. The number is how sure each tool was that the image is AI generated.
| Image | Sightengine | Hive |
|---|---|---|
| My cat photo | 0% | 1.6% |
| AI cat | 99% | 71.8% |
| My lantern photo | 0% | 0% |
| AI lantern | 99% | 99.9% |
| My Ferris wheel photo | 0% | 0.1% |
| AI Ferris wheel | 99% | 99.9% |
Tested on 9 October 2026, using the original files.
Both tools got all six right. The closest call was the AI cat: Sightengine was 99% sure, Hive only 72%.
The tools also tried to guess which generator made each AI image, and they didn't always agree. For the Ferris wheel, Hive was sure it was Flux, while Sightengine leaned towards Krea. A detector can be right about the yes or no and still be guessing about the details.
My takeaway: a free detector is a good second opinion, and on these six images both did well. But the AI cat shows how much two tools can differ on the same picture, which is why I wouldn't trust a single score on its own. Check two tools, then look at the picture yourself.
Signs
1. The lighting is too perfect
AI images love studio light, even in scenes where no studio could exist. The Mysterious Cat is probably my favourite photo in the CozyPhoto4u shop: a cat crossing a wet street at night, completely black against golden reflections on the ground. I had maybe two seconds, it didn't look at me and it didn't stop.
Next to it, Midnight Cat Silhouette has a clean glowing outline all the way around the cat, the sort of rim light a photographer sets up with two lamps and a lot of patience. On a real street at night, light comes from wherever it happens to be. In my photo it comes from below, bounced off puddles.
Be careful with this sign the other way round, though. In the Microsoft study, the real photos people most often mistook for AI were the ones with unusual lighting, and night photos like mine are exactly that kind of picture. Lighting is only a clue when it's too clean, not when it's merely strange.
2. Things appear that nobody asked for
An image generator doesn't photograph a scene, it predicts what a scene like that usually contains. So it adds things. Light Beam is an old lantern on a pole with a bush in front of it and not much else going on. The closest AI match, Illuminated Garden Lantern, added a rustic stone wall and a neat layer of ivy on its own, the kind of setting you'd expect in a catalogue for garden lights.
The Ferris wheel comparison below is even funnier. The AI version's own description mentions carousel horses blurred underneath the wheel. Nobody asked for carousel horses, but for an image generator a Ferris wheel means a funfair and a funfair means carousels, so they appear. If a picture contains every object you'd expect in that kind of place, and nothing you wouldn't, be suspicious.
3. The place could be anywhere
My photo Up & Around is a Ferris wheel in the Prater at night, in black and white, with lit spokes running out from a flower shaped hub. I cut off the bottom of the wheel on purpose.
The AI version, Luminous Carnival Night, is also monochrome, also against a black sky, also full of light radiating from the centre. But it's a Ferris wheel in general, not a particular one. It had no way of knowing about the flower in the middle of the hub, or how many gondolas this wheel has. Real places have specific details you can check: a street sign, a shop name, a building you can find on a map. AI landmarks tend to look familiar and belong nowhere.
4. The background stops making sense
A few years ago you could zoom in and find six fingers or gibberish letters on a shop sign. The newest generators have mostly fixed that. They write readable text and draw hands correctly most of the time, so clean lettering is no longer proof that a photo is real. Where they still slip is the logic of the background: a railing that melts into a wall, windows that don't line up from floor to floor, a path that leads nowhere, a pattern that repeats a little too neatly. Look at the boring corners of the image, not the subject in the middle.
In my lantern photo, zooming in shows tiny air bubbles in the glass and paint on the bracket worn to a kind of copper brown. Those details are odd and uneven, but they're consistent. That's what real objects look like up close.
5. Reflections and shadows don't match the scene
Light in a real photo has to come from somewhere, and everything in the frame obeys the same sources. In AI images, shadows sometimes point in different directions, reflections show things that aren't in the scene, or a glowing object lights up its surroundings in a way that doesn't add up. Wet streets and windows are good places to check. In The Mysterious Cat, every bit of gold on the cat's edges matches a reflection on the ground below it.
6. There are no imperfections at all
Real photos keep the small mistakes of a real moment. In Light Beam there are two blurry lights in the background that I would have removed if I could, but I couldn't, so they stayed. The lamp is off centre, the bush is slightly out of focus and the glass is a bit dirty, because that's how the lantern looked. The cat in my photo is a little chubby and has one paw in the air mid step. The AI cat has a pose straight out of a stock photo of an elegant cat.
When every element of an image is placed exactly where a designer would want it, with nothing awkward, cut off or out of place, that perfection is itself a clue.
With people in the picture, the same rule applies to skin and hair. AI skin often looks waxy and poreless, strands of hair blend into the background, and earrings or teeth don't quite match from one side to the other. I mostly photograph places rather than people, but in a portrait that's the first thing I'd check.
Why it matters: what the research says
Research shows that people value art more when they believe a human made it, even when they can't tell the difference.
Because "it just feels more real" isn't much of an argument, I also looked for studies. In a 2023 paper in Cognitive Research: Principles and Implications, researchers showed people a set of paintings that were all made by AI, but labelled some of them as human created. The human labelled ones were rated higher, and while the difference in beauty ratings was small, it got much bigger when people were asked how profound and how valuable the work felt.
Research from Columbia Business School in 2025 found something similar with money involved: art labelled as AI made was valued about 62% lower than the identical piece labelled as human made. Getty Images, in a survey of more than 30,000 people across 25 countries for their Building Trust in the Age of AI report, found that almost 90% want to know whether an image was made with AI, and 76% agreed that it's getting to the point where they can't tell if an image is real.
I don't think AI images are useless. For mood boards and quick sketches of an idea they're handy, and some of them are beautiful. On my own walls I still prefer photos of things that were actually there, and that's the only kind you'll find in the CozyPhoto4u webshop, all shot by me.
Last updated: October 2026
Read next: Is the Nikon D600 Still Worth It in 2026? Why Cheap Gear Still Takes Great Photos and A Photo Gift for Someone Who Has Everything: Why a Real Print Works.
About the author
Soma is the photographer behind CozyPhoto4u, capturing Vienna architecture, city nights and nature with a real camera. Every photo in the shop is Soma's own work, never AI generated, and available as a canvas or framed print.
Frequently asked questions
Can you tell if a photo is AI just by looking?
Not reliably anymore. In Microsoft's 2025 study people picked correctly only 62% of the time. Looking still helps, but it works best together with a watermark check and a reverse image search.
Do ChatGPT images have a watermark?
They carry Content Credentials, and in May 2026 OpenAI announced they're getting Google's invisible SynthID watermark too. Content Credentials live in the file's metadata, though, which a screenshot or a social media upload can strip.
Can Google tell if an image is AI?
Partly. Google Search, Lens and the Gemini app can find the SynthID watermark, but images from tools without it may not be recognised, so a "not found" result doesn't mean the photo is real.
Can AI recognise AI generated images?
Often, yes. AI detectors are trained on the patterns generators leave behind, and in my test both Sightengine and Hive spotted all three AI images. But every new generator changes those patterns, so detectors are always catching up.
Is there a free AI image detector?
Yes. Google Search, Lens and the Gemini app check for SynthID, contentcredentials.org shows Content Credentials, and Sightengine and Hive, the two I tested above, let you check an image for free without signing up.
Are AI image detectors accurate?
On my six test images, both Sightengine and Hive got every one right, but they disagreed a lot on the AI cat: one was 99% sure, the other only 72%. They can also miss images from the newest generators or flag real photos, so check two tools, then look at the picture yourself.
How can I check a photo I saw on social media?
Look for an AI label on the post, run a reverse image search to find the original, and ask Google Search or the Gemini app whether it was made with AI. Most platforms strip metadata on upload, so missing camera data tells you very little.