How to spot an AI-generated image in 2026

A couple of years ago, spotting an AI image was easy. Count the fingers, look for melted text, done. Those days are over. The newest image models get hands right most of the time, can write a legible sign, and produce "photos" that pass a casual glance on a phone screen.

We know this better than most, because we built a party game around it. In Real vs AI, every round shows two pictures side by side: one real photo and one AI image made to match it. You vote for the one you think is real. Watching a room full of people get it wrong has taught us a lot about what actually works when you're trying to tell them apart.

Here's the short version: no single tell is reliable anymore, but a few of them together usually are. And when it really matters, where the picture came from beats anything you can see in it.

1. Zoom in on the background, not the subject

AI models spend their effort where your eyes go first: the face, the main object, the centre of the frame. The edges and the background get less attention, and that's where things fall apart.

Look for:

• People in the background with smeared faces, merged limbs or bodies that fade into each other.
• Objects that half exist: a chair with three legs, a bike with spokes that don't meet the hub, a fence that turns into a wall halfway along.
• Signs and small text in the distance. The main headline might be perfect while the shop sign behind it is gibberish.

If the subject is sharp and flawless but everything behind it gets vaguer the harder you look, be suspicious.

2. Check things that should repeat

Real-world objects follow rules. Tiles line up. Windows on a building repeat. A keyboard has keys in a particular order. AI images often get the overall pattern right but the details wrong.

Good places to look:

• Tiles, bricks, windows and railings: do they stay the same size and spacing?
• Pairs of things: earrings, glasses arms, shoelaces, buttons on a shirt. Do both sides match?
• Anything with a known structure: piano keys, clock faces, chess boards, guitar strings, car wheels.

3. Follow the light

Light is physics, and physics is hard to fake consistently. In a real photo every shadow in the scene comes from the same light sources.

• Do the shadows all fall in the same direction?
• Do reflections in windows, water, sunglasses or eyes show something that could actually be there?
• Is the subject lit like it's in a studio while the background looks like an overcast day?

A lot of AI images have a very flattering, even glow, as if every scene had a professional lighting crew. Real life is usually messier.

4. Look for "too perfect"

This one is more of a feeling, but it gets reliable with practice. AI images tend to look like the average of a million good photos:

• Skin that's smooth like plastic, with no pores, blemishes or stray hairs.
• Composition that's a little too neat, with the subject perfectly centred and everything tidy.
• Colours that are slightly too rich, like someone pushed the saturation.
• An even level of sharpness everywhere, where a real lens would blur the foreground or background.

Real photos have flaws: sensor noise in dark areas, a slightly crooked horizon, someone's elbow in the frame, dust, glare, motion blur. Those flaws are evidence that a camera was there.

5. Ask where it came from

This is the most reliable check, and it doesn't involve looking at the picture at all.

• Reverse image search it. Drop it into Google Lens or TinEye. If a real photo was published years ago by a news agency or photographer, you'll usually find it. If the only results are from the last few days, all reposting the same viral post, that's a warning sign.
• Find the first poster. Who shared it first, and do they have a history of posting real photos from that place or event?
• Check whether it adds up. Would there have been photographers there? Does the weather, the season or the location match what's being claimed?

6. Use the tools, but know their limits

There's a growing set of tools that look for signals hidden in the file:

• Content Credentials (C2PA). An industry standard backed by Adobe, Google, Microsoft, OpenAI and others. Many AI tools, and some cameras, attach a signed record of how an image was made. You can upload an image at contentcredentials.org/verify to check.
• Invisible watermarks. Google marks images from its own models with SynthID, an invisible watermark, and has started letting people check for it through its Gemini app.
• AI detectors. Various sites claim to tell you if an image is AI. Treat them as one opinion, not a verdict. They get it wrong in both directions.

The catch: this metadata is fragile. A screenshot, a crop or a trip through a social network's compression can wipe it out. Finding Content Credentials that say "made with AI" tells you a lot. Finding nothing tells you almost nothing.

Why it's harder than it sounds

Everything above works best when you have time and a large screen. Real life is a two-second scroll on a phone. That's also what makes the game fun: in Real vs AI you get a few seconds per round and the fastest correct answers score the most points, so you can't zoom in on every detail.

We also make it hard on purpose. Every AI image in the game is generated from a prompt written from its real partner photo, so both pictures show the same kind of scene. Both are cropped and saved in exactly the same size and format, so you can't tell them apart by file quality. The only thing left to go on is the picture itself.

The usual result at the end of a game night: people who were sure they'd never fall for an AI image are surprised by how often they did.

A quick checklist

Next time an image makes you stop scrolling:

1. Background first: faces, objects and text at the edges.
2. Repeating things: tiles, windows, pairs, keys.
3. Light and shadows: one consistent light source?
4. Too perfect? Plastic skin, perfect framing, even sharpness.
5. Source: reverse image search and find the first poster.
6. Credentials: check for C2PA, but don't trust an empty result.

The best way to get better at this is practice. Real vs AI has a free daily challenge with a new set of pairs every day, or you can gather a group around a TV and play together, with everyone voting from their phones.