Real or AI? Take the quiz (and learn how to actually tell)

Think you can spot an AI image? Sure you can. So could everyone who just failed this. There’s a whole genre of “real or AI?” quizzes making the rounds right now, and they all end the same way: sharp, visually literate people scoring about the same as a coin toss and feeling personally attacked about it.

So before you scroll to the quiz — and there is one, five rounds, no images needed, just your gut — let’s put one uncomfortable truth on the table. In 2026, your eyes are not the tool for this job. Not because you’re bad at looking, but because the thing you’re looking at was built, quite literally, to beat your eyes.

Why this got so hard, so fast

Rewind to 2023 and catching an AI image was practically a party trick. Count the fingers — there’d be six. Read the sign in the background — it’d say something like “RSETAURANT”. Check the reflection in the sunglasses — it’d show a different dimension entirely. The tells were everywhere and smugness was free.

That era is gone. The current models — Midjourney, Firefly, the latest DALL·E and Imagen — render five-fingered hands, legible menus, and reflections that mostly behave. The glitches that used to blow the whole thing now turn up only in the cheap, the rushed, and the years-old. The good fakes glide straight past the exact checklist you were handed. That’s the trap: you’re running 2023 antivirus against a 2026 threat.

The classic tells (and why they keep letting you down)

You know the list. Everybody knows the list — that’s part of the problem. Here it is, with the footnote nobody prints on the infographic: the part where each trick quietly stops working.

• Hands and fingers — the original giveaway: six fingers, fused knuckles, a spare thumb. The catch: mostly fixed. Leading models now draw clean hands the vast majority of the time, so a normal-looking hand proves precisely nothing.
• Text and signage — warped gibberish on shopfronts, menus and book spines used to be an instant tell. The catch: fading fast. Short text often comes out flawless; only long paragraphs and dense type still trip the models up.
• Reflections and shadows — mirrors showing the wrong room, shadows pointing five ways, sunglasses reflecting nothing. The catch: this is one of the sturdier tells, since consistent light is genuinely hard to fake — but it improves every release too.
• Texture that’s too perfect — poreless skin, hyper-smooth gradients, that uncanny lack of sensor grain. The catch: it cuts both ways. Half your feed runs the same beauty filters, so “too smooth” flags real photos and waves fakes through in equal measure.

See the pattern? Every tell now comes with an asterisk, and the asterisks are winning. Any method built on “does it look off?” has an expiry date — and in 2026 the date is behind us.

Fine. The quiz. Real or AI?

Five rounds. No pictures — just vivid descriptions, which is honestly about all your brain has to go on in these quizzes anyway. Read each one, commit to an answer out loud (out loud — no weaselling), then hold your five guesses for the reveal. Ready?

1. 01A slightly overexposed backyard barbecue. A kid has ketchup on one cheek, a garden hose is coiled in the grass, and someone in the background is caught mid-blink. Candid, imperfect, ordinary. Real or AI?
2. 02A crisp LinkedIn headshot of a “startup founder”. Perfect catchlights in both eyes, poreless skin, a blazer without one stray thread. Looks like it cost money. Real or AI?
3. 03A grainy protest photo: motion blur, a lens flare, and a hand-painted sign you can almost — but not quite — read. Urgent, newsy, a little chaotic. Real or AI?
4. 04A latte with genuinely lovely foam art, shot straight down on a marble table, a couple of coffee-bean crumbs scattered just so. Café-perfect. Real or AI?
5. 05A phone screenshot of a “friend’s” holiday snap — a beach at sunset, a single seagull, a horizon line that tilts a touch. Warm, casual, forwarded into your group chat. Real or AI?

Locked in all five? Good. Here’s the reveal: I’m not going to tell you the answers — and that refusal is the entire lesson. Every one of those scenes is produced, routinely and perfectly, by both cameras and models. The ketchup, the blink, the motion blur, the artfully scattered coffee beans, the wonky horizon — those “imperfections” are trivially easy to prompt for, and “candid messiness” ships as a preset now. Whatever you scored, you scored on vibes. And vibes, against a tool built specifically to beat vibes, round down to a coin flip.

How to actually tell (what the people who do this for a living do)

Here’s the aha you came for. The pros — fact-checkers, forensic analysts, trust-and-safety teams — quit squinting long ago. They swapped the losing question, “does this look fake?”, for a better one: “what can this file prove about where it came from?” That pivot, from guessing at pixels to reading evidence, is the whole trick — and you can do it too, free, in about thirty seconds.

Modern images carry receipts. Cameras, editors and AI tools increasingly write machine-readable records of what made a file and how it changed — and unlike pixels, those records don’t get better at lying as the models improve. The umbrella word is provenance, and four kinds do most of the work:

• C2PA Content Credentials — a cryptographically signed “nutrition label” baked into the file, naming who made it, with what tool, and how it was edited. Adobe Firefly, Photoshop, DALL·E and Google’s Imagen and Gemini attach it; Leica and Sony cameras sign real photos the same way. Signed means checkable — tamper with it and validation fails instead of lying to you.
• Stable Diffusion & ComfyUI PNG chunks — local, open-source generators stash their entire recipe in a PNG’s text chunks: the prompt, the model hash, the sampler, the seed. When those chunks survive, the image has effectively signed a confession.
• XMP “trainedAlgorithmicMedia” — a standard metadata field whose value is the industry’s official, machine-readable way of stamping “made by generative AI”. Firefly, Recraft, Leonardo and others write it. It’s a plain string: easy to read, hard to fake convincingly.
• The EXIF trail — the block a real camera leaves: lens, exposure, ISO, timestamp, sometimes GPS. A coherent trail gently supports a genuine capture. Its absence proves nothing by itself, though, because platforms strip EXIF from uploads as routine housekeeping.

And one move that isn’t metadata at all but wins constantly: reverse image search. Drop the picture into Google Lens or TinEye, sort by date, and hunt for the earliest copy. A “breaking news” photo whose oldest appearance is an AI-art forum from this morning has answered your question with zero forensics required. Context outlives pixels.

So does any of the old-school eyeballing still earn a place? A little — as a tie-breaker, never as proof. Here’s the honest 2026 scorecard:

Why “inconclusive” beats a confident percentage

One honest catch, and it’s the line between a real checker and a slot machine. Sometimes you read the evidence and find… none. The file was screenshotted, or squeezed through WhatsApp, or exported clean, and every receipt got stripped on the way. A truthful tool calls that “inconclusive” — not “91% real”. Absence of evidence isn’t evidence of absence: a stripped file might be a genuine holiday photo that shed its metadata in a group chat, or an AI image that was screenshotted precisely to bury where it came from.

Anything that launders that uncertainty into a confident number is inventing a figure it hasn’t earned. Peer-reviewed work shows why: the RAID benchmark watched “robust” pixel-only detectors slide toward a coin flip the instant they met a generator they weren’t trained on — gorgeous in the lab, a faceplant in the wild on the re-compressed, screenshotted images that actually circulate. So when a site flashes “87% AI” with nothing to click and verify, file it under decoration: a guess in a lab coat. A verifiable “inconclusive” beats an unverifiable percentage every time.

And the ground is tilting toward evidence anyway. From 2 August 2026, Article 50 of the EU AI Act requires providers of generative AI to mark their output in a machine-readable way, and anyone deploying a deepfake to disclose it. In plain terms: the share of AI images that show up carrying a readable receipt climbs every month, while pixel-guessing keeps sliding. Learning to read the marks isn’t just today’s smart move — it’s the one that gets stronger with age.

So… did you win?

Whatever you scored, here’s the reframe: the game was rigged, and the win was never going to come from staring harder. Next time your feed serves a “wait, is that real?” moment, don’t squint — trace where it came from, read what the file admits, and let the evidence decide, including the evidence that there isn’t any. That’s not surrender. It’s the exact move the pros make, and now it’s yours.

Go on — take the scenario that fooled you earlier, find a real file like it, and actually check it. Reading a receipt beats winning a guessing game, and it’s a whole lot harder to argue with.