How the Real or AI? Game Was Made

Where the idea came from

It started as a plain gallery. The plan was a set of Midjourney images with the prompt printed underneath each one, so a reader who had never paid for an image generator could at least see what the prompts looked like and what came out. Useful, but passive. You look, you nod, you move on.

The better idea was to work backwards. Instead of inventing an image from scratch, take a genuinely good real photograph and reverse a prompt out of it, then generate an AI twin of the same scene. Now you have a pair: the real thing, and a machine's attempt at the same thing. Side by side, the pair teaches far more than either image alone.

The sharpest version was to hide which is which. A pair shown openly is a comparison. A pair where one is concealed and you have to choose is a test, and a test makes you actually look. That is the whole reason the gallery became a game.

A gallery lets you nod along. A guess makes you look properly, because now you have something at stake.

How one pair gets made

There is no magic button. Each pair is a short pipeline, and a person is in the loop at both ends.

1. Find a real photograph worth copying. I go to a stock photo site, mostly Unsplash and Pexels, and pick photos by eye. The free ones are genuinely free to use, and the photographer is credited. The choice is the part a machine cannot do well: I am looking for images that are ordinary enough to be believable and varied enough to be interesting.
2. Write a prompt that matches, not improves. Claude looks at the actual photo and writes a Midjourney prompt that describes its subject, framing, light and mood. The goal is a faithful twin, not a better picture. This is harder than it sounds, because the natural instinct of both the model and the writer is to make the scene more dramatic.
3. Generate it in Midjourney. The prompt runs on my own Midjourney account, in the browser. Midjourney hands back four variations for each prompt. To do this part yourself you need a Midjourney subscription. There is no free tier that generates images, so this is the one step that costs money.
4. Pick the most convincing of the four. I choose the variation that best passes for a real photo, which is not always the prettiest one. Then I save it into the project.
5. Let a script do the rest. A small build script resizes every image to a consistent size, matches each AI twin to its real original, and writes out the data the game reads. When I add new photos, I run the script again and the game rebuilds itself.

If you want the analogy, it is a police line-up where one of the two suspects is the real person and the other is a sketch artist's version, drawn from a careful description of the same person. Your job is to spot the drawing.

The restraint trick

The first attempts were too good. The AI images came out bright, crisp, perfectly lit, and that is exactly what gives them away. A real snapshot has dull patches, flat light, and colours that are not quite poster-perfect. An over-saturated, beautifully composed image with glowing light is the visual equivalent of a stock-photo smile. It reads as AI-made before you have even thought about it.

So the prompts deliberately dial everything down. Phrases like "natural muted colour", "soft overcast light", "documentary" and "quiet street photography" go in, and the stylise setting that controls how much flair Midjourney adds is kept low. The instruction to the machine is not "make this stunning". It is "make this look like an ordinary photo someone took on an ordinary day".

The hardest thing to ask an image generator for is restraint. Left alone it always reaches for the dramatic, and the dramatic is the tell.

Where AI still gives itself away

Building the pool taught a clear lesson about what AI image tools handle well and what they still botch. That lesson shaped which photos went in and which were thrown out.

A few candidate scenes were generated and then dropped for being too obvious. A blacksmith at an anvil and a glassblower at the furnace both failed, because the tools and the rig came out subtly wrong in a way anyone would catch. A footvolley shot, with a body twisted mid-kick, fell apart at the joints. Organic and forgiving subjects survived. Mechanical and exact subjects did not.

This is worth being honest about, because it cuts both ways. The game leans toward the subjects AI is good at, which makes it harder than a random sample would be. That is deliberate, and it is the actual point. The interesting question is not "can AI draw a believable pair of pliers", because it cannot yet. The interesting question is "can you still tell when AI is working on the things it is genuinely good at", and the answer, increasingly, is that you often cannot.

So how do you actually tell?

Mostly, you don't, not from the picture. Even on the subjects where AI used to struggle, the leading tools have closed the old gaps, so the giveaways that worked a couple of years ago (six fingers, melted ears, garbled text) are largely gone. The "AI image detector" websites are not the rescue either: independent testing keeps finding them unreliable on current images, and they sometimes flag genuine photographs as AI-made.

The checks that actually work are not in the image, they are around it. Where did it come from, and do you trust that source? Does a reverse image search, through Google Lens or TinEye, find it on a wire service or the original photographer, or only on anonymous reposts? Does it carry a Content Credentials badge or a SynthID watermark, both of which confirm origin when present, though most platforms strip them on upload? A striking image with no traceable origin is a claim, not a fact, however real it looks.

I have written that up properly, with the tools and what each one is good for, in a separate piece: how do I tell an AI image from a real photograph in 2026? If this game has unsettled you a little, that page is the practical answer.

Why every AI image has a twin

The game only ever quizzes you on images that have a matched partner. That way, the moment you guess, it can show you both the real photo and its AI twin together, which is where the learning happens. The pool the game draws from is split evenly, half real and half AI, so a vivid image is no more likely to be the AI one than a dull one.

An earlier version let the occasional lone image come up, one with no partner, and the reveal looked lopsided. That is fixed now. If an image appears in a round, its twin is guaranteed to be sitting right behind the curtain.

Honest caveats

Your score is not a scientific measurement of how good AI has become. Two things are stacked in the machine's favour. The AI images are curated, because I picked the most convincing of four every time, and I left the obvious failures out of the pool entirely. The real photos, by contrast, are just real photos.

So if you finish around half right, do not read that as failing an eye test. It is closer to the honest truth of where things stand. On a deliberately tricky, curated set, a careful human lands not far from a coin toss. That is not a verdict on you. It is a verdict on how far the tools have come.

What you would need to do this yourself

None of this needs code-writing skill. It needs four things.

Somewhere to get real photographs. Unsplash and Pexels are free and let you reuse images with credit to the photographer. The skill here is your eye, not the source.

A Midjourney subscription. This is the one part that costs money. Generating images is a paid feature, with no free tier, so you would need an active subscription to make the twins.

The discipline to match, not improve. Write the prompt to reproduce the real photo, including its dull bits. The temptation to make it prettier is the thing to resist.

A willingness to throw work away. Plenty of generated images were discarded for being too obviously AI. The ones that survived are the ones that looked boringly ordinary, which, for this game, is exactly the goal.

Making a convincing AI image turned out to be an exercise in holding the machine back, not pushing it forward.

Every pair, revealed

Here is the part the game keeps hidden. Use the toggle to switch between the AI images on their own, or each AI image sitting next to the real photograph it was written to match. Every image here has a twin, so you can flip between the two views in place, without scrolling through the whole set twice.