Spot the Fake: Real Landmark or AI-Generated?

This quiz features 16 world-famous landmarks from across the globe. Each round shows you 4 images of the same landmark — find the real one.

Famous buildings are both easy and hard for AI. The models have seen thousands of photos of these landmarks during training, so they know the general shape and style. But getting every architectural detail right — window patterns, stone textures, surrounding environment — is where AI still struggles.

Some landmarks are easier to fake than others. Simple, symmetric structures like the Brandenburg Gate look convincing as AI images. But complex buildings like the Sagrada Familia, with their intricate details, often reveal the AI.

Landmarks sit in the lower half of our image categories by raw accuracy — the live figures are above, and they move as people play. That ordering is the opposite of what almost everyone predicts. Buildings have hard edges, fixed proportions and countable elements; they ought to be the easiest thing in the world to verify. And in principle they are — the problem is that people do not verify them.

Recognition is not verification. When you look at a photograph of the Taj Mahal, the question your brain answers in the first fraction of a second is “is this the Taj Mahal?” — and the answer is yes for all four images, including the three fakes. That yes feels like confirmation, and most players stop there. With an unfamiliar building you have no such shortcut, so you actually look at it.

The practical consequence is uncomfortable but simple: the landmarks you know best are the ones you check worst. The Eiffel Tower example below exists because it is the most recognisable structure in the quiz, and therefore the one people get wrong most confidently.

There is a compensation, and it is the reason this category rewards practice more than any other: a landmark has a correct answer that exists outside the image. A fox in long grass could plausibly look a hundred different ways. The Sydney Opera House has a fixed number of shells, the Leaning Tower has a fixed number of arcades, and Big Ben's clock face has fixed proportions relative to the tower.

This means you can prepare. Pick one countable, checkable feature per landmark and commit it to memory — not the whole building, one feature. The number of tiers. The shape of a window arch. Whether the columns are fluted. Then in a round, you are not judging whether an image looks right; you are checking a fact you already know against what is in front of you.

Generators are strong on overall silhouette precisely because that is what dominates their training data. They are weaker on the specific count, because a plausible-looking number of arches satisfies the objective just as well as the correct one.

The Eiffel Tower is a perfect trap, because everyone "knows" it — which makes it easy to accept a wrong one. Run these three checks before you commit:

1. The lattice. The real tower is a precise iron lattice that follows a consistent curve from the four legs up to the tip. AI versions frequently muddle the ironwork halfway up, or let the two sides climb at slightly different angles.
2. The count of levels. There are two clearly defined observation platforms below the summit. Fakes often add a phantom third tier or blur the platforms into the structure.
3. The base and crowd. Look under the arches at the people and the Champ de Mars behind. Cloned, faceless figures and a smeared park are the giveaway — the tower itself is usually the last thing AI gets wrong.

The same approach — memorise one true detail, then check it — works for every landmark in the quiz.