AI People Quiz: Real Person or AI-Generated?

Modern models like Flux, Seedream 4.5 and DALL-E 3 generate portraits that look strikingly real. But the People quiz goes beyond single portraits — it includes everyday situations: couples, groups, street scenes, people at work.

These scenarios force AI to get clothing, backgrounds, lighting, shadows and human interaction all right at the same time — which is exactly where the cracks still appear.

Each round contains one real photo and three AI-generated images from different models:

• Juggernaut Pro Flux — excellent skin texture and lighting, often too polished
• Kling IMAGE 3.0 Omni — strong composition, occasional artifacts in backgrounds
• Z-Image — highly realistic, struggles with group interactions
• DALL·E 3 — consistent quality, sometimes over-saturated colors
• Seedream 4.5 — sharp detail, hands and fingers occasionally off

Enable "Show model names" before starting to reveal which AI generated each fake after every round.

The authentic option in every round is a licensed photograph, not an image that merely “looks real.” The three synthetic alternatives are generated for the same described scene and retain their model label in the quiz data. This matters most for people scenes: comparing a studio portrait with an AI street photo would test photographic style rather than authenticity.

Before a round is published, the four options are checked for matching subject matter, comparable image quality and an unambiguous source label. We also avoid relying on face quality alone. Current models often render a convincing face while failing in secondary evidence such as overlapping hands, repeated background people, inconsistent accessories or impossible contact between bodies and objects.

Across our recorded quiz answers, players find the real photograph in the People category about 52% of the time. In a four-image round where guessing alone would yield 25%, that is clearly better than chance — but it sits in the lower half of our image categories, well behind food and paintings, and barely ahead of animals.

The reason is not that faces are unbeatable. It is that faces are the part everyone checks, and the part models have most thoroughly solved. Skin pores, hair strands, catchlights in the eyes — all of these are now rendered convincingly enough that a careful look at a face usually confirms nothing either way. Players who spend their attention there come away with a confident feeling and a coin flip.

What still fails is everything around the person. That is why this quiz uses couples, groups and street scenes rather than isolated portraits: a single figure gives a generator one thing to get right, while two people in contact give it a relationship to get wrong.

A café scene, two people mid-conversation. Faces look fine in all four images — as they usually will. This is the order that actually resolves the round:

1. Find where the two bodies interact. A hand on a shoulder, elbows on the same table, one person leaning toward the other. Generators place figures convincingly but negotiate the space between them badly: an arm that rests on a surface without deforming it, or a shoulder that occupies the same space as the chair behind it.
2. Follow one object from hand to table. A cup, a phone, a fork. Check that the fingers actually wrap it, that the object casts a shadow onto the surface, and that the surface is where the shadow says it is.
3. Count people in the background — then count them again. Background figures are cheap for a model to render and easy to get wrong: duplicated faces, a person with no legs behind a table, a waiter whose apron merges into a wall.
4. Check accessories for symmetry. Glasses with mismatched arms, one earring of a pair, a watch strap that changes width. These survive when faces do not, because they are small, structured objects the model treats as texture.
5. Read any text last. A menu, a sign, a logo on a cup. If there is legible text it is often the fastest answer of all — but in people scenes there frequently is none, which is why it comes at the end rather than the start here.

Note what is missing from that list: counting fingers. It was the defining AI tell for years and current models handle hands correctly in most outputs — we wrote about the tells that stopped working separately. Hands are still worth looking at, but only where they are doing something.