Why generated faces fail, and what fixes them.

The usual complaint about generated people is that they look fake. That is not quite it. Look closely at a bad AI portrait and the giveaway is rarely a sixth finger or a melted ear, because those are mostly solved. The giveaway is that the skin is too even, the symmetry is too clean, and every pore has been quietly averaged into its neighbours.

The result is not a face you distrust. It is a face you do not remember. That is the more expensive failure, because nobody writes in to complain about it. The campaign simply does not land, and the brand concludes that the medium does not work.

A face reads as a person because of what is wrong with it

Real faces are asymmetric. One eye sits fractionally lower. Brows disagree with each other. There are broken capillaries at the side of the nose, freckling that goes patchy at the temple, a scar nobody remembers getting, a tooth that never quite straightened. None of it is flattering and all of it is load-bearing.

Strip that out and you get the visual equivalent of a voice with no accent. Technically clear, regionally nowhere, impossible to place or recall. So the standard we work to runs the other way: the imperfections are not tolerated, they are specified. They get directed as deliberately as the lighting.

Cast on the things a brief actually specifies

The other half of the problem is that most generated people are not cast at all. They are summoned. Someone types an adjective and accepts whatever arrives, which is why so much of it converges on the same handsome, ageless, ethnically ambiguous non-person.

A casting director does not work that way. They work to variables: age band, colouring, bone structure, hair, height and build, and the register the brief needs, whether that is warm or severe or entirely ordinary. Those variables exist because clients genuinely need to choose between them. Generated casting only becomes useful when it answers to the same list.

The practical test is boring and reliable. Shoot the whole board on one lighting setup, one lens, one ground, and change nothing but the person. If the faces still look like different people under identical conditions, the casting is real. If they start to blur into one another, it was never casting, it was a style.

What it is actually for

Two things, mostly. The first is seeing a decision before paying for it: the expensive way to test a casting direction is a shoot day, and the cheap way is a board. The second is duration. A booked face is a licence with a clock on it, a territory, a term and a renegotiation the moment the campaign works. A generated face has none of that, so it can carry a brand into a market that was never in the original buy, and come back next season without a rebook.

None of which is an argument against photographers. The best of this work is made by people who have art-directed real shoots and know what a real face does under a real light. The tool is new. The judgement is not, and the judgement is still the whole job.

See the casting board →