A picture of a person who does not exist still resembles somebody

The invented face is not invented from nothing

Generated people are now good enough that a reader assumes a photograph. That is the useful property and it is also the source of every difficulty, because an image that reads as a photograph of a person makes an implicit claim: somebody stood there.

The figure produced is not a copy of any individual, and it is also not unrelated to real faces, and the resemblance question is not settled by the fact that no particular person was intended. A close likeness to a real person can occur, and it is more likely when the description narrows the field — a role, a setting and a distinctive physical detail together can converge on somebody who exists.

For most uses this is a low risk that is worth naming rather than a reason to stop. For anything where a viewer might identify the figure with a real individual in a specific context, it stops being low.

Asking for a named person is a separate decision

There is a clear line between generating an unnamed figure and generating a recognisable public individual, and the second is a different act regardless of how easy it is. Producing an image of a real named person doing something they did not do is the mechanism behind most of the harm in this area, and the technical ease of it has no bearing on whether it is acceptable.

The awkward middle is a person who is not named and is identifiable anyway: a colleague, a client, somebody whose appearance is distinctive within a small community. The rule that holds in practice is about identification rather than about naming.

Where a real person is genuinely required, the answer is a photograph taken with their agreement. That is not a workaround, it is the correct instrument, and it costs less than the conversation that follows an image nobody consented to.

The default figure carries assumptions you did not state

Ask for a nurse, an engineer, a criminal, a chief executive, and you will get a person with an age, a build, a skin tone and a bearing that you never specified. Those defaults come from what is most commonly depicted, and using them without noticing means a set of images that quietly reproduces a narrow idea of who does what.

The correction is to specify, and to specify across a set rather than within one image. A single picture cannot be representative; a run of twenty for a publication can be checked for what it collectively shows, and that is the level at which the problem is visible.

Specificity also improves the pictures. A described person with an age, a build, a posture and a plausible relationship to the setting is a better image than a generic figure, quite apart from anything else.

Be careful with cultural and religious specifics in particular, where the default is the most circulated version rather than an accurate one, and where the details on clothing and setting are read closely by the people they depict.

Expression and hands are where the illusion breaks

The characteristic failures with human figures cluster around expression, gaze and the small articulations. Faces come out slightly composed, gazes land somewhere near the camera without meeting it, and expressions that should be complicated arrive as pleasant. In a picture meant to convey difficulty, sadness or concentration, that flatness is the whole problem.

It is more reliable to ask for the situation and the posture than for the emotion. A person looking down at their hands, a figure half-turned away from the group, someone standing while everyone else sits: the body carries the feeling more dependably than a named emotion does.

Check people at full size, and check the hands, the ears, the teeth, the join between hair and forehead, and any jewellery or lanyard, which are the details a reader will notice precisely because they scan faces for a living.

Say what it is, and know where it does not belong

Where a generated figure stands in for a real one — a testimonial, a case study, a staff page, a photograph illustrating an account of events — the picture is making a claim about the world and should be labelled or replaced. This is the boundary where the practice moves from illustration into misrepresentation.

A short caption is usually enough for the legitimate cases: an illustration rather than a photograph, a composite, a depiction. The label costs nothing in a decorative context and it is the entire difference in a factual one.

And there are contexts where the answer is simply no. News images, evidence, anything depicting an identifiable person in a situation they were not in, and anything involving children. In those, the correct instrument is a camera and a person who agreed to be in front of it.