The Glitch in the Mirror: What AI Still Cannot See
How AI learns to see, and what it still cannot recognise
I was not running an experiment this time.
I never set out to test AI in this way, even though I had been noticing the discrepancies for a while: the struggle to generate certain kinds of images, the inconsistency, and the gaps.
Often, I would ask two or three different systems the same question, using the exact same prompt, just to compare, to choose, or sometimes to merge results for whatever I was working on.
This time was no different.
I used the same prompt.
But when the results came back, I couldn’t ignore it anymore.
It made me think.
Because it became clear, again, that we are being under-represented, misrepresented, and excluded here too.
Except when something is needed from us.
Except when beauty wants to market itself as our solution.
For a while mow, I have been testing some of the most advanced AI systems, trying to generate images of people with visible differences.
Not abstract prompts.
Not artistic interpretations.
Specific realities.
At one point, even getting vitiligo right was difficult.
Then it started improving.
Now vitiligo is almost the default.
More recently, I’ve noticed the same with Alopecia.
Freckles too.
You can almost see the system learning what it has been exposed to.
But beyond that, it struggles.
When I ask for something like Scleroderma, Lupus, Psoriasis, Ichthyosis…, on dark skin, what I get back is not accuracy. It is an approximation.
Sometimes it produces textures that suggest “something is different” but cannot be linked to any real condition.
Sometimes it generates something that looks unfamiliar, even unrecognisable.
And after a while, you realise:
The issue is not the prompt.
It is visibility.
What AI is actually doing
AI does not imagine.
It reconstructs.
It draws from what has been documented, shared, repeated, and made visible.
So when certain conditions appear again and again, like vitiligo or alopecia, it learns them well.
When others remain under-represented, it hesitates.
It simplifies.
It substitutes.
It guesses.
And when those guesses are repeated at scale, they begin to shape perception.
A system that reflects selectively
AI is increasingly becoming a mirror.
Not a perfect mirror.
But a powerful one.
It reflects what it has been shown, and in doing so, it begins to influence what is seen as normal, recognisable, or even real.
So what happens when that mirror cannot reflect you clearly?
Or worse, when it replaces you with something easier to generate?
This is where the issue moves beyond technology.
It becomes about representation.
It becomes about who is visible and who is not.
What remains unseen
What remains unseen
There is something deeper here.
Some conditions are documented.
Some are aestheticised.
Some are repeated often enough to become familiar.
Others remain largely unseen.
Not because they do not exist.
But because they have not been captured, shared, or represented enough for systems like this to learn them.
So the system does what it can.
It fills the gap.
But filling the gap is not the same as telling the truth.
It is not the same as representation and inclusion.
Where this becomes real
This is not only about AI.
It is about how human experience is recorded, and what happens when it is not.
Through my work with The Appearance Positive (TAP) and the ASWALK Festival, I will be engaging more closely with individuals living with visible differences and appearance-related challenges.
What becomes clear very quickly is this:
Appearance is not a surface issue.
It is a lived reality.
It carries social weight.
Emotional weight.
Mental weight.
Psychological weight.
Identity.
And it shapes how people move through the world, how they are seen, how they are treated, and how they see themselves.
Yet much of this does not make it into datasets.
A quiet shift
A quiet shift
If AI is going to play a meaningful role in the future, then the systems it learns from must become more complete.
Not just clinically complete.
But humanly complete.
Because what is not seen cannot be fully supported.
And what is not understood cannot be meaningfully improved.
Closing
The question is no longer whether AI is powerful.
It clearly is.
The question is:
What does it know how to see?
And what does it still miss?