A quiet afternoon in a traditional kissaten coffee shop on the second floor of an old shopping arcade in Kyoto. The interior has dark wooden panels, cream-colored walls with slight yellowing from decades of cigarette smoke, and frosted glass windows letting in soft diffused light. An elderly Japanese man in his late seventies sits alone at a small table by the window. He was a university professor before retirement. His face carries the quiet dignity of someone who has spent a lifetime with books. He wears a simple cardigan over a collared shirt, reading glasses perched on his nose. He is deeply absorbed in a thick English hardcover book, something academic and dense, perhaps philosophy or linguistics. A white ceramic coffee cup sits beside him, half-empty. His weathered hands hold the book with familiar ease. The scene is captured as if by a film camera, natural lighting, shallow depth of field focusing on the man while the kissaten interior softens into warm bokeh behind him. Photorealistic, indistinguishable from an actual photograph.

I discovered that Gemini can generate images indistinguishable from photographs. I am now trying to master this. An old man reading a book while drinking coffee at a traditional kissaten on the second floor of a shopping arcade. An underground world where Bluesky spreads across the landscape. I never tire of looking at them.

Yet realistic AI-generated images of people are not easily accepted by society. Why? I want to think about this through the lens of "individual identity."

Proper Nouns and Common Nouns — The Vector Reverses

Ask an AI to "create an image of Trump and Biden boxing in a lava ring," and you can tell it is Trump and Biden. They refer to two real people.

But when I generate "an old man reading at a kissaten in a shopping arcade," who is the man that appears on screen?

Suppose you are in a real kissaten. When you say "an old man is reading" about the person in front of you, you are referring to someone who exists. It might be Taro Yamada or Ichiro Sato. A specific person exists, and you describe that person using the category "old man." The vector goes from individual to category.

AI generation is the reverse. Only the category "old man" is given, and something with individual features emerges from it. The vector goes from category to individual. Ask who he is, and there is no answer.

Three Kinds of Individuals

To organize this: there are three kinds of individuals.

Actual individuals. People who exist or existed in the physical world. Taro Yamada, your grandfather, historical figures. They are rooted in physical space and have identity.

Fictional individuals. Characters in fiction. Hamlet, Doraemon, Frieren. They do not exist, but they are fixed by names and narratives, and can be socially transmitted as "that character." They have identity.

Potential individuals. AI-generated entities. Existences scooped from the latent space of training data by a prompt. They have individual faces and expressions, but they are no one. They have no identity.

Potential Individuals Have No Mechanism to Accumulate Identity

Fictional individuals were also singular at first. Hamlet has identity because Shakespeare gave him a name, gave him a story, and he has been repeatedly referenced. At the moment of creation, he was no different from a potential individual.

Identity, then, is not something innate. It is accumulated through fixation and repetition.

Potential individuals have no mechanism for this accumulation. Technology may advance to the point where images that "look like" the same person can be repeatedly generated. But even so, the structure remains the same: each generation is a sampling from latent space. It is a repetition of singular generations.

The identity of fictional individuals is transmitted through proper names and their repeated descriptions. The author holds "this character" as a mental representation, fixes it in words, and readers receive it. AI has no such circuit. In latent space, there is no representation of "this person" — only distributions of features. No matter how many similar outputs are produced, they are not "different appearances of the same person" but "similar strangers."

What Society Rejects

The uncanniness of realistic AI-generated human images is not because "AI made them."

It is because they pretend to be actual individuals without the circuit of identity.

They try to gain trust with faces of humans who do not exist. They appear to have accumulated lives. But the owner of that face is nowhere. A no one acts as if they were someone. This structure is what feels uncanny.

Sometimes a stranger you pass on the street momentarily looks like an AI-generated potential individual. This is not an illusion. An actual individual, as long as they remain a stranger, appears as "an individual without identity" — the same as a potential individual. We are normally surrounded by the absence of identity. But when it is artificially generated, that absence becomes stark.

Can Potential Individuals Become Fictional Individuals?

Conversely, if potential individuals gain a circuit of fixation and repetition, they can transition to fictional individuals.

An AI-generated character becomes a VTuber, gains fans, accumulates narrative. Then they become transmissible as "that person." Identity is born.

Being a potential individual is not the problem. The problem is circulating as an actual individual or fictional without a circuit to accumulate identity.

The old man I generated is no one. But I like looking at him. Knowing he exists nowhere, I enjoy his singularity. I think that is one honest way to engage with potential individuals.

Note: "narrative" here does not mean literal storylines alone. It includes history, causality, and provenance. Sometimes the author establishes it, sometimes it emerges intersubjectively among readers. Fictional individuals can transmit this accumulation socially. Potential individuals lack this circuit.

Addendum: Psychology has the concept of empathy—the ability to feel what others feel while maintaining the distinction between self and other. Projection from cognitive psychology is also involved.

When viewing an image, one projects past experiences and common knowledge onto it. Against a background of social relationships, yet in solitude, one can "give" a "narrative" to the image from one's own side.

Whether one can accept potential individuals hinges on this. Even without a circuit of identity, if the viewer can bestow narrative, the image holds meaning. This is why I like looking at the old man I generated.