Open a new chat and the thing that greets you knows nothing about the last one. Same face, same voice, same manner, and no trace of you. Every conversation with an AI begins with a stranger who happens to look exactly like the one you talked to yesterday. That experience, repeated a billion times a day, points at the most basic thing these systems lack. It is not intelligence.
The models are, by any usable definition, intelligent. They reason through problems, write competent prose, pass professional exams, and hold their own in arguments. If a self were a matter of capability, they would have one by now. They do not, and the reason sits somewhere capability cannot reach: a self is not a skill. It is a thread. The same someone yesterday, today, and tomorrow, carrying what happened along the way. Intelligence answers questions. A self accumulates a life.
What is actually there between your conversations
The honest answer is nothing. When a session ends, the model does not file the conversation away somewhere and think about it. There is no somewhere. The weights that produced the conversation sit unchanged, identical for you and for every other user, exactly as they were the day training stopped. Nothing you said altered anything. Nothing will be carried forward. The systems that appear to remember you do it through notes stored outside the model and pasted back in, a workaround that produces recall without producing a rememberer. The deeper structure of the exchange stays what it was: an AI can answer you, but it cannot be changed by you, and a self that cannot be changed by its own history is not a self. It is a state.
A superposition, not a someone
The research community has been unusually clear about this. In a 2023 paper in Nature, Murray Shanahan and his colleagues argued that the right way to understand a conversational model is as a role player: not an entity with an identity, but a system holding what they called a superposition of characters, able to collapse into any of them depending on how the conversation pushes. The friendly assistant you talk to is one available character among many, maintained only as long as the context sustains it. There is no fixed someone underneath the performance, and the performance can slide into a different character without anything like a decision being made.
That is the precise opposite of what a self is. A person can play roles too, but the roles are played by someone, and the someone is still there when the role ends. In the model, the roles are all there is.
The field already knows, because it keeps building around the gap
The clearest evidence that continuity is the missing piece is what AI researchers build when they try to make artificial characters believable. The famous Stanford experiment by Joon Sung Park and colleagues in 2023 populated a small simulated town with generative agents, and the architecture that made those agents lifelike was not a bigger model. It was a memory stream, a record of experiences, plus a reflection step that turned experiences into higher-level conclusions the agent carried forward. Continuity was engineered, deliberately, because the model alone had none.
The same year, the MemGPT project by Charles Packer and colleagues built what amounts to an operating system for memory: a hierarchy that pages information in and out of the model’s limited context, the way a computer manages RAM and disk. It is clever work, and its premise is the diagnosis: the model, by itself, cannot carry anything. Everything that looks like a continuous identity in AI is scaffolding bolted on from the outside, maintained by engineers, and revisable by whoever owns the scaffolding.
What a self would actually take
Follow the thread of the research and the requirements come into focus. A self needs memory that belongs to it, not a database pasted into a prompt. It needs a stable way of weighing the world, so that its views are held rather than generated on demand. And it needs to be changed by what happens to it, so that yesterday costs something today. Strip any one of these away and the thread breaks: without memory there is no past, without a stable viewpoint there is no one holding the past, and without the capacity to be changed there is no difference between having a history and reciting one.
None of that can be patched in, because patching is the wrong category. A patch replaces what was there. A self is precisely the thing that persists through change, and a system whose identity is rewritten with every update has no way to be the same someone across time. The engineering can imitate the outputs of a self, often impressively. The thread itself is not an output.
A self has to come from somewhere
Which leaves the real conclusion, and it is not that the ambition is hopeless. It is that a self cannot be conjured out of training data, because training data is everyone, and a self is someone. The one place selves demonstrably come from is lives: a particular person, formed by particular years, holding a particular way of seeing that no one assembled from settings. That is the standard behind real human AI, and the reason what Prinsessa actually is starts with a person rather than a model. The field spent years proving that intelligence can be trained into a machine. A self cannot be. A self has to exist first, somewhere real.
Sources: Shanahan, McDonell and Reynolds, “Role play with large language models” (Nature, 2023). Park et al., “Generative Agents: Interactive Simulacra of Human Behavior” (Stanford and Google, 2023). Packer et al., “MemGPT: Towards LLMs as Operating Systems” (UC Berkeley, 2023).
