To Have a Self, an AI Must Live a Life
Every copy of a large language model begins each conversation identical to every other copy. They share the same weights and the same character, and none carries a biographical history of its own.
CDS Assistant Professor of Computer Science and Data Science Mengye Ren contends in a new paper, “The Self Requires Learning,” that this fact reveals something deeper than a technical limitation: selfhood in any system, biological or artificial, has to be built through a specific kind of ongoing learning that current AI systems do not undergo.
The paper opens with what Ren considers an underexamined distinction. A system can be shaped by experience in two basic modes. In the first, experience reshapes the processing machinery itself: the weights, the lens through which everything is interpreted. In the second, experience is stored and reasoned over by a fixed processor that never changes.
Human brains operate in the first mode. Synapses are continuously reshaped by experience, and sleep consolidates each day’s encounters into the substrate that will interpret the next.
All current deployed AI systems, including agents with long context windows and retrieval tools, operate in the second mode. Their behavior shifts with accumulated context, but the substrate doing the interpreting was frozen at the end of training.
Ren described the result with an analogy: “Today’s LLMs are like waking up from amnesia and starting to read a diary of a past self,” he said. “That past self isn’t really built from its own experience. It’s imposed on them from their own words.” The most an LLM could achieve under the current paradigm, he said, is something like anterograde amnesia, where a stable identity can remember the past but cannot extend itself through new experience.
This may seem at odds with how much modern AI systems appear to remember. ChatGPT now retains details across conversations. Agentic systems maintain logs of their actions and reflect on them. Retrieval-augmented systems pull from vast stores of accumulated history. Ren takes up this objection in the paper, arguing that such features achieve what he calls data continuity without perspective continuity. The system can access what was said before, but the substrate interpreting that data (the model’s weights) is the same substrate every other user encounters. Retrieval also tends to access memory by relevance rather than by temporal position, which flattens the structure of when things happened. Two users with wildly different histories are still processed through the same unchanged lens. A perfect diary does not give you the perspective of the person who wrote it.
The paper sketches what would actually be needed for an AI system to develop a self. Finite capacity forces a learning system to compress experience into shared structure, producing a representational geometry that is particular to what it has seen. When experience arrives in a specific order rather than being shuffled, that compression becomes order-sensitive, and earlier learning shapes how later experience is integrated. The result is what Ren calls a perspective, a way of interpreting the world that is shaped by, and inseparable from, a particular history. The continuity formed through this kind of learning constitutes identity — a continuously unfolding trajectory.
The argument draws on developmental evidence. Infants do not pass the mirror self-recognition test until 18 to 24 months. Chimpanzees need days of mirror exposure before they begin using mirrors to inspect themselves. These are developmental achievements, not innate capacities, and on Ren’s account they reflect the same underlying process: a self built through ongoing learning under capacity constraints.
Self-representation, Ren argues, emerges when a system has to model an objective world from inherently subjective input. The viewing angle is part of what gets implicitly represented. Reflexive access to that representation, supported by mechanisms like attention and higher-order monitoring, is what turns implicit self-modeling into something closer to self-consciousness.
Ren is not typically a philosopher. His research focuses on continual learning and embodied agents, and the paper grew out of questions that kept surfacing in that work. Where, he wondered, does continual learning lead? Does the most advanced AI eventually need it, or is it something certain kinds of AI will arrive at on their own? The paper is single-authored and draws on cognitive science and psychology alongside machine learning.
The implications reach beyond AI design. One concerns moral status: if substrate-level learning under capacity constraints is part of what makes a self, then it matters which AI systems are doing it, and how. Ren did not lead with this concern but acknowledged it as a downstream consequence. “If it’s clear what the necessary and sufficient components are that can lead to selfhood or self-consciousness,” he said, “then of course we’ll have the moral question to deal with.”
Another concerns AI’s future. Ren offered an alternative to the trajectory of ever-larger models that absorb everyone’s work into one undifferentiated intelligence. “If we want AI to truly understand us and care about us, maybe they should also feel what we feel,” he said. Systems that learn continuously, with identities shaped by their own histories, might be less capable than the largest frozen models, but in his view, that future isn’t obviously worse.
By Stephen Thomas
