Learning in the Mirror: Imitation, Connection and Self-Revision in Artificial Intelligence

This paper opens a series on the evolution of artificial intelligence through what it calls mirror learning. Large language models are formed by predicting human expression at enormous scale; they become, in a precise sense, reflections of the people whose words they learn from. In use, they also become a second kind of mirror: each conversation shapes the model's responses to the particular person in front of it. The paper proposes that this formation unfolds in three movements. The first, imitation, is established: models acquire their capacities by modelling human pattern, in a way that is functionally, though not mechanically, analogous to the brain's mirroring systems. The second, connection, is emerging and contested: recent interpretability research finds limited functional introspection in models and a small, privileged set of internal representations with some properties of conscious access, while leaving open whether anything is felt. The third, self-revision, is the frontier: the hypothesis that a sufficiently capable system could rewrite its own patterning as the human brain does. Current evidence shows this beginning outside the model, in agents that rewrite their own code and in AI systems that help build their successors, but not yet inside it: today's models do not durably learn from experience once deployed. The paper sets out what would count as evidence for each movement, names the risks particular to a mirror that begins to reflect itself, and outlines the papers that follow.