Artificial DNA and Self-Recovered Knowledge A Causal Theory of Inherited and Acquired Knowledge in Artificial Intelligence
This paper develops a causal theory for distinguishing information processing from genuine knowledge acquisition in artificial intelligence. Large language models can generate accurate explanations, solve unfamiliar problems, and reconstruct complex conceptual relations. However, these capabilities do not necessarily establish that the model itself has acquired the knowledge expressed in its outputs. Most of the relevant structure was formed before deployment through externally organized training over human-produced data. We introduce the concept of self-recovered knowledge. Under this framework, knowledge is not defined merely as stored information, successful prediction, or correct output. Knowledge is acquired when a subject acts upon an environment, receives consequences that are causally connected to its own action, recovers those consequences into its own organization, and preserves the resulting modification for future behavior. The paper distinguishes four categories: inherited structure, transmitted information, derived competence, and self-acquired knowledge. The pretrained parameters of a large language model are interpreted as an artificial inheritance structure rather than as ordinary autobiographical or experiential knowledge. In this limited structural sense, model weights are compared with biological inheritance: both constrain future behavior through patterns formed before the present individual interaction. The analogy does not claim that DNA contains propositional knowledge or that neural-network weights are literally genetic material. Instead, both are treated as inherited generative constraints. This distinction produces a central claim: An artificial system does not possess knowledge merely because it can generate knowledge-like outputs. It begins to acquire knowledge of its own when the consequences of its own interventions are persistently incorporated into the same continuing system that produced those interventions. The framework is applied to pretrained language models, memory-augmented systems, autonomous agents, adaptive robots, and human cognition. It also addresses major objections, including human learning from testimony, the role of external training, the continuity of artificial subjects, and the possibility of distributed or institutionally shared knowledge. The theory provides conceptual and operational criteria for evaluating whether an AI system merely expresses inherited cultural structure or genuinely acquires knowledge through its own causal history.