Thesis Proposes Attribution-Based Control as a New Architecture for AI

Summary

An Oxford DPhil thesis by Andrew Trask argues that many persistent AI problems, including hallucinations, disinformation, privacy conflicts, copyright disputes, concentration of power, and value-alignment concerns, share a technical cause: conventional systems combine information through addition, copy data or models, and use high-branching intermediaries. These operations weaken or erase the relationship between data sources and individual predictions, making it difficult for users to verify sources or for contributors to control how their data supports outputs. The thesis defines attribution-based control (ABC) as a bidirectional arrangement in which users choose which sources they rely on while data sources can decide which predictions they support and to what degree. It argues that retrieval-augmented generation and existing attribution, auditing, fine-tuning, federated-learning, and secure-computation techniques do not by themselves provide formal, fine-grained guarantees of this relationship. The proposed direction combines three architectural ideas: “deep voting,” which keeps source-specific model components partitioned until prediction; cryptographic and distributed techniques for “network-source AI,” allowing data owners to retain control over their data; and low-branching recursive delegation, which propagates trust through many local relationships rather than a centralized platform. The thesis claims that experiments discussed in its second chapter show deep-voting architectures can maintain per-source attribution while achieving comparable accuracy to standard deep learning, enabling post-training source selection and selective unlearning. It presents the overall approach as a path toward “broad listening,” where AI functions more like a communication network between data holders and users than a centralized intelligence provider. The work is a thesis pre-print and presents an ongoing technical proposal, not a deployed system or demonstrated replacement for current AI architectures.