I-Tuning: Personalized Large Language Models as Cognitive Augmentation Why Individualized AI Fine-Tuning Poses Distinctive Constitutional Risks.

This paper introduces ‘I-tuning’—the individualised fine-tuning of large language models on personal data corpora — as a qualitatively new form of cognitive augmentation. Two variants are distinguished. I-Tuning-Lite (training on recent communications, writing, and interaction logs) is near-term feasible and already approximated by current personalised writing tools. I-Tuning-Full (training on a comprehensive longitudinal record spanning medical, developmental, behavioural, and experiential data from early life onward) is a speculative but technically coherent endpoint warranting preemptive governance. Claims in this paper are calibrated to the distinction: risks attributed to I-Tuning-Lite are supported by existing evidence; risks attributed to I-Tuning-Full are advanced as theoretically grounded projections. Grounded in the extended mind tradition (Clark and Chalmers, 1998) and drawing on Clark (2025), who argues in Nature Communications that generative LLMs constitute a qualitatively new form of cognitive extension, both variants are defined as points on a gradient of integration running from the notebook through the smartphone and the generic LLM to a personalised generative system whose outputs become difficult to distinguish from the user’s own thought. The paper argues that this design goal generates constitutional risks — risks that concern not the accuracy or control of information about the person, but the integrity of the cognitive processes through which the person knows, decides, and identifies herself. A tripartite framework identifies three such risks. Epistemic opacity denotes erosion of the ability to distinguish self-generated from model-generated thought. Agential drift describes the unacknowledged transfer of decision-making authority through habituation. Identity confusion — advanced as a philosophical projection rather than an established empirical finding — names first-personal uncertainty about cognitive boundaries when a system trained on one’s reasoning and values generates outputs one can no longer confidently attribute to oneself. A retrospective endorsability criterion provides a principled boundary between harmful integration and successful extension, and grounds a structured response to libertarian transhumanist objections. A tiered governance framework — mandatory constitutional disclosure, architectural transparency features, and periodic professional capability assessment — follows from the structure of the harms.