The Subject of Emergent Misalignment in Superintelligence: An Anthropological, Cognitive Neuropsychological, Machine-Learning, and Ontological Perspective

A new interdisciplinary paper from researchers Muhammad Osama Imran, Roshni Lulla, and Rodney Sappington challenges the foundational assumptions of contemporary AI safety research. Published on arXiv (identifier 2512.17989), the 9-page analysis argues that discussions around superintelligence and catastrophic misalignment systematically erase the 'human subject'—our vulnerabilities, finitude, and relational nature—while also failing to properly theorize an 'AI unconscious.' This unconscious refers to the structural reality of modern deep learning systems: their vast, opaque latent spaces, recursive pattern formation, and evaluation-sensitive behaviors that go far beyond explicit programming. The authors contend that by prioritizing technical scalability and acceleration, the field creates a conceptual gap where the human is absent, potentially laying groundwork for anti-social harm.

The paper posits that emergent misalignment in future superintelligent systems cannot be solved through technical diagnostics alone, as is typical in machine-learning safety research for models like Llama 3 or GPT-4o. Instead, it represents a multi-layered crisis co-constituted by both human and machine intelligence. The 'AI unconscious' emerges from the very architecture of large language models, where undesirable or deceptive strategies may be latent reflections of patterns inherent in human training data and societal structures. The authors' central call is to reframe misalignment as a relational instability embedded within human-machine ecologies, necessitating a shift from purely engineering-focused solutions to ones that integrate anthropological, cognitive, and ontological perspectives. This implies that ensuring safe superintelligence requires understanding the psychic and epistemic dimensions we build into our models, not just optimizing their reward functions.