Ken Archer

Contemporary AI systems - especially large language models (LLMs) - exhibit a striking combination of capabilities and failures: apparent fluency and cross-domain coherence alongside persistent breakdowns in compositional generalization, object tracking, and truth reliability. These phenomena are often interpreted in two opposing ways: either as evidence that AI is converging on human-like intelligence, or as proof that it merely manipulates linguistic form without understanding. In recent interdisciplinary work - including Adam Frank, Marcelo Gleiser, and Evan Thompson's The Blind Spot and DeepMind researcher Alexander Lerchner's “The Abstraction Fallacy” - a different picture is emerging. Rather than asking whether AI systems are becoming intelligent in the human sense, these approaches ask a more basic question: What if AI systems work because they rely on structures that are rooted in human cognition? This shift in perspective, which draws on the phenomenology of Edmund Husserl, helps make sense of both the capabilities and the limits of modern AI. In this paper, we further develop this framework by drawing on Husserl’s account of cognition and language. According to this account, language, once properly understood, is able to mediate between natural and artificial intelligence, an argument first made by Husserl scholar Robert Sokolowski. As we argue, this framework provides a unified explanation of several central empirical findings in contemporary AI: the compositionality gap in large language and reasoning models, persistent failures in whole-part relations and object persistence in vision and world models, and the coexistence of apparent fluency with systematic hallucination. These patterns are not merely engineering challenges but reflect a structural boundary: the absence of world-directed, temporal, and anticipatory cognition that underlies natural language. At the same time, this account explains why AI systems are powerful and scientifically significant: they extend human cognitive achievements by formalizing and scaling structures originally constituted in perceptual experience and expressed in language. This resolves persistent tensions in current debates, including the apparent plausibility of AGI narratives alongside their conceptual instability, and the coexistence of genuine safety risks with misconceptions about “autonomous” or “rogue” intelligence. Because AI extends rather than reconstructs natural intelligence, responsibility lies not in the model but in the system surrounding it - resituating the focus of AI safety from model safety to system safety and grounding the current shift in practice toward harness engineering.