AI systems extend human intelligence rather than replicate it, research shows

Microsoft researchers argue that modern AI systems work not by replicating human intelligence, but by extending structures already present in human cognition and language—a shift in perspective that helps explain both their remarkable capabilities and their persistent boundaries.

Understanding AI through human cognition#

The research, drawing on philosophical traditions including Edmund Husserl's phenomenology, proposes that large language models learn statistical patterns embedded within language itself. These patterns reflect how humans express stable structures rooted in lived experience—concepts like "round," "red," or spatial relationships that originate in perception and movement through the world.

"[Words] articulate relationships that originate in lived experience," the researchers explain. This framework helps clarify why AI systems excel at tasks like writing essays and generating code—they're extending conceptual patterns sedimented in human language—while struggling with intuitive human abilities like tracking objects through change or compositional reasoning in novel situations.

Why hallucinations and brittleness emerge#

A key insight: humans remain answerable to the world through continuous experience. AI systems, by contrast, extend patterns within text itself. They can generate fluent reasoning without the "lived engagement with the world" that anchors meaning and truth—explaining why they hallucinate plausible but false outputs.

This also clarifies the "compositionality gap." Language models perform well on familiar patterns but fail when asked to combine concepts in genuinely novel ways. Larger models improve fluency and recall faster than true compositional reasoning—not just an engineering limitation but a structural boundary reflecting the systems' origins in human linguistic patterns.

Similar brittleness appears in multimodal systems combining vision and language. They learn correlations between visual patterns and language rather than perceiving stable objects the way humans do, making them impressive but fragile outside familiar contexts.

Reframing AI safety#

The research reframes AI safety as a system-level challenge. Public discourse often swings between fears of "rogue superintelligence" and dismissals of risk. The researchers argue both extremes misunderstand current systems: the immediate risks aren't from AI possessing human-like intentions, but from systems that extend reasoning patterns without reflective accountability to reality.