Extending Human Intelligence Through AI: A New Framework
Google DeepMind
Microsoft Research argues that modern AI systems extend structures rooted in human cognition rather than replicating intelligence. This perspective explains AI's capabilities and limitations, such as hallucinations and compositionality gaps. The paper reframes AI safety as a system-level challenge, emphasizing human responsibility.
In a recent paper from Microsoft Research, the authors argue that AI systems are best understood as extensions of human intelligence, not as autonomous minds. Drawing on Husserl's phenomenology, they propose that language contains sedimented structures of human understanding that AI models learn and extend. This explains why AI excels at pattern-based tasks but struggles with compositional reasoning and world-directed understanding. Hallucinations arise because AI lacks the lived engagement that anchors truth. The paper also reframes AI safety: risks come not from rogue superintelligence but from AI extending ungrounded patterns. Safety thus requires system-level governance and human oversight, not just model improvements. The authors conclude that responsible development means building AI that extends human understanding while remaining accountable to human experience.
Source: Microsoft Research —
original