Extending Human Intelligence Through AI
| Source: Microsoft Research Blog
Tags: Microsoft Research, AI safety, LLMs, phenomenology, cognitive science, AI alignment
Microsoft Research argues that LLMs work precisely because they extend cognitive structures sedimented in human language — not because they replicate human intelligence — reframing AI safety as a governance challenge rather than a rogue AI problem.
Details
A new paper, 'The Origins of Artificial Intelligence in Natural Intelligence,' co-authored by Microsoft Research affiliates, draws on Edmund Husserl's phenomenology to argue that modern AI systems extend structures already present in human cognition and language. Rather than asking whether AI is becoming human-like, the authors ask why it works at all. Their answer: human language encodes stable conceptual structures built from lived experience — object permanence, spatial relations, causal concepts — and LLMs learn to model and extend those structures. This framing explains both the capabilities and the persistent failures of AI. Systems that reason fluently in language struggle with tasks requiring perceptual grounding (tracking objects through change, compositional reasoning in novel situations) because those require what language alone cannot fully capture. Hallucinations and reasoning failures are predictable consequences of AI extending a specifically linguistic, human-grounded view. The practical implication is for AI safety: if AI systems are extensions of human cognitive structures embedded in sociotechnical systems, safety is a system-level engineering and governance problem — not a story about rogue machines developing independent goals. The authors cite DeepMind researcher Alexander Lerchner's 'The Abstraction Fallacy' and Adam Frank et al.'s 'The Blind Spot' as related work. The blog post is a summary; the underlying paper offers the full technical and philosophical argument.