AI can write compelling essays, generate complex code, and carry on sophisticated conversations, yet it still struggles with tasks that seem intuitive to humans—like tracking objects through changes or distinguishing truth from fiction. This paradox has sparked heated debates: Are we witnessing the birth of artificial minds, or just sophisticated autocomplete on steroids?
Groundbreaking research from Microsoft suggests we've been asking the wrong question entirely. Instead of debating whether AI systems are becoming human-like, we should ask: What if AI works precisely because it builds upon structures rooted in human cognition?
The Hidden Foundation: Language as Crystallized Human Understanding
According to researchers Ken Archer and Harald Wiltsche in their paper "The Origins of Artificial Intelligence in Natural Intelligence," modern AI systems succeed by extending patterns that already exist in human language and thought. Drawing on the philosophical work of Edmund Husserl, they argue that language itself contains "sedimented structures" of human understanding—crystallized ways we make sense of the world.
Think about how you perceive a coffee cup. As you move around it, your brain automatically maintains the understanding that it's the same cup, despite seeing different angles. Language captures these stable structures through words like "round," "solid," or "container." Large language models learn to model and extend these linguistic patterns across vast amounts of human writing.
This explains both AI's remarkable fluency and its peculiar limitations. AI systems can generate coherent responses across many domains because they're working with the accumulated patterns of human understanding. But they also hallucinate because, unlike humans, they're not grounded in direct experience of the world that originally created those patterns.
Why AI Hits Walls: The Compositionality Problem
This framework illuminates persistent challenges in AI development. Take the "compositionality gap"—AI's tendency to excel at familiar reasoning patterns while failing when asked to combine concepts in genuinely novel ways. Research shows that larger models improve at factual recall much faster than at true compositional reasoning.
From this perspective, it's not just an engineering problem to be solved with more data or compute. It's a structural limitation: AI systems can extend existing patterns in language, but they lack the world-directed understanding that allows humans to create genuinely new conceptual relationships.
Similarly, multimodal AI systems that combine vision and language often label images correctly but struggle with robust reasoning about objects and their relationships. They're learning correlations between visual patterns and words rather than perceiving stable objects unfolding through time as humans do.
Reframing AI Safety: From Rogue Robots to Responsible Systems
This research fundamentally reframes AI safety debates. Instead of oscillating between fears of "rogue superintelligence" and dismissive claims that AI poses little risk, we can focus on the real challenge: AI systems can extend patterns of reasoning without the reflective responsibility to the world that grounds human judgment.
The most immediate risks aren't from AI developing malicious intentions, but from systems that can:
- Generate persuasive but ungrounded outputs
- Automate flawed decisions at massive scale
- Execute harmful actions when embedded in poorly governed environments
This is why the industry is shifting from "model safety" to "system safety"—implementing layered safeguards, monitoring mechanisms, and governance structures that the authors call "harnesses." These aren't temporary patches but fundamental architectural requirements for trustworthy AI deployment.
The Path Forward: Extending, Not Replacing
Understanding AI as an extension of human intelligence rather than a replacement offers a more grounded path for development. It clarifies why human oversight, governance, and responsibility remain central to AI systems—not because the technology is primitive, but because that's how it fundamentally works.
As the researchers note, "The central societal risk of AI turns out to be kicking away the ladder of its origins in human experience and cognition." If we mistake AI for autonomous minds, we risk over-trusting them. If we dismiss them as mere tricks, we miss their genuine potential.
The real question isn't whether AI will replace human intelligence—it's how we can responsibly build systems that extend human understanding while remaining grounded in the world from which that understanding arises. This perspective offers both humility about AI's limitations and clarity about its remarkable promise as a tool for amplifying human cognition.
For practitioners building AI systems and prompts, this research suggests focusing not just on what AI can do, but on how to harness its pattern-extending capabilities while maintaining human judgment and oversight. The future of AI isn't about creating artificial minds—it's about thoughtfully extending our own.
Based on research by Ken Archer and Harald Wiltsche, published on Microsoft Research.