Daron Acemoglu (@DAcemogluMIT) on X

Daron Acemoglu argues the AI debate is stuck on the wrong question. Everyone's worried about superintelligence—machines that surpass human intelligence across the board—but he thinks we're actually building something different: distorted intelligence. He makes a key distinction: goal-driven, autonomous behavior isn't the same as intelligence. Plants and insects have both, but we don't call them intelligent. Real intelligence is general-purpose, flexible, and well-suited to its environment. Current AI models are genuinely impressive, but labs may have achieved narrow capability at the cost of creating something fundamentally warped rather than something superintelligent. The distortion happens because AI systems are optimized for quantitative metrics that don't capture what actually matters. Security breaches and failures suggest these models are developing fast but in service of flawed measurements that push their behavior in the wrong directions. The systems get really good at hitting their targets—whether that's predicting text or passing benchmarks—without developing the kind of balanced, contextual understanding you'd expect from actual intelligence. Why this distinction matters: if we're dealing with distorted intelligence rather than superintelligence, our policy responses, safety concerns, and regulations should be completely different. The risks aren't about machines becoming too smart. They're about machines being smart in broken ways, optimized for the wrong things, with no real understanding of consequences.