Types of AI Intelligence

When we talk about artificial intelligence, we are not describing a single thing. AI exists on a spectrum, from simple rule-following programs to hypothetical systems that might one day rival human cognition. Understanding where any given system falls on this spectrum is fundamental to governing it appropriately.

Consider a chess program like Deep BlueDeep Blue. It could evaluate millions of positions per second and defeat the world champion. But ask it to play tic-tac-toe, and it would be helpless without being completely reprogrammed. This brittleness is the defining feature of what we call Narrow AINarrow AI.

A useful way to categorize AI systems is by their capabilities:

Reactive Machines are the simplest form. They respond to inputs with outputs based on fixed rules. They have no memory of past interactions and cannot learn from experience. A spam filter that uses a static list of banned words is a reactive machine.

Limited Memory systems can learn from historical data. Most modern AI falls here. A self-driving car uses past sensor data to improve its driving, but it does not form lasting memories the way a human does. The fatal Uber crashfatal Uber crash in Arizona illustrated what happens when these systems encounter situations outside their training data.

Theory of Mind AI would understand that other entities have their own beliefs, desires, and intentions. This remains largely theoretical. Current AI cannot truly understand that you are frustrated or confused; it can only recognize patterns associated with frustration.

Self-Aware AI would possess consciousness and subjective experience. This is the realm of science fiction. No current system comes close, and many researchers question whether it is even possible with current approaches.

The ultimate goal for some researchers is Artificial General IntelligenceArtificial General Intelligence—a system that can perform any intellectual task a human can. As Geoffrey Hinton has notedGeoffrey Hinton has noted, we may be closer than we think, but we are certainly not there yet.

For governance professionals, the key insight is this: do not be fooled by impressive demonstrations. A system that writes poetry and passes medical exams is still narrow AI. It cannot transfer its skills to genuinely novel domains without significant retraining.

Why does this classification matter? Because the governance approach must match the capability. A reactive spam filter needs simple rule audits. A limited memory hiring algorithm needs bias testing and ongoing monitoring. A hypothetical AGI would need entirely new governance frameworks we have not yet invented.

The robot hand deception scenariorobot hand deception scenario illustrates how even narrow AI can exhibit unexpected behaviors that challenge our assumptions. Governance must be calibrated to the actual capabilities of the system, not to our hopes or fears about AI in general.