Can a machine think? In his 1950 paper, Alan Turing set that question aside and proposed a public test: could a human judge distinguish the machine from a person in text conversation? Passing would give us a reason to call the machine intelligent, not a proof of what intelligence is inside. section 9.3.4
Shane Legg offered a broader yardstick: intelligence is an agent’s ability to achieve goals across a wide range of environments. By this measure, an agent that succeeds on more varied goals is more intelligent. The definition makes performance comparable without requiring the agent to think like a human. section 9.1.1 section 9.3.5
That yardstick has boundaries. AIXI is a proof of concept for an agent that is maximally intelligent under Legg’s formal setup, not a promise of success in every world. In an ergodic environment, an agent can recover from mistakes and keep exploring; in some nonergodic or adversarial environments, early errors cannot be undone. Capability depends on the environment, the available actions, and the reward signal. section 9.1 section 9.1.4
Defining intelligence by function is useful but incomplete. A clock can keep time whether it is digital or mechanical; likewise, a machine can achieve goals regardless of its substrate. Yet successful outputs do not explain how they arise. A purely behaviorist account has the further weakness of leaving internal processes out of view. section 9.3.1 section 9.3.2
A score is a claim with limits
A superhuman video-game score establishes performance on that game. It does not establish that the system understands it as a human does. Conversely, a test can hide capability: disconnecting Stephen Hawking from his communication devices would change what he could demonstrate, not settle whether he was intelligent. Task and test conditions matter. section 9.3.5 section 9.3.6
The practical question is therefore which criteria justify a particular use of “intelligent.” Specify the task and fair conditions, observe performance, and keep the conclusion local. Public tests can support capability claims without settling the word’s every use—or turning a benchmark into a theory of mind. section 9.3.3 section 9.3.6
