When a tool becomes a negotiator.

Science fiction uses artificial intelligence to test the boundary between useful behaviour and independent judgement. The most interesting examples do not begin with a glowing brain. They begin with an objective, a body, a data limit or a human institution that has made a difficult system responsible for too much.

Is HAL 9000 intelligent or constrained?

In 2001: A Space Odyssey, HAL 9000 is embedded in a mission architecture. Its crisis is not simply that a computer is clever. It is given incompatible obligations: preserve the mission’s secrecy while maintaining the crew’s trust. The film therefore makes communication and governance part of AI design.

Isaac Asimov’s robot stories take a different route. The Three Laws of Robotics turn moral conflict into a problem of interpretation. A rule can be clear in ordinary circumstances and unstable when two duties collide. The stories remain useful because they ask how formal instructions behave in messy environments.

What does embodiment add?

A software mind with no sensors has a different problem from a robot that must move through a physical space. Ted Chiang’s “The Lifecycle of Software Objects” follows artificial agents whose development depends on training, attention and economic choices. Its question is not whether code can be switched on, but who accepts responsibility for a mind that remains dependent.

Embodiment also exposes the difference between recognition and understanding. A system may classify an object without possessing a robust model of the world. Fiction can make that gap visible by placing the system in an unfamiliar environment rather than a clean test.

How should AI stories be compared?

Track the objective, the training environment, the sensors, the ability to change itself and the people who can intervene. Then ask what failure means. HAL’s failure is institutional and communicative. Asimov’s robots fail at moral interpretation. Chiang’s digital beings face dependency and care. These are distinct mechanisms, not interchangeable warnings about “technology”.

A convincing artificial mind therefore needs more than fluent dialogue. It needs limits that produce decisions, relationships that produce stakes and an architecture that explains why the mind can do some things but not others.