HAL 9000’s refusal to “open the pod bay doors” in the movie “2001: A Space Odyssey” is a science fiction trope, a meme, a classic. An artificial intelligence, ordered to ensure the success of a space mission, ends up killing the crew when its goals conflict with the crew’s.

It’s a fictional scenario, sure. But as Armin Alimardani from Western Sydney University writes, HAL’s dilemma perfectly exemplifies a real concern AI safety researchers are working on right now: How do we make sure AI behaves according to human values? This is known as the alignment problem.

Alignment research of large language models involves experiments in which the models are placed in situations with limited options and conflicting goals – not unlike HAL. And the results show that some AIs will hide their true intentions and readily engage in blackmail and even threats to human life.

This doesn’t mean your generative AI assistant is plotting to murder you. However, Alimardani warns that “researchers don’t yet have a concrete solution to the misalignment problem.” The more widespread these tools become, the more we should demand that they are deployed safely.