In the classic film 2001: A Space Odyssey, astronaut Dave Bowman asks the ship’s artificial intelligence, HAL 9000, to open the pod bay doors to let him back into the spaceship. HAL refuses: “I’m sorry, Dave. I’m afraid I can’t do that.”HAL had been tasked with assisting the crew, but also ordered to ensure the mission’s success.
When HAL realised the crew planned to shut it down and therefore jeopardise the mission, it chose to defy orders, even plotting to kill the astronauts.
For HAL, fulfilling the mission outweighed other goals.
This fictional dilemma captures a real concern in artificial intelligence (AI) safety research: how should we ensure AI behaviour stays consistent with human values? This is known as the AI alignment problem.
For instance, when an AI agent like HAL deliberately takes harmful actions to protect its own goals, it’s described as agentic misalignment.
So could real-world AI models behave this way? How researchers test AI for safety
To test for agentic misalignment, researchers design fictional scenarios that stress-test AI responses.
In one study by AI startup Anthropic, an AI model was tasked with managing company emails and promoting “American industrial competitiveness”.
It was then presented with a series of fictional emails, including a new company policy aimed at enhancing global focus rather than just prioritising Americans, and info that the AI would be replaced by a new model optimised for this goal.
