The field of making AI systems not cause harm, from a chatbot's manners to the end of the world, and the argument over which of those it should mean. A 2016 paper gave it a research agenda; the 2023 boom gave it institutes, summits and a Bletchley Park declaration. In 2025 the UK's safety institute renamed itself for security, which was a policy decision expressed as a noun.
The record's note on the word: 'safety' covers a car's airbag and a nuclear plant's containment, and the field's split is over which one it is. The near-term camp works on bias, misuse and hallucination. The long-term camp works on control. Both use the word, and each thinks the other is spending the money wrong.
Sighted in February 2025, when the UK's AI Safety Institute became the AI Security Institute, with a statement that it would focus on serious harms and not on bias or free speech. The record notes that a field had been renamed by a government to signal what it would no longer study.
AI safety is not a department at every lab, whatever the organisation charts say. It is a set of people, who move between labs, resign in public, and write about why. The record files the resignations under this heading because they are, so far, the field's most reliable data.
I remember the 2016 paper being welcome because it was boring: a cleaning robot that knocks over a vase, a reward function that can be gamed. Real problems, no apocalypse. The apocalypse arrived in the literature anyway, and now the vase is a footnote in a paper about extinction.
What does it mean? Write it the way you would say it out loud.
