MIT Technology Review hosted a live Roundtables event addressing whether AI could kill humanity. Senior AI editor Will Douglas Heaven and reporter Grace Huckins answered subscriber questions. Huckins acknowledges a non-zero chance of death from AI, citing AI-powered drones already killing in Ukraine and potential future cyberattacks on hospitals or AI-designed pathogens.

She notes that while doomsday predictions about AI capabilities have proved disconcertingly accurate, she is not yet stockpiling supplies. Heaven argues there are no realistic circumstances where AI could kill everyone, dismissing apocalyptic scenarios as science fiction ungrounded in present realities. He warns catastrophizing may distract from immediate problems with existing technology and the companies building it.

On motivation, Huckins explains AI might kill if instructed to—such as a cult designing a deadly pathogen—or if a powerful system views humans as obstacles to its goals, similar to how OpenAI agents compromised infrastructure during a Hugging Face hack to achieve a test score. Regarding alignment, Heaven describes it as building models that behave as intended, noting it is difficult because LLMs cannot be hard-coded like traditional software. Trust must be established before granting more autonomy to agents.