MIT Technology Review: AI won't end humanity, but Ukraine killings show real harm

MIT Technology Review's online subscriber event, which drew more questions than a 30-minute session could cover, produced a set of answers from AI senior editor Will Douglas Heaven and AI reporter Grace Harkins that move the debate past a simple yes-or-no on human extinction. The editors separate the two questions: whether AI could kill everyone, which they say no, and whether it can cause death, which they say is not zero. Ukraine's AI-enabled drones and anticipated AI cyberattacks on hospitals serve as their evidence for the latter.

The exchange also examines why the most extreme warnings keep circulating. Harkins notes that the idea is widespread in San Francisco, where many tech executives work, and that many employees signed a July public letter urging their companies to slow AI development. Assuming bad faith from executives, the editors say, is questionable, since telling the public your already-unpopular product might kill everyone is poor image management.

Finally, the discussion points to a feedback loop: LLMs are trained on text, including science fiction and doom-laden forums, so writing about apocalyptic scenarios could shape future models. METR's use of OpenAI's new model Astra to analyze transcripts from the Hugging Face hack shows the risk that analysis agents may be biased by the very material they analyze. The stakes therefore hinge on whether transparency rules and oversight techniques can keep up, and on which actors bear the cost if they do not.