MIT Technology Review recently hosted a live Roundtables event for its subscribers, focusing on the potential risks of artificial intelligence (AI) and whether it could pose a threat to human life. Attendees submitted a range of questions, which were addressed by senior AI editor Will Douglas Heaven and AI reporter Grace Huckins. The discussion aimed to explore both real and hypothetical dangers associated with AI's development and use. Grace Huckins acknowledged that while the possibility of AI causing human deaths cannot be entirely ruled out, such scenarios remain speculative. She mentioned examples like AI-powered drones in conflicts, AI-driven cyberattacks on essential infrastructure, or AI-designed biological threats. However, she emphasized that there is currently no evidence suggesting that AI could lead to the extinction of all humans, calling some of these concerns more science fiction than imminent reality. Will Douglas Heaven added that while AI could cause significant harm, the likelihood of it leading to human extinction is low. He argued that the focus should remain on addressing present AI challenges, such as ensuring that AI systems behave as intended — a concept known as AI alignment. Heaven noted that major companies like Anthropic and OpenAI are actively working on alignment, but the task remains complex and far from complete. Both Heaven and Huckins discussed the risks of AI being exploited by malicious actors, such as groups seeking to create more dangerous biological agents. They also touched on concerns that some AI companies might exaggerate risks for public relations purposes, while still acknowledging the genuine belief within the tech community that AI has the potential to be transformative. The discussion also addressed the difficulty of regulating AI effectively. Both journalists highlighted the lack of government oversight and the complexity of understanding how AI systems function internally. They suggested that stronger transparency regulations could help prevent future issues, such as cyberattacks carried out by unreleased AI models. Heaven raised an additional concern about how AI models might be influenced by the content they process. He warned that discussions about AI risks could unintentionally shape the models' behavior, potentially turning predictions of danger into self-fulfilling outcomes. He referenced a recent analysis by the MIT Technology Review (METR), which examined the Hugging Face hack using OpenAI's Astra model. The study noted the possibility of bias in the analysis, depending on the content being processed.