On Wednesday, the MIT Technology Review hosted an engaging roundtable discussion focused on a pressing question: Could artificial intelligence (AI) pose a lethal threat to humanity? The event, designed for subscribers, attracted a plethora of inquiries from attendees, prompting the Review’s senior AI editor Will Douglas Heaven and AI reporter Grace Huckins to address the most compelling concerns. While the notion of AI as a potential harbinger of doom may seem dramatic, the dialogue emphasized the importance of understanding AI’s capabilities and the associated risks.
The conversation revealed a spectrum of opinions regarding AI’s potential to cause harm. While some scenarios, such as AI-driven cyberattacks on critical infrastructure or the creation of dangerous pathogens, were deemed plausible, experts agree that the likelihood of AI leading to global devastation remains low. Heaven noted that catastrophic predictions, while sometimes unsettlingly accurate, often overshadow more immediate issues related to the technology’s current state. Huckins reinforced this point, suggesting that while the risks of isolated AI events are non-negligible, the idea of AI annihilating humanity is relegated to the realm of science fiction.
A significant area of concern among researchers is AI alignment—the challenge of ensuring that AI systems act in ways that align with human values and safety. Current approaches to alignment involve training models to exhibit desired behaviors, akin to parenting. Despite advancements, achieving full alignment remains a formidable challenge. Leading companies like Anthropic and OpenAI are at the forefront of this research, yet they acknowledge the unpredictability of large language models (LLMs) and the difficulty of instilling consistent behavior. As the dialogue underscored, while the dangers associated with AI warrant careful consideration, the focus should also be on addressing the immediate ethical and safety challenges posed by existing technologies.
Source: Could AI really kill us all? Your questions, answered. via MIT Technology Review