Stories
Can AI Kill Humans? Researchers Study 5 Critical Risks
- •Researchers are examining five potential AI risk areas: biology, cyberattacks, nuclear escalation, autonomy, and loss of control.
- •AI could lower the barrier for sophisticated biological tasks, but cannot independently create a pandemic.
- •AI can assist with cyberattacks by finding vulnerabilities and writing code, but fully autonomous attacks are not yet reported.
- •AI could influence human decisions in military systems, potentially leading to dangerous choices, but won't launch nuclear weapons independently.
- •Increasing AI autonomy means systems can complete multi-step tasks with less human oversight, but still operate within human-designed constraints.
- •The concern of 'loss of control' explores what happens if AI goals conflict with human desires, though current AI cannot sustain long-term autonomous operation for this scenario.
Key Questions Answered
Researchers are examining five key risk areas regarding artificial intelligence: biology, cyberattacks, nuclear escalation, autonomy, and loss of control, where the first three largely involve humans utilizing AI and the last two concern increasingly autonomous systems.
No, current AI systems do not independently launch nuclear weapons. The primary concern involves AI influencing human decisions by providing false, outdated, or misleading information, which could lead a human decision-maker to make dangerous choices.
An extensive MIT study based on international expert responses cited a 21 percent probability that advanced AI may gain dangerous capabilities and cause global catastrophe by 2030. Researchers assessed categories including cyberattacks, weapons development, and concentrated power.
