Key facts
- Five AI risks identified as most severe: dangerous capabilities, competitive pressures, weapons/cyberattacks, concentrated power, false information.
- Information and finance sectors are particularly vulnerable to AI risks.
- 18 AI risk domains had at least a 10% probability of catastrophic outcomes under a business-as-usual scenario.
- Catastrophic outcomes include over 1 million deaths or $100 billion in financial losses.
- AI is well-suited for coding and pattern recognition, accelerating cyberattacks.
- AI advances can make dangerous tasks like deepfake creation and weapon development easier.
A new study from MIT FutureTech and the University of Queensland, titled “Prioritization of Risks From Artificial Intelligence,” surveyed 272 international AI experts to assess 24 AI risks. The experts evaluated risks based on their likelihood and severity of harm, identifying which sectors and actors are most vulnerable. The study utilized the Delphi method, a structured process to gather expert judgments over multiple rounds. Under a business-as-usual scenario, 18 of the 24 risk domains had at least a 10% probability of catastrophic outcomes, defined as harms including over 1 million deaths or $100 billion in financial losses. Even with pragmatic mitigation efforts, five domains retained at least a 10% probability of catastrophic outcomes. AI-enabled weapons and cyberattacks were ranked high due to AI's capabilities in coding and pattern recognition, while dangerous capabilities encompass broader risks like persuasion, surveillance, deepfakes, and assistance in building weapons.