Dive Brief:
- Rogue or weaponized artificial intelligence poses 10% odds of causing 1 million deaths, at least $100 billion in losses or other “catastrophic harm” by 2030, with finance an especially vulnerable sector, according to a study based on estimates by 272 AI experts worldwide.
- AI deployed in finance could trigger widespread losses by turbocharging capabilities in market manipulation, fraud, cyberattack, computer sabotage and breaches of privacy, according to the study by MIT FutureTech and the University of Queensland.
- “Dangerous capabilities, weapons and cyberattacks, power centralization, inequality and unemployment, and environmental harm have a greater than 10% chance of catastrophic outcomes even if pragmatic mitigations are put in place,” the researchers said. Even with mitigation efforts, “experts judge more than a 5% chance of catastrophic outcomes” across 24 “risk domains.”
Dive Insight:
The MIT/Queensland findings echo other studies in recent years that have sounded an alarm over AI risks to humanity, sparking public opposition and prompting steps toward regulation.
For example, AI experts in the Existential Risk Persuasion Tournament in 2022 estimated 3% odds that AI will lead to human extinction.
AI poses the biggest risk on the prospect that by 2030 it will gain capabilities in mass deception, weapons development, public manipulation, political abuse and cyber-offense, according to the MIT/Queensland study.
Other significant risks loom large, including the possibility that nations or companies in an “AI race” seek to gain an advantage by rushing to deploy systems that are inadequately checked for errors or safety threats, according to the study.
Bad actors could use AI to launch cyber attacks or create chemical, biological, radiological or nuclear weapons, according to the study.
AI may concentrate power and resources, leading to inequitable distribution of its benefits and social inequality, or it may inadvertently generate false or deceptive information, warping people’s outlook and provoking physical, emotional or material harm, according to the study.
Efforts to blunt AI risk run up against an obstacle — AI users and stakeholders are more vulnerable to the various threats than the AI developers, regulators, lawmakers and standards-setters with the authority to reduce the risks, according to the researchers.
“Providers of compute, cloud and data infrastructure are largely insulated from many of the risks of AI,” they said.
The gap in vulnerability has created “misaligned incentives in addressing the most important AI risks,” according to the study.
“Even well-understood risks can persist when responsibility is diffuse, misaligned with capability or disconnected from those who bear the harm,” the researchers said.
Moreover, “commentators have argued that the emerging regulatory landscape resembles a patchwork rather than a coordinated framework, and it remains unclear which risks governments should prioritize,” they said.
