MIT FutureTech Identifies Five Major Catastrophic AI Risks

Summary

A study conducted by MIT FutureTech, involving 272 experts from 37 countries, ranked 24 artificial intelligence risk categories. It was found that under current trends, 18 of these categories have a greater than 10% probability of leading to catastrophic scenarios (1 million+ casualties or $100 billion+ in losses). The list is topped by dangerous AI capabilities, AI-assisted weapons and cyberattacks, competitive dynamics, power centralization, and sophisticated disinformation.

Why it matters

This study is one of the largest structured expert assessments regarding the tail risks of artificial intelligence. A key finding is that even after practical measures are taken, 5 risks still exceed the 10% catastrophe threshold. This result directly challenges the assumption that current governance frameworks are sufficient. The study also highlights a structural accountability gap. Those best positioned to mitigate risks (laboratories, governments) are different actors from those who will suffer the most severe consequences (users, the public). This situation has urgent implications for policy design and international coordination.

Highlights

• Top 5 risks: Dangerous AI capabilities, AI-assisted weapons and cyberattacks, competitive dynamics, power centralization, sophisticated disinformation
• Under current trends, 18 of the 24 risks exceed the 10% catastrophe probability threshold. Even after practical measures are taken, 5 risks remain above this threshold.
• The information, finance, and national security sectors are identified as the most vulnerable. Users and the public are defined as those who will experience the most severe consequences; however, these groups have the lowest capacity to mitigate risks.

Critical take

The methodology relies on a three-round anonymous Delphi survey. Therefore, the 'probabilities' in question are not predictions derived from empirical data, but rather a synthesis of expert opinions. The authors specifically emphasize that the 24 risk categories overlap and cannot be aggregated into a single global figure. Consequently, the headline '18 out of 24' should be read as a signal of widespread concern rather than a precise measurement. Additionally, the study merely identifies who should be responsible and detects the misaligned incentives of these actors. Since it does not propose concrete mechanisms, it does not effectively address the policy gap it diagnoses.