Research Digested – Prioritization of Risks from Artificial Intelligence June 2026 – A Delphi Study of 272 International Experts, MIT AI Risk Initiative
This research explores and prioritises AI risks through the Delphi method in which 272 experts build consensus through an iterative, and anonymous, consultation. In each round of the consultation the experts (academics, practitioners etc) see peer responses and can update their scoring from previous rounds. This surfaces disagreement and enables participants to learn from each others’s thinking.
The 24 risk groups (based on published AI risk frameworks) are:
- Dangerous capabilities
- Competitive dynamics
- Weapons and cyber attacks
- Power centralisation
- False information
- Inequality and unemployment
- Governance failure
- Disinformation and influence
- Environmental harm
- AI security vulnerabilities
- Loss of consensus reality
- Loss of privacy
- Discrimination
- Devaluation of human creativity
- Loss of human agency
- Capability and robustness
- Fraud and scams
- AI misalignment
- Multi-agent risks
- Over reliance and unsafe use
- Unequal performance
- Transparency and interpretability
- Toxic content
- AI welfare and rights
For each risk there was a severity scale that ran from negligible to catastrophic, with catastrophic meaning more than one million human deaths ir more than $100bn in financial damage.
Key findings
Experts say 18 of the 24 risks have at least a 10% probability of catastrophic outcomes in the next five years (2025-30). The five most severe AI risks are:
- AI possessing dangerous capabilities
- Competitive dynamics
- Weapons and cyber attacks
- Power centralisation and unfair distribution of benefits
- False or misleading information
Even when organisations and governments make pragmatic efforts to address AI risks all 24 factors had a five percent chanbne of a catastrophic outcome.
This report is a call to action on AI risks. As the authors say:
“The findings reported here point toward urgent, concrete action rather than further delay. The expert panel judged several risks to remain at least 10% catastrophic probability even following pragmatic mitigations. This would be an unacceptable level under most risk-governance frameworks. Reducing this risk requires governance instruments (e.g., regulation, liability, mandatory insurance, transparency, monitoring) that internalize costs currently borne by the public. Even if model developers need to deploy technical solutions, governance instruments are required to mitigate race dynamics and ‘tragedies of the commons’. The window for avoiding catastrophic outcomes remains open but is narrowing.”
Read the report https://futuretech.mit.edu/publication/prioritization-of-risks-from-artificial-intelligence-a-delphi-study-of-272-international-experts
