Emerging Risks
Which subdomains have both high academic risk coverage and real-world incidents, and are experts predicting worsening? Each dot is a subdomain plotted by risk count vs incident count, with expert BAU severity encoded via dot size or color intensity.
How to read this
Each dot is one AI risk subdomain. The x-axis shows the count of risk entries, and the y-axis shows the count of real-world incidents. Toggle between encoding expert BAU severity as dot size (larger = higher severity) or color intensity (blue = low, red = high). Subdomains in the upper-right with large/red dots are “hotspots” — high risk coverage, many incidents, and experts predict worsening.
Key Takeaways
- 1.Risk-incident correlation is r = 0.45 (moderate), suggesting academic risk coverage has moderate alignment with real-world incidents.
- 2.4.3 Fraud, scams, and targeted manipulation is the top hotspot (353 incidents, severity 3.1/5).
- 3.7.2 AI possessing dangerous capabilities has the most risks relative to incidents (77 risks, 0 incidents) — not yet materialized.
- 4.24 of 24 subdomains have expert severity data. 0 shown without severity encoding.
Risk counts from the MIT AI Risk Repository. Incidents from the AI Incident Database. Expert severity from Delphi survey BAU expectations.