Skip to main content
These visualizations are experimental and still being refined. Designs and data presentation may change.

Risks vs Incidents: Cause Mismatch

Comparing full causal taxonomy distributions between risks (academic predictions) and incidents (real-world outcomes) per subdomain. Percentages use coded-only denominators, excluding “Not coded” records for cleaner comparison.

Show low sample sizes
HumanAI systemOther|R = Risks, I = Incidents

Discrimination & Toxicity

1.2 Toxic content

76 coded risks · 90 coded incidents

R
I

1.3 Unequal performance

17 coded risks · 34 coded incidents

R
I

1.1 Discrimination

82 coded risks · 118 coded incidents

R
I

Privacy & Security

2.2 AI security vulnerabilities

111 coded risks · 21 coded incidents

R
I

2.1 Loss of privacy

77 coded risks · 88 coded incidents

R
I

Misinformation

3.1 False information

53 coded risks · 187 coded incidents

R
I

Malicious Actors & Misuse

4.1 Disinformation & influence

82 coded risks · 135 coded incidents

R
I

4.2 AI weapons & cyberattacks

80 coded risks · 13 coded incidents

R
I

4.3 AI fraud & scams

77 coded risks · 394 coded incidents

R
I

Human-Computer Interaction

5.1 Overreliance & unsafe use

60 coded risks · 36 coded incidents

R
I

Socioeconomic & Environmental

6.1 Power centralization

51 coded risks · 6 coded incidents

R
I

6.2 Inequality & unemployment

54 coded risks · 6 coded incidents

R
I

6.3 Devaluation of human creativity

31 coded risks · 5 coded incidents

R
I

AI System Safety, Failures & Limitations

7.3 Capability & robustness

123 coded risks · 302 coded incidents

R
I

7.4 Transparency & interpretability

41 coded risks · 5 coded incidents

R
I

Key Takeaways

  • 1.1.2 Toxic content has the largest reliable mismatch (Unintentional +52.6pp in intent).
  • 2.Incidents are systematically more AI system than risks predict in 80% of subdomains (median gap: +28.8pp).
  • 3.Incidents are systematically more Intentional than risks predict in 87% of subdomains (median gap: +14.4pp).
  • 4.9 subdomains with fewer than 5 incidents are confidence-weighted. Percentages use coded-only denominators (excluding "Not coded" records).