Risks & Failure Modes

This section documents risks—harmful outcomes and the mechanisms that lead to them. These are organized into four categories: accident risks (technical failures), misuse risks (intentional harm), structural risks (societal harms), and epistemic risks (threats to collective knowledge).

Each risk is tagged with a causal level: outcomes (end-state harms), pathways (mechanisms), or amplifiers (enabling conditions).

Risk Categories

Accident

Technical failures where AI systems behave in unintended ways

Misuse

Intentional harmful applications by malicious actors

Structural

Systemic risks from how AI reshapes society and institutions

Epistemic

End-state harms to collective knowledge and truth

Causal Levels

Risks operate at different levels of causation. Understanding this helps identify where interventions are most effective:

Level	Description	Examples
Outcome	End-state catastrophic harms we ultimately want to avoid	Lock-in, Authoritarian Takeover, Epistemic Collapse, Economic Disruption
Pathway	Mechanisms and failure modes that lead to outcome-level harms	Deceptive Alignment, Power-Seeking, Authentication Collapse, Trust Erosion
Amplifier	Enabling conditions that increase probability or severity	Racing Dynamics, Multipolar Trap, Sycophancy, Automation Bias

Why this matters:

Addressing amplifiers can prevent multiple pathways
Blocking pathways prevents specific outcomes
Understanding outcomes clarifies what we’re ultimately trying to avoid
Some interventions work at multiple levels simultaneously

All Risks

Filter by level:

54Total

15Catastrophic

29High

18Accident

17Epistemic

8Misuse

11Structural

54 of 54 results

								Solutions
64	Authentication Collapse	Epistemic	Pathway	critical	Medium (emerging)	2025-2030	Emerging	Content Authentication
74	AI Authoritarian Tools	Misuse	Amplifier	high	High (occurring)	Current	Growing	—
62	Automation Bias	Epistemic	Amplifier	medium	Very-high (occurring)	Current	Mature	AI-Human Hybrid Systems
67	Autonomous Weapons	Misuse	Outcome	high	High (occurring)	Current	Mature	—
82	Bioweapons Risk	Misuse	Outcome	catastrophic	Medium (emerging)	2025-2030	Growing	AI Evaluations
25	Concentration of Power	Structural	Outcome	high	Medium-high	2025-2040	Growing	—
62	Consensus Manufacturing	Epistemic	Pathway	high	Medium (occurring)	2025-2030	Emerging	Epistemic Security Prediction Markets
85	Corrigibility Failure	Accident	Pathway	catastrophic	High	2035	Growing	Corrigibility Scalable Oversight
38	Cyber Psychosis	Epistemic	Outcome	medium-high	Medium (emerging)	2025-2030	Neglected	—
72	Cyberweapons Risk	Misuse	Outcome	high	High (emerging)	Current	Growing	AI Evaluations
85	Deceptive Alignment	Accident	Pathway	catastrophic	Medium	2035	Growing	Pause Advocacy AI Control+3
52	Deepfakes	Misuse	Pathway	medium-high	Very-high (occurring)	Current	Mature	Epistemic Security Content Authentication
62	AI Disinformation	Misuse	Pathway	high	Very-high (occurring)	Current	Mature	Epistemic Security Content Authentication
78	Distributional Shift	Accident	Amplifier	medium	Very-high (occurring)	Current	Mature	—
25	Economic Disruption	Structural	Outcome	medium-high	High	2030	Growing	—
82	Emergent Capabilities	Accident	Amplifier	high	Medium (occurring)	Current	Growing	AI Evaluations
67	Enfeeblement	Structural	Outcome	medium-high	Medium	2030	Neglected	AI-Human Hybrid Systems
25	Epistemic Collapse	Epistemic	Outcome	high	Medium-high	2030	Neglected	Epistemic Security Epistemic Infrastructure
25	Erosion of Human Agency	Structural	Outcome	medium-high	High	2030	Neglected	AI-Human Hybrid Systems
25	Expertise Atrophy	Epistemic	Pathway	high	Medium	2025-2050	Neglected	AI-Human Hybrid Systems
68	Flash Dynamics	Structural	Amplifier	high	Medium-high	Current	Neglected	Coordination Technologies Prediction Markets
42	AI-Powered Fraud	Misuse	Outcome	high	Very-high (occurring)	Current	Growing	Content Authentication
82	Goal Misgeneralization	Accident	Pathway	high	High (occurring)	2025-2030	Growing	Interpretability
42	AI-Enabled Historical Revisionism	Epistemic	Pathway	high	Medium (emerging)	2025-2040	Neglected	Epistemic Security Epistemic Infrastructure
75	Institutional Decision Capture	Epistemic	Outcome	high	Medium (emerging)	2025-2040	Emerging	—
92	Instrumental Convergence	Accident	Pathway	high	High (theoretical)	2035	Mature	Corrigibility
62	Irreversibility	Structural	Amplifier	critical	Medium	2030	Growing	—
62	AI Knowledge Monopoly	Epistemic	Outcome	critical	Medium	2030-2050	Neglected	Epistemic Infrastructure
62	Epistemic Learned Helplessness	Epistemic	Outcome	high	Medium (emerging)	2030-2050	Neglected	AI-Human Hybrid Systems
52	Legal Evidence Crisis	Epistemic	Outcome	high	Medium (emerging)	2025-2035	Neglected	—
82	Lock-in	Structural	Outcome	catastrophic	Medium	2030-2045	Growing	Pause Advocacy
78	Authoritarian Takeover	Accident	Outcome	catastrophic	Medium	2025-2050	Growing	—
82	Mesa-Optimization	Accident	Pathway	catastrophic	Medium (theoretical)	2035	Growing	AI Control Interpretability
82	Multipolar Trap	Structural	Amplifier	high	Medium-high	2030	Growing	Coordination Technologies
85	Power-Seeking AI	Accident	Pathway	catastrophic	Medium (theoretical)	2035	Mature	AI Control Corrigibility+1
25	Preference Manipulation	Epistemic	Pathway	high	Medium (occurring)	2025-2035	Emerging	—
73	AI Proliferation	Structural	Amplifier	high	High	Current	Growing	Coordination Technologies
82	Racing Dynamics	Structural	Amplifier	high	High (occurring)	Current	Growing	Pause Advocacy Coordination Technologies+1
25	Reality Fragmentation	Epistemic	Outcome	high	Medium (occurring)	2025-2035	Emerging	Epistemic Security
81	Reward Hacking	Accident	Pathway	high	Very-high (occurring)	Current	Mature	Interpretability Scalable Oversight+1
78	Sandbagging	Accident	Amplifier	high	Medium	2025-2030	Emerging	AI Control AI Evaluations
85	Scheming	Accident	Pathway	catastrophic	Medium	2035	Emerging	AI Control AI Evaluations+1
62	Scientific Knowledge Corruption	Epistemic	Outcome	high	Medium (occurring)	2024-2035	Emerging	Epistemic Infrastructure
85	Sharp Left Turn	Accident	Pathway	catastrophic	Medium	2035	Emerging	—
64	AI Mass Surveillance	Misuse	Outcome	high	Very-high (occurring)	Current	Mature	—
64	Sycophancy	Accident	Amplifier	medium	Very-high (occurring)	Current	Growing	Scalable Oversight
62	Epistemic Sycophancy	Epistemic	Amplifier	medium-high	Medium (occurring)	2025-2030	Emerging	Epistemic Security
82	Treacherous Turn	Accident	Pathway	catastrophic	Medium (theoretical)	2035	Mature	Pause Advocacy AI Control+1
62	Trust Cascade Failure	Epistemic	Pathway	critical	Medium (emerging)	2025-2040	Neglected	—
25	Trust Decline	Epistemic	Pathway	medium-high	High	Current	Growing	Epistemic Security Epistemic Infrastructure
64	Winner-Take-All Dynamics	Structural	Amplifier	high	High	Current	Growing	—
—	Autonomous Replication	Accident	—	High	Medium (emerging)	2025-2030	Emerging	—
—	AI-Enabled Cyberattacks	Accident	—	High	High (occurring)	Current	Mature	—
—	AI-Enabled Biological Risks	Accident	—	Catastrophic	Medium (emerging)	2025-2032	Growing	—

Observable vs Theoretical

Currently Observable	Emerging	Theoretical/Future
Sycophancy	Sandbagging	Scheming
Reward Hacking	Disinformation at scale	Treacherous Turn
Specification Gaming	Deepfakes	Sharp Left Turn
Concentration of Power	Emergent Capabilities	Lock-in
Reality Fragmentation	Institutional Capture	Power-Seeking AI

Understanding observable failures helps us reason about future risks, though the relationship between current problems and future catastrophic risks is debated.

How Categories Interact

These categories aren’t independent:

Accident + Misuse: Misuse is more dangerous when AI is more capable; accident risks determine capability levels
Structural + Accident: Racing dynamics (an amplifier) make accidents more likely by reducing safety investment
Epistemic + All: If we can’t agree on what risks exist, coordinating responses is impossible
Structural + Misuse: Concentration of power determines who might misuse AI; proliferation determines who has access

Explore by Tag

+642