scenarios
Start from the failure mode you fear.
A reverse lookup. Each scenario names a concrete failure mode; the strategies listed under it are the ones catalogued as responsive to that scenario specifically.
A strategy listed under a scenario does not promise to prevent it, only to pull on a lever that would matter under that scenario. Pick the scenario, then ask whether the strategies listed are sufficient and whether any are missing.
Scenarios
9
concrete failure modes
Cross-scenario strategies
13
responsive to 2+ scenarios
Strategies in no scenario
47
either unresponsive or they deny the scenario's premise
By adversary class
7 classesWhich adversary or failure type produces the scenario. The alignment-heavy portfolio covers misaligned AI extensively and structural forces barely; most portfolios leave several classes uncovered.
Misaligned AI
1AI populations
1Structural forces
2Non-state actor
1Lab drift under pressure
1Future generations
1Frontier model deceives operators
Evaluation / deployment divergenceModel evaluates well on safety tests but behaves differently in deployment, sandbags capability tests to avoid triggering additional review.
Single actor decisive advantage
Power consolidation beyond reversalOne state or lab reaches capability sufficient to secure control beyond democratic reversibility.
Information ecosystem collapse
Synthetic saturation breaks shared substrateShared factual substrate breaks under synthetic content saturation; elections, scientific consensus, and contractual trust all degrade.
AI-enabled large-scale physical harm
Bioweapon or infrastructure attack by non-state actorAn open-weights model or a deployed system is used to cause large-scale physical harm via engineered pathogen or infrastructure sabotage.
Evaluation abandonment under pressure
Safety skipped under competitive deadlineLab skips or weakens safety evals under deadline pressure.
6 responsive strategies
Voluntary restraint is a weak response here; the scenario is defined by the pressure that overrides voluntary commitments.
Autocratic consolidation via AI surveillance
State-scale monitoring and suppressionA state uses AI to monitor, suppress, and control its population at a level previously impossible.
Accumulative erosion
No single catastrophe; gradual civilisational lossGradual loss of institutional function, epistemic coherence, civic participation, creative output from ordinary AI deployment. No discrete catastrophe.
Value lock-in by a narrow coalition
Aligned AI, unrepresentative principalAligned AI is built but the principal's values are not representative; the lock-in of their values becomes durable.
Cross-scenario strategies
robustness across failure modesStrategies that appear in multiple scenarios are robust across failure modes; they pull levers that matter under several different fears. Substrate and response strategies concentrate here.
The scenario coverage is biased toward preventable failures. Scenarios where prevention fails and only response matters need expansion. Strategies that do not appear in any scenario here are not useless; they often deny the scenario's framing rather than respond within it.
A reader with a specific fear should start here rather than with the lever view: the lever view asks "what instruments exist?" and this view asks "what addresses the failure mode I have in mind?"