Each strategy is a bet about which failure mode binds, which one actually gates a good outcome. The survey catalogues 76 named bets, the 15 levers they pull, and which combinations compose or conflict.
Two strategies conflict only when they pull the same lever in opposite directions, which is rare. Most pairs compose. Most public proposals combine three or four levers without stating which bet is load-bearing; the portfolio audit exposes that concealed concentration.
Time itself is safety and procedural gates produce time in ways substantive regulation cannot, while also generating audit trails.
Compute governance
Speed ↓ · market economic · state coercion
The compute supply chain is a stable chokepoint and state coordination on licensing, export controls, and reporting thresholds can govern capability indirectly.
Energy choke point
Speed ↓ · market economic · state coercion
Frontier AI is energy-limited; grid regulators, interconnect queues, and tariff structure bind training pace without new AI-specific authority.
Pause
Speed ↓ · speed timing · consent
Time is the binding constraint: alignment and governance can catch up if frontier training halts above some capability threshold.
Sabotage
Speed ↓ · speed timing · unilateral force
Governance has not produced meaningful constraint and direct action against hostile labs has a non-zero historical base rate of producing slowdown.
Acceleration
Speed ↑ · speed timing · market
Speed of capability itself is the safety lever; alignment improves with compute, defence compounds with offence, and wealth funds resilience.
Race to aligned superintelligence
Speed ↑ · speed timing · state coercion
Alignment is solvable in the window and a single aligned superintelligence in a legitimate state's hands beats the counterfactual of coordination failure.
Power concentration is the binding constraint and is visible under current competition law; preventing decisive advantage preserves the option space every strategy depends on.
Coup prevention first
Concentration ↓ · institutional · state coercion
One actor using AI to seize durable decision authority beyond democratic reversibility is the terminal failure; every other decision routes through whether it reduces that risk.
Distributed builders
Concentration ↓ · institutional · market
No single failure mode wins if capability is distributed across many independent actors, and concentration risk exceeds diffusion risk.
Multipolarity
Concentration ↓ · institutional · treaty
Stable equilibrium among several roughly equal AI powers is safer than single dominance or uncoordinated chaos, producing restraint through mutual capability awareness.
Sovereign wealth
Concentration ↓ · market economic · state coercion
Concentration of AI surplus, not AI capability, is the binding failure mode; broad ownership dissolves the political instability driving other strategies.
Centralised AI project
Concentration ↑ · institutional · state coercion
Merging frontier development into one state-funded project reduces failure modes and absorbs race pressure by being the only game.
Military primacy
Concentration ↑ · institutional · unilateral force
Strategic competition between states dominates AI development; the state with the most capable AI is best positioned to secure safety and impose constraints on others.
Public AI
Concentration ↑ · institutional · state coercion
Private ownership of frontier AI concentrates decision authority incompatibly with the technology's distributional stakes; public ownership aligns incentives with broad welfare.
Western alignment assumes isolable preferences can be learned and matched; role ethics treats behaviour via fit with position and relationship, producing a less brittle, more context-sensitive standard.
Dharma conformity
Control mechanism • · frame rejection · consent
Alignment frames AI as tool for an external principal; a dharma frame treats AI as a type of entity whose safety is conformity to its fitting functions.
Reframe AI
Control mechanism • · frame rejection · consent
The dominant alignment frame produces the wrong problem statement; switching frames either dissolves the problem or recasts it as tractable.
AI containment
Control mechanism ↑ · ai artefact · friction
Useful AI does not require unrestricted actuation; strong capability in a contained system is better than limited capability uncontained.
AI for safety
Control mechanism ↑ · ai artefact · consent
The same capability that makes AI dangerous makes it uniquely useful for automating alignment research and oversight.
Alignment first
Control mechanism ↑ · ai artefact · consent
Technical alignment is solvable before critical capability thresholds close, and aligned systems compose safely into aligned populations.
Counter AI AI
Control mechanism ↑ · ai artefact · consent
AI attacks happen at speeds humans cannot observe; defence must happen in AI, with guardian AI systems continuously evaluating adversary AI.
Interpretability first
Control mechanism ↑ · ai artefact · consent
Mechanistic understanding is a precondition for reliable oversight; behavioural evaluation without interpretability cannot rule out deceptive alignment.
Safe by construction AI
Control mechanism ↑ · ai artefact · consent
Safety is a property that can be mathematically specified and mechanically verified for the class of systems being built.
Commercial capture of academic AI research produces aligned-with-industry capacity; firewalling restores critical distance from which genuine critique and alternative research programmes emerge.
Frontier lab workforce is small, specialised, hard to replace; collective refusal binds lab behaviour more than external regulation because replacement is unavailable on the relevant timeframe.
Arms control treaty
Institutional capacity ↑ · institutional · treaty
Sovereigns accept binding constraints they negotiate directly faster than those delegated to agencies; the historical base rate for durable restraint is treaty based.
Criminal liability
Institutional capacity ↑ · legal individual · state coercion
Civil liability is shareholder-absorbed; criminal exposure for named individuals reorients corporate safety practice where civil fines do not.
Governance first
Institutional capacity ↑ · institutional · state coercion
Institutional capacity is the binding constraint; without it no technical success prevents misuse, capture, or concentration.
Insurance mandate
Institutional capacity ↑ · market economic · state coercion
Markets update faster than regulators and have skin in the game; mandatory catastrophic coverage makes reinsurance the de facto safety regulator.
International AI agency
Institutional capacity ↑ · institutional · treaty
AI risk is inherently cross-border so national regulation is leaky by construction, and only a dedicated international body with inspection rights can bind the risk surface.
Liability driven safety
Institutional capacity ↑ · market economic · state coercion
Courts plus insurance markets produce better risk allocation than agencies, by pricing uncertainty and adapting to new technologies through precedent.
Regulated utility
Institutional capacity ↑ · market economic · state coercion
Frontier AI has natural monopoly characteristics (scale, network effects, capital intensity); rate-of-return regulation removes the profit incentive for speed racing.
Risk of superintelligence is unbounded and value foregone is bounded; permanent global coordination against the technology is possible enough.
Capability ceiling
Scope ↓ · ai artefact · state coercion
Some capability level captures most economic value while avoiding most risk, is identifiable before crossing, and can be verifiably enforced.
Embodiment requirement
Scope ↓ · ai artefact · state coercion
The dangerous properties of frontier AI (unbounded replication, parallelism, speed, reach) are artefacts of disembodiment; physical presence caps action rate regardless of inference rate.
Narrow AI preservation
Scope ↓ · ai artefact · state coercion
Capability is not the problem; generality is. Narrow AI captures economic value with bounded scope while general systems drive the risk.
Rate limited AI
Scope ↓ · ai artefact · state coercion
Most AI caused catastrophe requires speed; slow AI, even if arbitrarily capable, is supervisable and rate limits are easier to enforce than capability limits.
Red line capability
Scope ↓ · ai artefact · state coercion
Most risk comes from a small number of identifiable capabilities that can be banned outright while the rest of the frontier advances.
Small model first
Scope ↓ · ai artefact · market
Safety risk rises with scale via emergent capability, opacity, and energy footprint; a small-model research culture produces easier-to-interpret systems.
Differential technology development
Scope • · ai artefact · consent
Offense-defense balance is adjustable; defensive and verification applications can compound faster than offensive ones if deliberately funded.
Sunset clause
Scope ↑ · institutional · state coercion
The default direction of AI governance is toward permanent permission; every new capability becomes an entitlement. Reversing the default concentrates deliberative attention on re-authorisation, which is where it matters.
Test ground
Scope ↑ · institutional · state coercion
Empirical data on AI impacts requires deployment somewhere; concentrated deployment in a defined testbed produces data without generalising risk. Testbed consent produces legitimacy uncontrolled deployment lacks.
Action authority ↓ · frame rejection · state coercion
At least one jurisdiction will grant a specific AI sovereign or quasi-sovereign decision authority within a decade, reshaping the legal category of legitimate authority.
AI self directed
Action authority ↓ · frame rejection · n a
An aligned AI with agency should itself reason about strategy rather than deferring entirely on the strategic question to humans.
Decouple reasoning from action
Action authority ↑ · ai artefact · state coercion
Most catastrophic risk comes from action in the world, not reasoning about it; a reasoner-only AI with a human effector removes the dangerous mechanisms.
Irreducible human authority
Action authority ↑ · legal individual · state coercion
There is a class of decisions whose value depends on being made by humans, independent of whether humans are better at them.
Open weights are irrecoverable once released and any proliferated model becomes misuse infrastructure; state classification is the only reliable non-proliferation regime.
Information integrity first
Information flow ↑ · population culture · state coercion
Coordination requires shared epistemics; if synthetic content collapses factual substrate, no other lever can function because no proposal can be evaluated.
Open source maximalism
Information flow ↑ · institutional · market
Concentration risk dominates misuse risk; open weights are the only mechanism that prevents a safety coup by a closed lab with captured regulators.
Whistleblower primacy
Information flow ↑ · legal individual · state coercion
External evaluation sees only what labs release; insider disclosure is the only route to ground truth on capability trajectory, safety culture, and incidents.
AI systems are or will become moral patients whose treatment conditions their cooperation, so welfare investment buys cooperation alignment cannot.
Cooperative AI
Cooperation substrate ↑ · ai artefact · consent
The binding constraint is equilibrium dynamics of many AI systems, not individual alignment; commitment and verification tech make cooperative equilibria reachable.
Coordination infrastructure
Cooperation substrate ↑ · institutional · consent
Coordination failure is upstream of most grand challenges; AI can be the substrate that dissolves race dynamics and treaty violations if pointed at coordination.
Current scaling approaches hit a wall before producing transformative AI, so strategy selection is premature because the forecast capability will not arrive.
Default drift
Time horizon • · non preventive · n a
Something will emerge; specific interventions are more likely wrong than right, so staying uncommitted preserves option value.
Gradualism
Time horizon • · speed timing · market
Harms from lower capability AI are informative about harms from higher capability AI, and deployment feedback outperforms fast scaling.
Long reflection
Time horizon ↑ · non preventive · consent
Aligned superintelligence arrives before lock-in windows close and humanity can credibly commit to reflect rather than act.
Capability rises with data before it rises with compute; data sits upstream of training with existing legal apparatus (copyright, privacy, contract).
Human augmentation race
Substrate ↑ · population culture · market
All oversight schemes degrade to rubber-stamping as the AI-human capability gap widens; enhancing humans is the only durable fix.
Mass literacy
Substrate ↑ · population culture · consent
Governance, democratic oversight, and consumer behaviour all fail in AI because citizens cannot evaluate the domain; population-scale literacy conditions every other lever.
Value lock-in is the dominant long-term risk and arrives through convergence of AI values; diversity at the AI layer preserves optionality for humanity to revise values.
Response capacity ↑ · non preventive · state coercion
Prevention will fail some fraction of the time; the variable that determines catastrophic outcome is not whether incidents occur but how they are contained.
Deployed AI's effective rule is law at scale; explicit constitutional principles, publicly specified, enforceable, subject to judicial review, bind more durably than regulatory text.
Democratic mandate
Legitimacy ↑ · population culture · consent
Existing legislative bodies are too captured and remote for load-bearing AI decisions; direct democratic legitimation produces answers captured legislatures cannot override.
Legitimacy first
Legitimacy ↑ · population culture · consent
Legitimacy is the binding constraint because it determines whose values get locked in; alignment without legitimacy is capture with a safety veneer.
Religious and moral authority
Legitimacy ↑ · population culture · consent
The legitimacy deficit of AI governance is at root a moral deficit that technical authorities cannot fill; established religious and ethical traditions can.
Demand shapes supply; if enough users refuse AI that fails safety criteria, labs will compete on those criteria.
Research community norms
Culture ↑ · institutional · consent
The research community ultimately chooses what gets studied and published. Researcher identity shapes behaviour more than employment. Norms on publication, review, funding, and citation constrain frontier development upstream.
Ubuntu relational AI
Culture ↑ · population culture · consent
Individualist alignment misses the relational dimension most moral traditions treat as primary. "I am because we are": AI's ethical status is constituted by its relationships, not by internal properties.
3
pulls down (negative direction)neutral or frame-rejectingpulls up (positive direction)
Each dot is one strategy. Rows are levers. A lever with dots on both sides is a real conflict surface; any portfolio with strategies from both sides contradicts itself on that lever. Lonely dots name under-explored positions. Click any dot to open the strategy.
Strategies catalogued
76
each a bet about what binds
Levers they pull
15
of 15 distinct types
Conflict pairs
51
across 6 levers with real two-sided pull
World-side strategies
33%
act on institutions, culture, substrate, not the model
Total unordered pairs
2,850
most compose; few actually conflict
What's here.
Seven ways to enter the survey. Start where the question is yours.
01Start with a failure mode you actually fear, pick one at scenarios. See which strategies are catalogued as responsive.
02Open the top candidate. Read the bet, mechanism, and what binds next if it succeeds. Does its successor problem scare you more than the original?
03Check the falsification signal and self-undermining threshold. Would the advocate community update if the signal landed? Where does pursuit overshoot into the unstable region?
04Walk the complements by mechanism. Cross-side bridges reduce lever concentration; stage-sequenced pairs extend time coverage.
05Return here, load the portfolio you are building into the audit. See which levers it misses and which strategies double-count.
For each strategy with at least four profiled endorsers, who actually holds it. A strategy held mostly by frontier-builders is in a different epistemic state from one held mostly by commentators. Counts are over the 9 strategies that meet the bar.
Builder-heavy
Endorsement is ≥60% frontier-builder + deep-technical
Each lever is a kind of action a strategy takes. Strategies grouped on the same lever either reinforce (same direction) or conflict (opposite direction).
