Longterm Wiki
Updated 2026-01-28HistoryData
Page StatusContent
Edited 2 weeks ago32 words
12
QualityStub
25
ImportancePeripheral
2
Structure2/15
00000%0%
Updated every 6 weeksDue in 4 weeks
Summary

Presents two core cruxes in the AI x-risk debate: whether advanced AI would develop dangerous goals (instrumental convergence vs. trainable safety) and whether we'll get warning signs (gradual failures vs. deception/fast takeoff). No quantitative analysis, primary sources, or novel framing provided.

Issues1
StructureNo tables or diagrams - consider adding visual content

Is AI Existential Risk Real?

Crux

Is AI Existential Risk Real?

Presents two core cruxes in the AI x-risk debate: whether advanced AI would develop dangerous goals (instrumental convergence vs. trainable safety) and whether we'll get warning signs (gradual failures vs. deception/fast takeoff). No quantitative analysis, primary sources, or novel framing provided.

QuestionDoes AI pose genuine existential risk?
StakesDetermines priority of AI safety work
Expert ConsensusSignificant disagreement
32 words
Crux

AI Existential Risk Debate

QuestionDoes AI pose genuine existential risk?
StakesDetermines priority of AI safety work
Expert ConsensusSignificant disagreement

This is the foundational question in AI safety. Everything else depends on whether you believe AI could actually pose existential risk.

Key Cruxes

What would change your mind on this debate?

Key Questions

  • ?If we built human-level AI, would it naturally develop dangerous goals?
    Yes - instrumental convergence applies

    Power-seeking emerges from almost any goal. Training won't reliably prevent it.

    X-risk is real; alignment is critical

    Confidence: medium
    No - we can train safe systems

    Goals come from training. We can instill safe goals and verify them.

    X-risk is manageable with standard safety engineering

    Confidence: medium
  • ?Will we get warning signs before catastrophe?
    Yes - problems will be visible first

    Weaker systems will fail in detectable ways. We can iterate to safety.

    Can learn from experience; less urgent

    Confidence: low
    No - deception or fast takeoff prevents warning

    Sufficiently capable AI might hide misalignment. Jump to dangerous capability.

    Must solve alignment before building dangerous AI

    Confidence: medium

Related Pages

Top Related Pages

Key Debates

Open vs Closed Source AIGovernment Regulation vs Industry Self-GovernanceIs Interpretability Sufficient for Safety?Should We Pause AI Development?Is Scaling All You Need?When Will AGI Arrive?

Organizations

Alignment Research Center