Pause AI vs Accelerate Capabilities
pause-vs-accelerate · AI-governance meta-debate
Should the global community impose temporary or capability-conditional pauses on frontier-AI development, or should development accelerate with safety work conducted in parallel?
Why it matters
Most-public manifestation of the catastrophic-risk debate. Shapes which regulatory architecture is even on the table: a pause-favouring policy stance (e.g., EU AIA's risk-tiering with prohibited practices) differs structurally from an accelerate-with-safeguards stance (e.g., NIST AI RMF voluntary framework).
Positions (3)
Catalogued in editorial order; not ranked. Each position carries its own primary sources.
Position 1
Pause or capability-conditional moratorium
Frontier-AI development outpaces alignment research; a temporary pause (FLI Mar 2023 open letter: 6 months on training systems more powerful than GPT-4) or capability-thresholded moratorium would create breathing room. Without such intervention, catastrophic-risk probability grows.
Proponents
Primary sources
Position 2
Accelerate with adaptive safeguards
A pause is unenforceable (China + other non-signatories continue), economically costly, and disincentivises the very safety research that justifies it. The right posture is accelerated development PLUS adaptive safeguards (RSP-style, capability-tiered) that scale with model capability.
Proponents
Primary sources
Position 3
Continue, but cap capability via compute / behavioral threshold
Neither pause nor unrestricted acceleration. Use compute or behavioural-capability thresholds (EU AIA 10²⁵ FLOPs, CA-SB-1047 10²⁶ FLOPs, behavioural-eval-based caps) to gate the most-capable models without halting development broadly.
Proponents
Primary sources
Instruments shaped by this debate
Topics this debate touches
Editorial note
FLI Mar 2023 pause-letter was the most-public articulation; CAIS May 2023 'Statement on AI Risk' (one sentence, lab-CEO signatories) was the moderating reply. By Bletchley + Seoul (2023-2024), the de facto policy convergence is capability-thresholded gating — pause + accelerate framings remain in discourse but are no longer the operative regulatory choices.