On the Limitations of Compute Thresholds as a Governance Strategy

Compute thresholds are an unreliable governance tool because the relationship between computational power and AI risk is highly uncertain and constantly evolving, making it difficult to predict which abilities will emerge at different scales.

Current policies may be based on flawed assumptions: Both the White House Executive Orders on AI Safety and the EU AI Act use compute thresholds to identify risky systems, but this approach "overestimates our ability to predict what abilities emerge at different scales," according to the paper's key conclusion.

The "bigger is always better" debate remains unresolved: The paper traces how decades of computer science debate about whether increased compute automatically leads to progress is still unsettled, yet policymakers have adopted compute thresholds without resolving this fundamental question.

Compute thresholds risk becoming outdated quickly: Since the relationship between compute and risk is "rapidly changing," regulatory thresholds set today may fail to capture tomorrow's actual risks or may unnecessarily restrict systems that don't pose the predicted dangers.

Better governance approaches are needed: The paper recommends moving beyond compute-based thresholds toward alternative strategies, though specific alternatives would require consulting the full recommendations section for implementation details.