Compute threshold

In plain English

Because the most capable models are trained with the most compute, lawmakers use the total number of operations in a training run as a simple trigger for safety duties. The EU AI Act presumes that a general purpose model trained with more than 10^25 FLOP has systemic risk, which brings duties to evaluate the model, report serious incidents and secure it. California's SB 53, New York's RAISE Act and Illinois SB 315 use 10^26 operations to define frontier developers, and South Korea's AI Basic Act sets a threshold by decree, reported as 10^26 FLOP.

Thresholds are easy to measure and hard to dodge for the largest runs, which is their appeal. They are also imperfect. Better algorithms reach the same capability with less compute every year, so a fixed threshold covers fewer dangerous models over time, and tests of actual capability are needed alongside it. Several laws therefore let regulators review or update the number.

Why it matters for rogue SI

Thresholds decide which models get tested and watched at all. A model capable of going rogue but trained efficiently enough to fall under the line would slip past the rules meant to catch it.

Laws and rules that use it

• In force EU AI Act (GPAI) European Union
• In force SB 53 California
• Passed, not yet in force RAISE Act New York
• Passed, not yet in force SB 315 Illinois
• In force AI Basic Act South Korea

Sources

1. Guidelines for providers of general-purpose AI models European Commission · Jul 18, 2025
2. SB 53: Artificial intelligence models: large developers CalMatters Digital Democracy