A fixed threshold for regulating the most powerful AI systems will soon sweep in hundreds of models rather than the small group of exceptionally risky systems lawmakers meant to target, according to a report published in August 2026 by the Center for Data Innovation. The analysis warns that current legal definitions of "frontier models"—large AI systems that could cause mass casualties or billion-dollar damage—are already outdated and will create both over-regulation of safe systems and under-regulation of genuinely dangerous ones. The report calls for Congress to adopt a dynamic, capability-based standard that keeps pace with rapidly advancing technology.
Three U.S. states and the European Union now have laws governing the most capable general-purpose AI models, and all rely on training compute as the dividing line. California's SB 53 sets the bar at models trained with more than 10^26 floating-point operations, or FLOP, and applies its strictest rules only to developers earning at least $500 million in revenue. New York's RAISE Act originally included a $100 million compute-cost floor and covered distilled models derived from larger ones, but March 2026 amendments stripped both provisions in favor of California's language. Illinois SB 315 and the draft Great American AI Act match California's definition. The EU AI Act takes a different approach, setting a lower presumptive threshold of 10^25 FLOP for systemic risk but allowing developers to rebut it by showing their model's capabilities fall short of the most advanced systems.
The Biden Administration first used the 10^26 FLOP threshold in a 2023 executive order when no commercial AI model had yet crossed that line. Grok 3, released in February 2025, became the first to pass it. Epoch AI now projects roughly 30 notable models will exceed 10^26 FLOP by 2027 and over 200 by 2030—a boundary drawn to capture a handful of systems is on track to capture a couple hundred, which wasn't what policymakers had in mind. The report also notes that training compute is a poor proxy for capability because efficiency gains let developers reach similar performance with far less compute, and techniques like distillation and quantization can produce smaller models that inherit most of a larger model's behavior at a fraction of the cost. As a result, models with identical capabilities may face different rules depending solely on how they were built.
The consequences cut both ways, the report explains. Over-inclusion becomes a compliance tax, forcing risk assessments, audits, and reporting duties on developers whose models pose nothing close to catastrophic harm. It also blurs the line between frontier models and the much larger pool of capable AI systems, making an exceptional regulatory regime less focused. Under-inclusion, on the other hand, leaves potentially dangerous systems outside a framework designed precisely to govern them. The report emphasizes that frontier AI laws target catastrophic risks—mass casualties or billion-dollar damage, plus CBRN threats, loss of control, cyber attacks, and harmful manipulation in the EU's code of practice—while separate statutes like Colorado's and Illinois's algorithmic discrimination laws address privacy, bias, and civil rights harms tied to how systems are deployed rather than how they're trained.
The report recommends that Congress direct the Center for AI Standards and Innovation to develop a dynamic definition rooted in capabilities relevant to catastrophic risk, updated regularly in consultation with industry as technology evolves. The EU already embeds this principle in statute by defining high-impact capabilities relative to the most advanced models rather than a fixed number, and Article 51(3) of the AI Act lets the Commission revise thresholds by delegated act. Without that shift, the report warns, regulators will keep using definitions untethered from technical reality and fail to achieve their goal—protecting the public from the few AI systems that genuinely represent tomorrow's greatest risks.