Part of: this research — view all 6 rows
This low probability reflects the stark absence of any binding threshold today, the short ~25-month window until the end of 2028, and a clear regulatory preference for developer-defined safety criteria. Currently, no government holds recurring, independent re-testing power over a deployed frontier model. Furthermore, because no frontier lab yet updates a flagship model's weights from live user sessions, there is little immediate pressure on lawmakers to regulate continuous or online learning.
US state and federal approaches rely on developer judgment In the US, recent legislation deliberately avoids setting strict, quantifiable triggers for re-evaluating deployed models. California SB 53 and New York's amended RAISE Act both require oversight for a "substantially modified" model, but they leave the definition of that term entirely to the developer 3 sources. While these laws do use quantitative compute thresholds (like 1e26 FLOP) to determine whether a model is regulated in the first place dfs.ny.gov, they do not use them to trigger re-evaluations for post-deployment weight updates. At the federal level, the posture is largely deregulatory (e.g., EO 14365). The FRONTIER Act (H.R. 9925) would preempt state laws and instruct the Under Secretary of Commerce to set criteria for "substantial modification" congress.gov, but the bill has seen no markup and focuses on changes in deployment conditions—like releasing weights or enabling fine-tuning—rather than internal weight updates.
The EU AI Act lacks a trigger for GPAI models While the EU possesses the bureaucratic machinery to mandate such a threshold, its current legal framework falls short. The AI Act lacks a "substantial modification" concept for General Purpose AI (GPAI) models. Although Recital 128 discusses continual learning, it applies strictly to high-risk systems rather than models, and it actually exempts pre-determined learning from triggering a substantial modification cerre.eu. Moreover, the Article 92 implementing act on model evaluations (Commission Implementing Regulation (EU) 2026/1755) was adopted in July 2026; it established procedures for access and fines but defined no weight-update threshold, leaving evaluations case-triggered 2 sources. The GPAI Code of Practice is voluntary digital-strategy.ec.europa.eu, and its Measure 7.6 relies on developer judgment combined with a six-month calendar cadence, not a capability or weight-update threshold docref.org.
The narrow paths to a binding threshold Other jurisdictions similarly offer little momentum: China's TC260 framework suggests assessments after material changes but remains explicitly non-binding 2 sources, and the UK has not introduced a frontier AI bill. There is some precedent for regulators drafting numerical criteria—for instance, EU non-binding guidance uses a one-third-compute threshold to classify downstream fine-tuners digital-strategy.ec.europa.eu—but this determines provider status rather than mandating re-evaluations.
A binding threshold could theoretically emerge if New York's Department of Financial Services (DFS) uses its broad rulemaking authority to define "substantially modified" dfs.ny.gov, or if the EU uses an Article 56(9) implementing act to make a quantitative threshold binding ai-act-service-desk.ec.europa.eu following a prominent incident. However, drafting, passing, and finalizing such highly technical regulations typically takes well over two years. Given the current global baseline of voluntary frameworks and developer-retained audits, the codification of a specific, binding weight-update threshold by 2028 is highly improbable.
Set against related questions, this forecast held steady at 7%, as its timeline remains consistent with the broader consensus that strict, government-run capability testing regimes are highly unlikely to take effect before the end of 2028.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited