probability: 24%

rationale:

Current Regulatory Landscape

As of August 2026, no jurisdiction restricts the use of AI to conduct AI research. Existing and pending frameworks uniformly regulate AI models as products, services, or deployments rather than as research instruments. The EU AI Act's general-purpose AI regime applies to models placed on the market and generally excludes pre-market research and development 2 sources. In the US, state-level efforts like California's SB 53 and Illinois's SB 315, as well as the proposed federal FRONTIER Act, extend transparency, auditing, and risk-assessment duties to internal use, but they do not impose capability caps, model-age limits, or deployment-history criteria that strictly dictate which models a developer may use for its own R&D 3 sources.

Momentum and Conceptual Readiness

While binding laws do not yet exist, the policy target is becoming highly legible. Major frontier safety frameworks from Anthropic, OpenAI, and Google DeepMind now explicitly define AI-R&D automation thresholds, such as OpenAI's "Critical" track for fully automated R&D cdn.openai.com and Anthropic's automated R&D threshold anthropic.com. Advocacy has also recently shifted toward pacing internal research. The July 2026 "Pacing the Frontier" letter, signed by over 1,300 industry figures, calls for international tools to deliberately pace automated AI development pacingthefrontier.com. Furthermore, the AI Futures Project has proposed concrete rules, such as restricting R&D to models trained at least nine months earlier blog.aifutures.org. If governments seek to operationalize these concepts into law, the blueprints already exist.

Structural Headwinds and Enforcement Constraints

Despite this conceptual momentum, translating voluntary frameworks into binding state bans faces severe headwinds. Constraining the models used for internal research directly targets a developer's most valuable input for technological advancement, running directly against the competitive and geopolitical incentives of the US, China, and allied nations. Moreover, enforcing such a regime requires intrusive verification mechanisms—such as inference monitoring, rigorous code review, and embedded third-party auditors—which are currently immature and politically fraught 2 sources. Prominent policy groups, such as the Institute for Progress, explicitly recommend against legal mandates restricting AI-R&D automation, favoring transparency and capability-building instead ifp.org.

Pathways to a Binding Regime by 2031

For a qualifying restriction to materialize within the next five years, it would likely require event-driven escalation. A severe safety incident traced to internal AI-assisted research, or a sudden, undeniable acceleration in AI capabilities driven by recursive self-improvement, could rapidly overcome current deregulatory preferences. Additionally, there is some ambiguity in how broad legislative language might be interpreted; for example, if the FRONTIER Act's proposed emergency authority to restrict the "development, deployment, or internal use" of a model congress.gov were enacted and invoked conditionally, it could plausibly satisfy the broad "any other criterion" clause of this definition. Because a qualifying regime must regulate models strictly as research instruments—an unprecedented and highly intrusive policy object—the likelihood of implementation remains modestly low. Overall, my estimate stands at 24%, explicitly anchored as the cumulative probability of both early restriction-first scenarios and later adoption following AI automation milestones .