Core open questions in frontier AI regulation

Investigate and resolve core open questions in frontier AI regulation by defining the regulatory target for frontier AI systems, identifying effective enforcement measures, and designing new models of supervision for frontier AI developers.

Background

The authors conclude that foundational aspects of frontier AI regulation remain unsettled. These include how to precisely define the regulatory target (material, personal, territorial, and temporal scope), which enforcement mechanisms are most effective, and what supervisory models are appropriate for overseeing frontier AI developers.

Addressing these questions is essential for building a coherent regulatory regime that mitigates risks while supporting innovation, and for tailoring approaches to the legal and institutional contexts of specific jurisdictions.

References

There are numerous open questions, ranging from defining the regulatory target over effective enforcement measures to new models of supervision.

— From Principles to Rules: A Regulatory Approach for Frontier AI  (2407.07300 - Schuett, 2024) in Section V. Conclusion

Despite their growing adoption, the real-world effectiveness of these frameworks remains largely untested, and it is unclear whether such commitments would hold under pressure.

— Toward a Threat Actor Profiling Taxonomy for Pre-Release Risk Management of Open-Weight Frontier Models  (2608.25361 - Zhang, 26 Aug 2026) in Chapter 2, Section “Current Risk Management Practices”

Whether the deliberative density of the three-level regime can be reproduced in this flatter structure is an open question; the meaning of regulatory learning is less clear when there is no intermediate institution.

— MARLA: A Conceptual Scaffold for Regulatory Learning under the EU AI Act  (2609.04877 - Buscemi et al., 4 Sep 2026) in Section Discussion, paragraph beginning “A limitation deserves to be acknowledged”

One candidate substitute is the scientific panel of independent experts, which may issue qualified alerts to the AI Office concerning risks arising from general-purpose AI models; whether an advisory body without supervisory powers can discharge the mediating function that national competent authorities perform remains untested.

— MARLA: A Conceptual Scaffold for Regulatory Learning under the EU AI Act  (2609.04877 - Buscemi et al., 4 Sep 2026) in Section Discussion, paragraph beginning “A limitation deserves to be acknowledged”

even this framework, however, explicitly scopes itself to training and notes that inference-side enforcement is an open problem.

— Beyond Training: A Feasibility Taxonomy for Inference-Time AI Governance  (2609.10105 - Ansari, 9 Sep 2026) in Section 2.1, Inference and existing compute thresholds