Effectiveness of generative AI regulation methods

Determine which regulatory methods for generative artificial intelligence are effective in practice, identifying policy instruments that reliably achieve intended outcomes for systems such as ChatGPT.

Background

The paper investigates one specific regulatory approach—geographic restrictions on access to ChatGPT—and finds that such restrictions did not reduce use in scientific preprints and were often associated with higher usage in restricted countries.

Despite this domain-specific evidence, the broader question of which regulatory methods work across contexts remains unresolved, highlighting a need for rigorous, empirical evaluation of policy instruments for generative AI.

References

Regulating AI is a key societal challenge, but which regulation methods are effective is unclear.

— Where there's a will there's a way: ChatGPT is used more for science in countries where it is prohibited  (2406.11583 - Bao et al., 2024) in Abstract

Whether license-based use restrictions can effectively prevent model misuse remains uncertain. Users who intend to misuse a model may not comply with license restrictions. At the same time, broad or ambiguous prohibitions may require users who seek to comply with the license to determine whether their intended uses are permitted, placing an additional burden and potentially discouraging research or socially beneficial applications.

— LLM-jp-4 Technical Report  (2610.19331 - LLM-jp et al., 14 Oct 2026) in Section Model License