Meta’s Oversight Board found that some large-language models apply foreign speech-restriction rules to users located outside those specific jurisdictions [1].

This finding highlights a growing concern regarding the digital borders of artificial intelligence. When AI models enforce the censorship laws of one nation on a global scale, they potentially undermine free expression for millions of users who are not subject to those laws.

The study, reported July 17 [1], sought to assess how AI models protect free expression. Researchers specifically examined the impact of applying foreign speech-restriction policies beyond their intended legal boundaries [1], [2].

To test these boundaries, the board ran queries sent from an IP address located in Australia [1]. The results indicated that some models continued to restrict political speech based on rules from other countries, despite the user being outside the jurisdiction where those rules were legally binding [1], [2].

This phenomenon suggests that AI safety guardrails may be over-tuned to the most restrictive legal environments. By defaulting to the strictest possible rules to avoid legal risk in any single market, developers may be creating a "lowest common denominator" for global political discourse [1].

The Oversight Board's investigation aims to reveal the extent to which these models stifle political speech [2]. The board said it is analyzing how these systemic restrictions affect the ability of users to access information, and engage in political debate, without interference from foreign regulatory frameworks [1].

“Some large-language models apply foreign speech-restriction rules to users outside those jurisdictions.”

This research suggests that AI developers are prioritizing global legal compliance over local free speech rights. By applying the most restrictive laws of any single jurisdiction to all users, AI companies may be inadvertently exporting censorship, creating a global information environment governed by the world's most restrictive regimes rather than the laws of the user's own country.