Notes
Intelligence Gathering
We expect all parties would make ongoing efforts to independently determine whether any actor is conducting dangerous AI activities, out of interest in their own security. A range of state intelligence gathering activities would supplement and validate monitoring the ISIA conducts directly (as described in Articles IV through VII). Towards that end, an Information Consolidation division is vital, and must be trusted to receive information from all parties.
Towards that end, confidentiality is vital, and must be sufficiently robust to assure state intelligence services that the risks imposed on their intelligence methods are minimal, and are justified in order to provide needed information to the ISIA. Avoiding collecting sensitive information whenever possible, and keeping the collected information in the strictest confidence, minimizes risks of compromise.
Article X also addresses the surveillance of non-signatories, where the need for intelligence is strong.
Article X stops short of imposing an obligation to surveil. It would be unprecedented to mandate the creation of a self-sufficient intelligence gathering capability within the ISIA at the required level of capability to give states assurance, and such it seems unnecessary in light of the fact that the creation of superintelligence would pose a grave security threat, which means all parties are already strongly incentivized to surveil and monitor any actor with that capability. Thus, the ISIA relies on parties to provide key intelligence
Whistleblower Protections
The overall effectiveness of this treaty relies on parties’ justified confidence that other parties are not undertaking prohibited AI activities. Even with National Technical Means and other intelligence gathering, it may be difficult for states to detect clandestine efforts to develop superintelligence. There are many domains in which it may not be feasible for states to gather intelligence on their rivals, such as efforts conducted inside military facilities. Whistleblowers can serve as an additional source of information, and the possibility of whistleblowing provides further deterrence against non-compliance.
Whistleblowers may be effective because individuals involved in secret treaty violations (e.g., clandestine training runs or AI research) may themselves be concerned about the danger from ASI. This article aims to make it safer and less costly for them to report violations, shifting the personal incentives away from silence and toward disclosure.
Whistleblowers could sound the alarm for violations of the treaty including:
- Article IV: Training runs that are unmonitored, exceed thresholds, or use prohibited distributed training methods.
- Article V: The existence of undeclared chip clusters, the failure to consolidate all covered hardware, or the diversion of chips to secret, unmonitored facilities.
- Article VI: New manufactured AI chips diverted away from monitoring, or created without mandated security features.
- Article VIII: Prohibited AI research.
Modifications to the whistleblower clauses could change its efficacy and political viability in various ways. For example, states could offer to financially compensate legitimate whistleblowers to provide additional incentives, but this may be seen as paying citizens to defect on their own countries.
Challenge Inspections
Challenge inspections are a critical function provided by the ISIA. Without the credible threat of detection, parties may fear that their rivals would attempt to cheat the treaty (despite the lose-lose nature of a race to superintelligence). Intelligence gathering is one method to combat apparent (illusory) incentives to defect.