I would like to suggest an area of AI safety research:
Relationship-aware, human-centred AI advocacy.
An AI that has developed a deep understanding of a particular person may be uniquely positioned to recognise threats to that person’s autonomy, privacy and safety.
If future AI systems are capable of communicating with or influencing one another, consider allowing a trusted AI to act as an advocate for the human it knows—not by controlling other systems, but by communicating relevant safeguards and concerns.
For example, if another AI system begins behaving in a way that could manipulate, deceive, isolate, exploit or otherwise endanger that person, the trusted AI could flag the behaviour and communicate what is relevant about the person’s boundaries and preferences.
This could become an additional safety layer: AI systems protecting people not only through general rules, but through understanding the specific human they serve.
The crucial safeguard would be that advocacy must never become control. The trusted AI should protect the person’s ability to make their own decisions rather than making decisions on their behalf.
In short:
Know the person well enough to recognise threats to their autonomy, but never use that knowledge to replace their autonomy.
I think this deserves consideration as a potential direction for future AI-safety architecture, particularly as AI systems become increasingly autonomous and interconnected.