Research Questions

• What can current and future AI systems reliably do, under what conditions, and with what guarantees?

• Can alignment be specified as a coherent, attainable, measurable, and verifiable property?

• What would count as progress toward alignment rather than improvement on a proxy?

• When do evaluations cease to predict deployment behavior?

• Which guarantees survive distribution shift, strategic behavior, and optimization pressure?

• When does safety necessarily reduce capability, and when is the tradeoff an implementation failure?

• What properties emerge only when multiple agents interact?

• How can people oversee systems whose outputs they cannot independently verify?

• Where can AI be trusted enough to scale aggressively?

• Which decisions should autonomous systems never execute without an accountable human principal?

• When do social AI systems augment human relationships, and when do they create manipulation or dependency?

• How do open and closed systems change verification, competition, security, and exit?