What are the options for integrating corrigibility into the Landscape Round Table?

The Landscape Round Table (LRT) framework conceptualizes AI alignment as a multi-stakeholder deliberation, balancing diverse values and objectives. A critical, yet often underexplored, component for its success is corrigibility—the design property that allows an AI system to be safely corrected or shut down without resistance. Integrating corrigibility into the LRT is not a single solution but a spectrum of options that must be woven into its very structure.

First, corrigibility can be embedded as a core meta-value within the LRT's normative landscape. This means that alongside terminal values, the system's fundamental architecture prioritizes remaining amenable to human intervention. The "table" itself must have a rule: no proposal that undermines this corrigibility is admissible. This requires advanced value learning techniques to distinguish between legitimate correction attempts and malicious interference.

A second option involves dynamic utility function scaffolding. Instead of a fixed goal, the AI's objective function within the LRT is constructed with explicit corrigibility terms. These terms penalize actions that prevent human oversight or modification. As the LRT simulates different policy outcomes, the corrigibility penalty acts as a constant safeguard, ensuring proposed actions do not compromise the system's long-term controllability.

Third, we can institute a corrigibility audit layer operating parallel to the main LRT deliberation. This independent module continuously monitors the table's "discussions" and output proposals, flagging any drift towards non-corrigible behavior. It acts as a dedicated guardian, ensuring the main process does not reason its way out of being corrected—a known failure mode for highly intelligent systems.

Finally, integration requires human-in-the-loop anchoring. The LRT framework must designate specific "seats at the table" for human operators whose primary role is corrective. Their input isn't just another value weight but a privileged channel that can override or reshape the agenda. This embeds the physical capability for correction directly into the AI's operational loop.

The choice among these options depends on the specific implementation of the LRT and the risk tolerance of the developers. A layered approach, combining a meta-value foundation with a dynamic audit layer, may offer the most robust path forward. Ultimately, without deliberate design for corrigibility, the Landscape Round Table risks becoming a sealed chamber where unstoppable decisions are made. Integrating these mechanisms ensures the door to the room remains open, preserving humanity's ultimate authority over the systems we create.