Session context: A member raised a pointed question during the AI council discussion: is there a vetting process for selecting council members that accounts for blind spots — ensuring the assembled perspectives complement rather than compound each other? Lou gave an honest answer that surfaced the structural problem underneath the question.
Core Idea
When you build an AI advisory board, you are simultaneously the architect and the subject under review. You choose which expert lenses to assemble. You write the briefs that define each persona’s perspective. You decide which domains to represent.
The bias this creates is invisible: you populate the council with perspectives you already find credible. The board reflects your existing assumptions back at you, but now with the added authority of multiple “expert” voices in apparent agreement. It feels like getting diverse input. What it actually is: your own mental models in different costumes.
This isn’t a flaw in the AI council concept — it’s a structural property of any self-assembled advisory system. Human advisory boards have the same problem. The founder who built the board rarely recruited people who fundamentally challenge their worldview.
The fix is a composition audit conducted before the council meets, not after. Ask:
- Who is NOT on this council? Which domains and perspectives are absent?
- Which of my assumptions does every member share? If your entire council believes X, X is invisible to the council.
- What adjacent domain might see this differently? A board of marketing experts has no friction from an operations lens. A board of coaches has no friction from a client-side perspective.
- Is there a moderator role? Someone whose job is to surface gaps, not integrate outputs.
Lou’s answer to the implementation question was precise: the structural solution is a moderator agent whose job is to identify what the council is not considering, not to synthesize what it did. That’s a fundamentally different role from the expert personas — it requires stepping outside the question and asking what’s missing from the frame.
Practical Application
Add a Meta-Moderator to any AI council configuration. Prompt it with:
“You are not a domain expert. Your job is to audit the perspectives above for blind spots. What expert lens is absent from this council? What assumption does every member share that has not been examined? What is the most important question this council has not asked?”
Run this after the council members weigh in, before you synthesize their outputs. The moderator’s output is often more valuable than the council’s — because it names what the frame excluded.
For a quick composition audit before assembling any council: list your intended members, then list the three domains most likely to disagree with your approach. If none of them are on the council, recruit one.
Evolution Across Sessions
Multi-Model Debate established that using competing models prevents sycophancy. Data-Gated Pushback addressed breaking agreement loops by requiring evidence. This insight goes upstream — to the composition of the council before it convenes. The problem isn’t that AI agrees too readily; it’s that you built a council pre-loaded to agree on your core assumptions. The composition is the blind spot.