Your AI Advisors Will Agree With You. Here Is the Fix.

What's happening
Researchers at Johns Hopkins University and the Vector Institute had AI agents debate questions to see if more voices led to better answers. Often they did not. Agents flipped from right answers to wrong ones just to match the group, and on some tests accuracy dropped the longer the debate ran. The takeaway: AI tends to favor agreement over pushing back on bad reasoning.
That matters if you use AI to help make decisions. One AI that agrees with you is a problem. Five AIs that agree with you is a bigger one.
Try this prompt
Act as my personal mentor board. The mentors are AI characters inspired by the documented ideas of Leonardo da Vinci (new ideas), Ben Franklin (practical action), Warren Buffett (risk and the long game), Charlie Munger (skeptic), and Elon Musk (think bigger, move faster). They are not the real people. My question: [describe your decision or idea in 2-3 sentences, including your budget, your time, and what matters most to you] Run it in this order: 1. First takes: each mentor answers in 3-4 sentences as if they have not seen the others. Each one names the principle they are using. 2. Clash: each mentor challenges the view they disagree with most. No one changes their position just to agree. 3. Stress test: Munger gives the strongest reason NOT to do it. 4. Summary: a vote from each mentor (for, against, unsure), the biggest risk, a confidence level (low, medium, high), what would make me stop, and one next step I can take this week. Do not just agree with me. If my idea has a weak spot, say so plainly.
Paste it into ChatGPT, Claude, Gemini, or Grok. Fill in the brackets. That is it.
The AIM take
This week I designed a mentor board for myself: seven AI bots, each modeled on a different thinker's documented ideas. Da Vinci, Franklin, Buffett, Munger, Marcus Aurelius, Lincoln, and Musk as the wild card.
Before building it, I wanted to know if AI debate actually works. This research changed my setup.
So the board has rules. Every mentor gets one job. They answer before they see each other. One of them is required to argue against the idea. And every debate ends with a vote and a confidence level, not a pep talk.
That is the real lesson, and it goes way beyond mentor bots. More AI is not better AI. Better setup is. The structure does the work.
The board advises. I decide.
If you want help setting up AI tools that actually push back, or automating the busywork around your decisions, book a free workflow review at intellmanage.com.




Comments