Stop Asking AI What It Thinks
Why one opinion feels like an answer and three opinions feel like research.
I watched a friend paste a city council transcript into ChatGPT last week and ask: “Should I support this rezoning proposal?”
He got a long, balanced answer that considered fiscal impact, community sentiment, and economic development. It was thoughtful. It was well-structured. It was completely useless.
Not because the answer was wrong. Because he had no idea what he was agreeing with.
Here’s the problem with asking one AI for an opinion: you get one opinion. And it sounds authoritative. And you nod along because it’s articulate and it cited some sources and it organized its thoughts better than you would have.
But you have no idea what it optimized for. Was it prioritizing fiscal responsibility? Community impact? Economic growth? You can’t tell, because it blended everything into one reasonable-sounding response. The seams are hidden.
Now try something different. Instead of asking one model for one balanced take, ask three systems that each care about different things.
One that only thinks about budget impact. Obsessively. Every dollar justified over a 5-year horizon or it’s a no.
One that only thinks about economic growth. Jobs, investment, regional competitiveness. Short-term costs are acceptable if the multiplier math works.
One that only listens to residents. Public comment sentiment, quality of life impact, neighborhood disruption. If 73% of adjacent homeowners oppose it, that’s the signal.
Now show me all three at once.
Something interesting happens when you see three opinionated answers side by side: you start thinking.
Not “which AI is right?” but “where do they disagree, and what does that tell me?”
If the fiscal analyst and the growth advocate both oppose it, that’s a strong signal. If the fiscal analyst opposes but the growth advocate supports, you’re looking at a real tradeoff and now you know exactly what the tradeoff is. If all three agree, maybe the decision is simpler than you thought.
The disagreement is the insight. Not the consensus.
This is obvious if you’ve ever worked in finance. When you read a Moody’s and an S&P report on the same credit, you don’t pick one. You read both, you notice where they diverge, and the divergence tells you where to focus your own analysis. The value isn’t in either opinion alone. It’s in the spread.
But we’ve collectively decided that the way to use AI is to ask one system for one answer and then either accept it or reject it. That’s like reading one analyst report and calling it research.
The fix isn’t better prompts. It’s more opinions.
I’ve been building what I call “crews”, sets of specialized agents that each analyze the same input from different angles. Not a committee that compromises. Not an ensemble that averages. Independent analysts who each render a verdict based on their own methodology and their own beliefs about what matters.
The output isn’t “here’s the answer.” It’s “here’s where they agree, here’s where they disagree, and here’s why.”
A few things I’ve learned building these.
The agents need to be opinionated, not balanced. If you tell three agents to “give a balanced analysis,” you get three versions of the same balanced take. Useless. Each agent needs a declared perspective, what it optimizes for, what it deprioritizes, what it refuses to weigh in on. The narrower the lens, the more useful the disagreement.
They should show their work. An opinion without reasoning is a guess. An opinion with reasoning is analysis. An opinion with reasoning and cited sources is something you can actually evaluate. I call these “receipts”, what evidence did this agent use to reach its conclusion? If the fiscal hawk says “oppose” but its receipts reference a budget memo from 2019, you can discount that. If the neighborhood voice says “strongly oppose” and cites 187 public comments with 73% opposition, that’s hard to argue with.
Disagreement should be measured, not just noticed. If you’re going to compare opinions, you need a way to quantify how much they diverge. Not as a judgment (”high divergence is bad”) but as a signal (”high divergence means this decision has real tension and your judgment matters more here”). The “widest gap”, the pair of analysts who disagree most, is where the real tradeoff lives. If the growth advocate says “support” and the neighborhood voice says “strongly oppose,” you now know exactly what you’re choosing between: economic development vs. resident quality of life.
My friend with the city council transcript doesn’t need a better AI. He needs a better pattern.
Paste the transcript. Get three opinions from three different lenses. See where they agree, that’s probably settled. See where they disagree, that’s where his judgment as a resident actually matters. Look at the receipts. Make his own call.
This works anywhere multiple opinions beat one balanced summary. Policy, credit, hiring, diligence.
One opinion feels like an answer. Multiple opinions feel like research.
And research is how grown-ups make decisions.
If you’re building agent systems, try this: before you optimize your single agent’s output, duplicate it three times and give each copy a different institutional belief. Run the same input through all three. The first time you see them disagree, you’ll never go back to asking one AI what it thinks.