The wrong people are driving your AI decisions
Why confidence keeps beating experience in the room
There’s a meeting happening somewhere right now about AI adoption. Someone on one side of the table says coding agents are transforming how their team ships. Someone on the other side says they’re generating garbage that senior engineers spend all day cleaning up. Both are speaking with conviction. Neither is obviously wrong.
And somewhere in the middle, a staff engineer who has actually used these tools for six months in production says something careful and qualified. The room moves on before she finishes.
This is not a coincidence.
The center is not neutral. It’s informed.
Armin Ronacher wrote something recently worth sitting with: the people who occupy the middle ground in any technology debate tend to look like adopters, not because they’ve made up their minds, but because forming an informed opinion requires actual contact with the thing. If you want to criticize something well, you first have to get close enough to dislike it for the right reasons.
This creates a structural asymmetry. One side has paid the cost of direct experience. The other hasn’t, or not to the same degree. From the outside, the cautiously experienced and the uncritically enthusiastic can look identical. They both “use AI.” But those are not the same position, and treating them as equivalent is where a lot of adoption decisions go wrong.
What Ronacher is describing is an epistemic gap, a difference in the quality and texture of knowledge between people who appear, on the surface, to be in the same camp.
Why the most calibrated voices get compressed
In any evaluative social context, and technology adoption meetings are deeply evaluative, people gravitate toward contributions that feel safe to make. Confident, simple positions travel easily. Nuanced, qualified positions require more cognitive work from the audience and more social exposure from the speaker.
This is the same mechanism at work in bike-shedding. Groups don’t avoid complexity because they’re lazy. They avoid it because complexity is cognitively expensive and socially risky when the room is watching. The person who says “AI agents are transformative” and the person who says “AI agents are garbage” are both making contributions the room can process quickly. The person who says “it depends on how you structure the handoff between human judgment and machine output, and here’s what I’ve learned about where that boundary should sit” is making a contribution that requires everyone to slow down.
Rooms don’t slow down voluntarily. So that voice gets compressed into whichever camp it sounds most like, and the nuance disappears.
The overconfidence problem runs in both directions
There’s a second mechanism compounding this.
Dunning and Kruger’s core finding is familiar: people with limited experience in a domain tend to overestimate their competence, because the skills required to produce good work are the same skills required to recognize good work. If you haven’t used AI tools seriously, you also lack the direct experience to accurately assess what you’re missing, or what you’re not missing.
The less-discussed mirror finding matters just as much. Highly experienced people tend to underestimate their competence, because the more you know, the more visible the landscape of your ignorance becomes. The engineer who has spent six months learning exactly where coding agents fail will speak with more qualification and more hedging than someone who has used them for two weeks and found them impressive.
I’ve written before about how confidence is not signal, that in any room where competence isn’t externally measured, the least qualified person will often be the most confident, and the most qualified will often be the most restrained. This isn’t a failure of human cognition. It’s a predictable output of how metacognitive sophistication works. The more accurately you understand a problem, the more clearly you can see what you don’t yet know.
Put Ronacher’s asymmetry together with that finding and you get a specific prediction: in AI adoption discussions, the people with the most useful signal will tend to present with the least apparent certainty, and will therefore receive the least weight in decisions that rely on confident advocacy to move forward.
What this actually costs
The practical cost is significant and largely invisible.
When organizations make AI adoption decisions by aggregating the loudest voices rather than the most informed ones, they tend to oscillate between two failure modes. Premature adoption: deploying tools into workflows that weren’t designed for them, producing the outcome we’re now seeing in many engineering organizations, more code output, lower signal quality, exhausted humans reviewing machine-generated work without the structural support to do that job well. Blanket rejection: dismissing tools that would genuinely improve how people work, based on criticisms that don’t survive contact with actual use.
Both failure modes share a cause. The decision was made without adequately weighting the people who’d paid the cost of direct experience.
The organizations getting AI adoption right aren’t the ones with the most enthusiastic leadership or the most vocal skeptics. They’re the ones treating it as a design problem: which tasks belong to machines, which belong to humans, and how do you structure the handoff so that humans operate where they’re actually strongest. That kind of design thinking requires input from people who’ve worked close enough to the tools to have calibrated views. It can’t be done from a position of pure enthusiasm or pure refusal.
The processing time gap
There’s a third dynamic worth naming. The people initiating decisions about new technology have usually had much more time to process the implications than the people receiving those decisions.
I’ve written about this pattern before, that change has a burn rate too. A VP who announces an AI-first engineering strategy has spent months thinking about the strategic logic. The team receiving the announcement has had fifteen minutes. That gap in processing time creates a predictable misread: the leader has already grieved the old way of working and gotten excited about the new one. The engineers who’ve actually used the tools are still working through what they’ve learned, what’s promising, what’s broken, and what would need to change for any of it to work at scale.
The most useful input, the calibrated, experienced, nuanced kind, often arrives too late to influence the decision. Or it arrives in a form that gets compressed into a camp before it can be properly heard.
What to actually do about it
The fix isn’t to distrust enthusiasm or distrust skepticism. Both contain real signal. The fix is to change how you weight input in technology decisions.
Seek out the people who’ve used it seriously enough to have a qualified view, not just a positive or negative one. The engineer who can tell you “it works well for X but breaks down at Y, and here’s the specific condition where the failure happens” is giving you more actionable information than either the evangelist or the refuser. Weight that input accordingly.
Create structural space for nuanced contributions. If your meeting format rewards fast, confident takes, you will get fast, confident takes. If you want calibrated judgment, you have to design for it: written pre-reads, structured questions, explicit invitations to surface the conditions under which something works versus where it fails.
Treat “I don’t know yet” and “it depends” as signs of expertise rather than as problems to be resolved. A useful data point from hiring research (Schmidt & Hunter, 1998): structured interviews predict job performance at nearly twice the validity of unstructured ones, precisely because they force the evaluator to slow down and apply consistent criteria rather than rewarding whoever presents most confidently. The same principle applies here.
Close the processing time gap. When making technology adoption decisions, acknowledge that the people being asked to implement the change have had far less time to develop a view than the people making the decision. Build in time for that processing to happen before commitments are made. The cost of that delay is almost always lower than the cost of implementing a decision that the most informed people in the room hadn’t finished thinking through yet.
The bottom line
The AI debate feels polarized because the actual center is hard to see. It doesn’t look centered. It looks like qualified adoption, which from a distance is indistinguishable from enthusiasm.
The people with the most useful signal are the ones who’ve gotten close enough to the technology to dislike it for the right reasons and value it for the right reasons, often simultaneously. They exist in your organization. They’re probably not the ones talking the most in your adoption meetings.
The question is whether your decision-making process is structured to hear them before it commits to a direction, or whether it will compress their nuance into whichever camp they sound most like and move on without them.
Getting that right isn’t a technology problem. It’s a meeting design problem, and underneath that, a question of whether the people with the most to offer feel safe enough to say the qualified, uncomfortable, not-quite-finished thing out loud.