Ask any major AI chatbot for a random number between 1 and 10 and you will almost certainly get 7. Researchers studying this pattern describe it as a kind of statistical gravity, a pull toward the most-expected answer baked into the training data. The model is not choosing 7 because 7 is correct. It is choosing 7 because the humans who generated its training data chose 7, over and over, until the entire probability space tilted that direction. The model is doing exactly what it was trained to do. It is just not doing what you actually asked.
That gap between "what the system optimizes for" and "what you actually need" is not an AI problem. We have watched it play out in product strategy, in agency vendor selection, in hiring, in tech-stack decisions, across hundreds of engagements over twenty-five years. The question is always the same: is this the right answer, or is it just the answer that everyone else is already landing on?
Here is where it gets interesting. The EV market just handed us a live case study in what happens when a single dominant consensus fractures. May 2026 data shows pure battery electrics up 15% year over year while plug-in hybrids dropped 15%, a simultaneous surge and collapse inside what most people still treat as one category. The market did not drift. It bifurcated sharply, and anyone who built their product, their installer network, or their service model around "EVs are EVs" just got caught on the wrong side of a clean line. The consensus answer, "electrification is the trend," was true and useless at the same time. The directional bet that mattered was inside the consensus, not on it.
This is the trap most founders who come to us are already inside. They read the same newsletters, follow the same operators on social media, and sit in rooms with people who all converged on the same stack, the same pricing model, the same customer acquisition playbook. It works until it stops working. When it stops working, everyone is surprised at the same time because they were all holding the same position.
The deeper issue is about what it actually takes to build something generative rather than something derivative. Scientists recently built the first synthetic cell capable of growing and dividing, which sounds like a biology headline until you understand what made it hard. The challenge was not assembling the parts. The parts were available. The challenge was building something that could sustain its own forward motion, that could replicate without being handed a template every generation. Consensus-optimized systems can assemble. They cannot replicate independently. They need constant prompting, constant input, constant context from outside themselves. Sound like any team you have managed?
The pattern we keep seeing is this: founders hit a wall around $1.5M to $2M in revenue and call us because growth has plateaued. When we dig in, we almost always find that the product, the ops, and the tooling have all converged on the same "safe" center. The CRM is whatever everyone else is using. The analytics setup tracks whatever the default dashboard shows. The content strategy is whatever the top SEO articles from three years ago said to do. Nobody made a bad decision. They made a sequence of reasonable decisions that all pointed the same direction and ended up at 7.
Breaking out of that is less about finding a clever new tool and more about rebuilding a culture of genuine divergence inside the team. Isamu Noguchi spent years designing playgrounds that were never built because city planners could not reconcile art and infrastructure, two categories they had decided did not belong together. His unrealized designs were not failures. They were proof that the system's consensus about what a playground should be was a constraint, not a fact. The founders we work with who actually break through their plateau are the ones who can do what Noguchi was doing: hold a genuine question open long enough to see where the consensus is arbitrary.
That is a discipline, not a talent. It requires building actual practices around dissent. Scheduled red-team sessions before major product decisions. A standing policy of asking "who is NOT buying this and why" before any feature ships. Explicit separation of "what do our best customers want" from "what does the market at large want," because those two answers diverge constantly and the best opportunity usually lives in that gap. The artist Guimi You does something analogous in paint: her work sits deliberately in the space between clear memory and atmospheric uncertainty, refusing the consensus resolution of either sharp representation or pure abstraction. That refusal is the work. The tension is not a problem to solve; it is the thing itself.
For a founder, that tension usually lives between what is already working and what the next version of the business needs to be. The instinct is to resolve it fast, pick a lane, stop feeling ambiguous. We have watched that instinct kill good businesses more reliably than any market downturn. The ones that scale well are the ones that learned to stay in the question a little longer than is comfortable, to resist the pull toward the obvious answer, to notice when the whole room has agreed too quickly.
Here is the practical version. When your team reaches fast consensus on a product direction, on a vendor, on a hiring profile, that is not a green light. That is a signal to apply pressure. Not because the consensus is wrong, it might be right. But because consensus-optimized decisions are, by definition, already priced in. Your competitor made the same call last quarter. The upside is gone before you ship.
The businesses we have seen survive the jump from scrappy to scalable are the ones that treat "everyone agrees" as a question, not an answer. The AI will give you 7. The market will tell you hybrids are EVs. The city planner will tell you playgrounds are not art. The only durable edge is the team that keeps asking what happens when you actually mean random.