We’ve all got that one friend. You know the one — charismatic, well-read, and devastatingly confident. He can explain the intricacies of the Roman Empire or the mechanics of a jet engine with such authority that you find yourself nodding along.
It’s only later, when you’re driving home, that you think: wait, did he say the Vikings invented the color orange after finding pumpkins in the fjords? That… actually sounded plausible at the time.
This is the core experience of working with a large language model. It is the plausible moron: fluent, assured, and untethered from any internal sense of whether it is right.
Fluency is not knowledge
An LLM is optimized to produce text that looks like a correct answer. It is not optimized to know one. Those two objectives overlap often enough to be useful and diverge often enough to be dangerous — and the model gives you no reliable signal about which mode it is in. The tone of a confident truth and a confident fabrication are, to the machine, identical.
The danger isn’t that the model is wrong. It’s that it is wrong with the exact same confidence it is right.
Your charismatic friend has the same bug. The fix, in both cases, is not to stop listening — it is to stop outsourcing your judgment. You keep the fluency and you supply the verification.
Working with it anyway
Treat the model as a brilliant, tireless intern with no shame: fast, broad, and in constant need of a check. Give it the boring, bounded, verifiable work. Reserve the judgment — the weighing, the pivoting, the knowing-when-something-is-off — for the human in the loop.
That judgment is exactly the layer Dominion is built to capture and preserve. Because the long-term answer to the plausible moron is not a bigger model. It is a model that has finally been shown how humans actually decide.