Essay · The values layer

Every model is an opinion

We talk about models as if they report the world. They argue about it. Every dataset is a selection, every objective a preference, every guardrail a value. Neutral is the one thing a model cannot be.

A model is built from choices: what data to include and leave out, what to optimize for, what to forbid. Each of those choices is a small opinion, and the finished system is their sum. This is not a flaw to be scrubbed out. It is the nature of the thing. The danger is not that models have opinions. It is that we keep calling them neutral, which leaves the opinions in place and removes the label that would let anyone question them.

Why "neutral" is the dangerous word

Calling a model neutral does not remove its point of view, it hides it, and a hidden opinion is one nobody can contest. The most consequential decisions get made twice, once in the choices that shaped the model and once when it runs, and the first time is invisible if we pretend it did not happen. Honesty starts by admitting the opinion is there.

The honest question

If every model is an opinion, the useful question is not whether but whose. Whose choices, whose values, whose interests are encoded, and do the people affected by the model's decisions have any claim on those choices. That question has an answer for every system in production right now. We just have not been in the habit of asking it out loud.