The Model Doesn't Have Values. You Do.

Seven years ago, I tweeted: algorithms are opinions, not truth machines, and demand the application of ethics. It didn't land. (I'd tell you to find it on Twitter, but I'd also tell you to leave Twitter, so let's call that a wash.)

Now a model's output sits in your hiring pipeline, credit assessment, content feed, and performance review. The phrase "the AI recommended" has become the new "the data shows", a way of presenting a conclusion that sounds too technical to argue with. It's a transfer of accountability dressed as efficiency.

Accuracy Is the Wrong Test

Every algorithm encodes a definition of success. Whoever built it decided what good looks like, implicitly, if not explicitly. Engage more. Convert faster. Score higher. These aren't neutral goals. They're value judgements, compressed into a function, delivered to you as output.

There's a well-documented pattern called algorithm aversion: people are less tolerant of an algorithm's mistakes than their own, even when their own mistakes are larger. The bias cuts both ways. You distrust the model when it errs, and defer to it when it's confident. Neither response involves asking the right question.

You're Asking the Wrong Question

The most asked question is whether the recommendation is accurate. The useful question is: whose values are encoded in this system, and do I agree with them?

That question can't be outsourced. It requires a person with a position, a stake, and real accountability to look at what the model is optimising for and decide whether that's the direction they'd have chosen.

AI can help you move faster. It cannot carry the ethical weight of the decisions you're accelerating toward.

Before you act on the next recommendation, ask what it was built to optimise. Then decide whether that's a value you'd put your name on.