Criterion Model Dimension

Hidden criterion is the main AI failure because a complete-looking answer often conceals the standard it optimized for, and users rarely notice until the output fails in the world.

Ask for 'the best page' and the system will still need a standard. If you did not name one, it will supply or infer one. That is not a bug. It is the structure of the task.

The failure happens when the user mistakes that hidden standard for neutrality.

Why hidden criterion is so common

Most users think they specified the task when they only specified the topic. Topic is not criterion. 'Write about X' is not the same as 'write about X under this standard, for this end, under these constraints.'

The model then optimizes toward proxies like readability, generic consensus, institutional convention, or persuasive completeness.

Why it matters more than many factual errors

A wrong date can be fixed. A hidden criterion changes the entire shape of the answer. It determines what gets emphasized, what gets omitted, and what looks like success.

That is why a page can be factually decent and still structurally wrong.

The correction

State the standard before asking for final output. Then ask the model to tell you what standard it appears to have optimized for anyway.

Go deeper inside Modern Discernment

Frequently asked questions