Judgment in the Loop

The cost of producing competent work has collapsed. The cost of deciding whether it is the right work has not.

A first draft, a working prototype, a plausible analysis: each of these can now be generated in the time it used to take to schedule a meeting about them.

What has not changed is the moment where someone must look at the output and say this is right, or this is wrong, or this is fine but it answers the wrong question.

That moment is now the bottleneck in most knowledge work. It is also the part almost nobody has been trained to do deliberately.

When production was expensive, judgment was distributed across the process. You made decisions slowly because each step cost something to take.

Cheap generation removes that friction, and with it the pauses that used to force reflection. The work arrives faster than the thinking that should precede it.

So judgment gets compressed into a single checkpoint at the end, where a person reviews output they did not construct and often cannot fully trace.

The failure mode is not that machines produce bad work. It is that fluent, well-formatted, confident output is much harder to reject than a rough draft. Polish reads as correctness.

The valuable skill shifts from producing to discriminating. From can you make this to should this exist in this form.

This rewards people who hold a clear standard. If you know what good looks like in your domain, abundant generation makes you faster. If you do not, it makes you prolific and directionless.

It also changes what expertise means. Expertise used to be knowing how. Increasingly it is knowing when something is subtly wrong, which is a slower and less transferable thing to build.

Organizations feel this as a strange new problem: more output, less certainty about whether any of it was the right move.

Separate the generating step from the deciding step, deliberately, in your own work.

Before you ask a tool to produce something, write down what a good answer would contain and what would make you reject it. Two sentences is enough. Do it first, so the output cannot define the standard retroactively.

Then review against your written criteria rather than against your impression of the output. Fluency is not evidence.

Keep a short record of the times you accepted something and were wrong. That record is how calibration gets built, and it is the one input no tool can supply for you.