The Job Became Reviewing, Not Writing

Every argument about AI slop starts from the wrong accusation. People act like getting caught means someone learns you used a model to write the first draft. That is not what actually triggers the reaction. Nobody has ever read a clean, accurate, well built piece of writing and felt betrayed to later learn a machine helped start it.

What triggers the reaction is the opposite. It is reading something wrong, generic, or slightly broken, and realizing that the person who sent it never looked at it again after the machine finished. The tool is not the offense. The absence is the offense.

I use AI constantly, for drafts, for structure, for a first pass at an automation script or a rough cut description. Nobody has ever pushed back on that when the result was solid. The complaints only show up when the result was not solid and it went out anyway.

So the starting question is wrong. It should never be did you use AI. It should be did you check what it gave you before you put your name on it.

People point at certain patterns and call them AI tells. A particular rhythm, an overused structure, a stray character that shows up more in generated text than human text. But none of these are actually evidence of a machine. They are evidence of nobody proofreading.

A tired human writes generic sentences too. A rushed human repeats a phrase without noticing. The difference is that a careful human, whether or not they started with a model, ends up removing the tell before anyone else sees it. The tell survives only when the review step never happened.

This means the visible signal everyone calls slop is not really about generation at all. It is a downstream symptom of skipped attention. Blaming the tool for the tell is like blaming a camera for a blurry photo nobody checked before printing it.

Once I saw it this way, my own process changed. I stopped worrying about hiding that a draft started with AI, and started worrying only about whether the final version was actually good. The origin was never the risk. The unedited output was.

You can hand a model the job of producing words. You cannot hand it the job of knowing whether those words are true, whether they fit the moment, or whether they sound like you. That part stays with the person whose name goes on the work, every time, no exceptions.

This is the same principle that runs through everything I build. A prompt can draft a page structure. It cannot decide if the structure actually serves the client. An automation can move data between systems. It cannot decide whether the data was ever correct in the first place. Generation is cheap. Judgment is not, and judgment cannot be generated on your behalf.

The mistake people make is treating a good output as proof that understanding happened. It is not proof. It is a draft that got lucky, or a draft that got caught. Either way, the understanding still has to come from a person who reads it and says yes, this is right, before it goes anywhere.

I do not think this is a limitation of current models that will disappear later. I think it is a permanent line. Thinking can be assisted. It cannot be delegated away entirely, because the responsibility for being wrong never transfers to the tool.

For a long time, writing meant producing sentences. Now, for a lot of work, producing sentences is the cheap part and reviewing them is the actual job. The center of gravity moved, and a lot of people have not noticed the shift yet.

Reviewing well is a real skill, not a lesser one. It means reading slowly enough to catch a fact that is wrong even though it sounds confident. It means hearing when a sentence is technically correct but does not sound like the person sending it. It means being willing to throw out a paragraph that is fine, because fine is not the bar.

In my own work this shows up constantly. A generated caption might be grammatically perfect and still wrong for the brand voice. A drafted automation step might run without errors and still make a decision I would never actually make. Passing does not mean finished. My review is what finishes it.

Treating review as a lesser task, something you rush through because the hard part is already done, is exactly how slop gets published. Review is not the leftover step after the real work. In this era, review is closer to being the real work itself.

Strip away the AI angle and slop is not a new phenomenon. It is the same old failure it always was, someone sending out work they did not actually stand behind. The technology just made it faster and more common, because generation got easier while review stayed exactly as effortful as before.

That gap, easy generation and unchanged review cost, is the entire crisis. It is tempting to skip the expensive step now that the cheap step got so fast. Every piece of slop anyone has ever complained about is a version of someone giving in to that temptation.

The fix was never going to be avoiding AI. The fix is refusing to let the speed of generation talk you out of the discipline of review. That discipline was always the actual craft, long before any of these tools existed.

When my name is on something, film, a page, an automation, a caption, the question I ask has not changed even though the tools have. Do I actually stand behind this. If the answer is yes because I checked, it goes out. If the answer is yes because it looked fine at a glance, that is exactly how slop happens, and I know better now.