The Slop Tax

The draft looks done.

That is the trap.

It has headings. It has bullets. It has that clean, harmless confidence that makes weak work feel civilized. You paste it into the doc, skim it once, and feel the little chemical reward of progress.

Then tomorrow comes. Someone asks what the claim means. A number has no source. The customer paragraph sounds like it was written by a committee that has never paid for anything. The code runs, but no one trusts the edge cases. The email is fluent enough to send and hollow enough to make the sender look careless.

The work was faster. The trust got slower.

The False Diagnosis Is Speed

The childish argument says AI is good because it makes things faster, or bad because it makes things worse. Both are too clean. The real problem is meaner and more useful: AI moves the cost from creation to judgment.

Creation used to be the bottleneck. You had to write the memo, build the first draft, outline the page, name the edge cases, make the ugly pass with your own hands. That friction was irritating, but it forced contact. You had to notice what you believed. You had to decide what mattered.

Now the first version arrives before your standards have woken up. It feels like leverage because the screen is full. But a full screen is not the same thing as a finished thought.

Harvard Business Review has a useful name for the office version of this: workslop, polished work that shifts cleanup onto someone else instead of reducing the burden of work. That phrase should make every ambitious builder a little nauseous. Not because AI is useless. Because plausible work is more dangerous than obviously bad work.

Obviously bad work gets rejected. Plausible work gets forwarded.

Then it breeds.

Plausible Is Expensive

A bad paragraph with dirt under its nails is honest. You can smell the weakness. You cut it, rewrite it, or admit the idea is not ready.

AI slop is worse because it wears a pressed shirt. It gives you the feeling of structure before the structure has earned trust. It gives you the rhythm of competence before the thinking has been paid for. It gives you the shape of a decision without the bruise of making one.

That is why smart people get caught. Not because they are lazy. Because the output is close enough to flatter them and vague enough to escape immediate punishment.

This is the same old founder disease in a newer suit. You already loved tools because tools let you feel like the person who moves. AI just made the costume more convincing. Instead of one blank page accusing you, you now have ten decent pages asking for review.

Congratulations. You automated the easy part and multiplied the part you were avoiding.

The Tax Shows Up Later

The slop tax rarely arrives at the moment of generation. That is why it survives. At the moment of generation, everything feels clean. The tool answered. The file exists. The page has body. The agent made progress while you made coffee.

The bill arrives downstream.

It arrives when a customer reads the page and cannot tell who it is for. It arrives when a teammate has to reverse-engineer the logic inside a shiny proposal. It arrives when a buyer senses that the case study was assembled from fog. It arrives when a developer spends an hour finding the one generated line that should never have been trusted.

In a randomized study of experienced open-source developers, METR found that developers using early-2025 AI tools took longer on the studied tasks, even though the developers expected the tools to help before the measured result contradicted the feeling. Do not turn that into a religion either. The point is not that AI always slows people down. The point is that feeling faster is not proof of leverage.

The body knows the difference before the spreadsheet does. Real leverage leaves the system lighter. Slop leaves the system full of small doubts no one wants to name.

Is this true? Who checked it? Does this claim belong to us? Why does the offer sound like everyone else's offer? What did the agent assume? What did the prompt fail to say? Which part would embarrass us if the buyer were more informed than we hoped?

Those questions are not friction. They are the tax office.

The Manual Pass Is The Moat

The amateur response is to ban the tool or worship it. Both are childish. The operator response is to separate generation from authority.

AI can draft. It can compare. It can find angles. It can make the ugly first version less precious. It can help a solo builder touch more surfaces without hiring a small department. Fine. Use it.

But never confuse output with permission.

Nielsen Norman Group's 2026 guidance on AI-assisted survey writing makes the same point in a narrow, practical way: AI can produce polished survey drafts quickly, but human expertise is still needed to catch subtle flaws that damage the data before the research pretends to be useful. That is the principle. The more polished the draft, the more explicit the review must become.

A manual pass is not nostalgia. It is where taste, context, liability, source discipline, buyer empathy, and strategic intent re-enter the room.

This is where you earn the right to automate. You read the draft against the wound. You cut the clever sentence that does not help the buyer. You trace the claim back to a source. You ask whether the example could be misunderstood. You test whether the thing sounds like a person with a point of view or a machine trying to avoid blame.

The manual pass is not the slow part. It is the part that makes speed worth having.

Build The Judgment Loop

You do not need a better prompt first. You need a judgment loop.

Before the tool touches the work, name the job. Not the task. The job. Is this supposed to persuade a skeptical buyer, reduce internal risk, explain a hard tradeoff, find edge cases, compress research, or produce options for a human choice? If you cannot name the job, the tool will produce activity dressed as help.

Then name the standard. What would make the output good enough to carry your name? Sources linked. Buyer language intact. Claims testable. No generic abstractions. No fake certainty. No borrowed voice. No paragraph that sounds like it could appear on a thousand other sites by lunch.

Then name the review. Who checks truth? Who checks taste? Who checks risk? Who checks whether the output made the next action clearer or just made the folder heavier?

Microsoft's 2025 Work Trend Index calls the emerging role an agent boss, with people directing AI systems rather than merely using assistants inside more agent-shaped workflows. Cute phrase. Dangerous job. A bad boss creates motion and calls it management. A good boss sets standards before the work begins and refuses to accept output that merely looks busy.

That is the shift. You are not trying to produce more. You are trying to make more work worthy of trust.

The Slop Tax Audit

Pick one workflow you have handed to AI. Writing. Research. code review, outbound, support replies, documentation, product specs. One is enough.

Now ask the rude questions. What did the tool make faster? What did it make less visible? What now requires checking that did not require checking before? Which mistakes look polished enough to survive a skim? Which outputs would damage trust if they reached a buyer unchanged?

Then price the tax. Count the review time. Count the rework. Count the awkward clarifying messages. Count the buyer confusion. Count the little loss of confidence every time someone opens an AI-assisted draft and silently wonders what else is wrong.

If the tool still leaves the system lighter after that, keep it. Tighten the prompt. Save the checklist. Build the template. Let the machine carry what it can carry.

If the tool only moved the work into inspection, stop calling it leverage. Call it what it is: a beautiful invoice with no line item.

The Cleaner Machine

The draft can still look done. That part does not change.

What changes is your relationship to the feeling. You stop treating a full page as relief. You stop letting fluency impersonate quality. You stop confusing the absence of a blank screen with the presence of a real decision.

The work gets stranger after that. Quieter. More exact. Less impressed with itself. A machine can give you ten versions, but it cannot decide which one carries your promise without making the buyer do cleanup in their head.

That is still your job.

Use the tool. Tax the output. Keep the judgment.