The Death of Slop: Why Detection Loses and Provenance Might Not

Slop Is An Economics Problem Wearing A Technology Costume

The volume of low-quality generated content on the open web is usually discussed as a failure of AI. It is not. It is the predictable result of a cost curve moving while an incentive stayed still.

For two decades, platforms paid for attention. Producing something worth attention cost time, skill, and money, which acted as a natural throttle on volume. Generative models removed the cost and left the payment in place. Anyone surprised by the flood was not watching the equation.

This matters because it tells you which fixes cannot work.

---

Why Detection Loses

There is a simple test for whether an open standard is actually open: can the person who wrote it take it back?

The instinctive response is to detect generated content and demote it. This fails for three structural reasons, none of which is going to be engineered away.

The classifier is always behind. Detection trains on the output of yesterday's models. Generation improves continuously. The gap is not closing. It is a permanent property of the arrangement.

The category is wrong. Machine-generated is not the same as worthless, and human-written is not the same as worth reading. Plenty of generated text is accurate and useful. Plenty of human writing is padding. A filter that sorts on provenance rather than value will misclassify in both directions, and the errors that hurt most are the useful things it silently removes.

The incentive to evade is unbounded. Whoever profits from the flood has an arbitrarily large budget for defeating the filter, and only needs to win once per item.

Detection is a defensible tactic. It is not a solution, and a platform that stakes its credibility on it is making a promise it cannot keep.

---

Most people who say "for humanity" build a kill-switch. I am removing mine.

The Two Fixes That Address The Cause

Move the burden from detection to attestation

Instead of asking whether something was generated, ask who is willing to put their name to it.

That is a different question with a different failure mode. Provenance is checkable. Quality is not. An attestation says a named party asserts specific things about a piece of content — its origin, what was reviewed, what was verified — and signs that assertion. It can be wrong, but it cannot be anonymous, and it cannot be quietly revised after the fact.

This does not make bad content disappear. It makes bad content attributable, which is the precondition for reputation to do any work at all. Anonymous slop is free. Attributed slop costs the attributor something.

Move curation to the reader's side

The deeper problem is that the filter sits with the party that profits from volume. A platform paid by engagement is structurally the wrong entity to decide what you should not see, and no amount of policy fixes a conflict of interest that lives in the revenue model.

It rewards activity, not squatting. It is a head-start with a sunset and a clock.

An agent that curates on the reader's behalf inverts that. It is accountable to one person, and its instructions are that person's instructions. It can be told to reject entire categories, to demand provenance before surfacing a claim, or to prefer primary sources over commentary about them — and it applies that consistently, which no human reader manages.

The important consequence is economic. Slop works because distribution is free and indiscriminate. Reader-side filtering makes distribution conditional. A flood that reaches nobody has no return, and the cost curve that produced it stops paying.

---

What This Does Not Fix

Three things, stated plainly, because the optimistic version of this argument is wrong.

Personal curation can build a very comfortable cage. A filter that only shows you what you already accept is not an improvement on a feed optimised for outrage. It is the same failure with better manners. Any honest curation agent has to be instructable to include material its owner will dislike, and most people will not instruct it that way.

Give away the seed. Get rich on the orchard. And make sure no one can burn the orchard down — including you.

Attestation moves trust, it does not create it. A signature proves who made a claim. It says nothing about whether the claim is true. The registry of attestors becomes the thing that matters, and registries can be captured, gamed, or turn into gatekeepers — including by whoever operates them.

Adoption is the whole game, and it is not won. Reader-side curation only changes the economics at scale. Below scale it is a useful tool for a small number of careful people while the flood continues everywhere else. Nothing in the argument above guarantees it gets there.

---

The Position Worth Holding

Slop will not be filtered out of existence, and anyone selling that outcome is repeating the promise spam filters made about email — a permanent, partially effective arms race described as a victory.

What can change is where the filter lives and who it answers to. Move provenance to something checkable, move curation to the reader, and the flood stops being profitable rather than stops being possible.

That is a slower and less satisfying answer. It is also the only one with a mechanism behind it.