The backlash against AI-generated marketing is now a named trend, and it is being widely misdiagnosed.
The common reading is that audiences dislike machine-written text and the response is either to hide the tools or to avoid them. Both responses misunderstand the complaint. Very little of what people call slop is identifiable as machine-written; what makes it slop is that it is obviously unconsidered. Nobody decided it was worth anyone's time. It exists because producing it was cheap.
That is a process failure, and it has a process fix.
The defect is the missing decision
Take one model and two workflows. In the first, it drafts, a person reads it, most drafts get killed, and the survivors get published because someone thought they were worth reading. In the second, it drafts and everything publishes on a schedule.
Same model. Same prompts. One produces a useful blog and the other produces the thing everyone is complaining about.
The variable is not authorship. It is whether anything stood between the output and the audience. That is why sourcing better models has not solved this and will not: the missing component is a judgement, and a judgement is a thing a person makes.
Why this got bad quickly
Production capacity rose by an enormous factor. Review capacity did not move at all, because it is bounded by human attention.
Any system where one input scales and the other does not will converge on skipping the one that does not. If you can generate forty posts a day and read three, either you publish three or you stop reading. Plenty of operations stopped reading, and the result is what the trend is named after.
The insidious part is that it looks like progress the whole way down. Output is easy to measure, judgement is not, so a dashboard reports a team getting more productive right up until the audience leaves.
Small teams have the advantage here
This reads as bad news for anyone marketing with AI. It is mostly the opposite, if you are small.
The operations producing slop are structurally committed to volume. It is how they are staffed, priced and measured, and unwinding it means unwinding the business model. You have no such commitment. You can publish once a month and be fine.
That asymmetry is worth using rather than apologising for. Being able to not publish is a competitive position when everyone else has to.
Building the decision into the system
If you use an agent for this, the review step has to be structural rather than aspirational. Aspirational review is what everyone has in month one and nobody has in month six.
Structurally means there is no path to publishing that does not pass a human decision. Not a setting that defaults to on, not a bulk approve that nobody reads — a genuine gate with no code path around it. The reasoning for that design is here and it is the same argument arriving from a different direction: the approval boundary that exists to prevent an agent doing harm also happens to be the thing that prevents it producing slop.
Two details make it survive contact with a busy week.
The agent should propose less than it could produce. A system that generates everything possible and relies on a human to filter has handed its worst problem to the person with least time. Filtering upstream is cheaper than filtering downstream, and an agent that bins its own weak work and says so is doing the expensive part.
And approval should be explicit about what it covers. Approving a batch with a stated ceiling is fine and often correct; approving everything forever by clicking once is how the gate quietly stops existing.
What this means for what you publish
Fewer things, each of which someone decided was worth publishing.
In practice that means writing what you actually know. The material that survives this filter tends to be specific: something you measured, something you built and can describe, a decision you got wrong. It is also, conveniently, the material that answer engines have a reason to cite, because it does not exist anywhere else. The approach we use leans on exactly that overlap.
The uncomfortable implication is that this makes marketing harder, not easier. The bottleneck moves back to having something to say. That is the correct place for it to be, and the reason the backlash is ultimately good news for anyone with real material and a small enough operation to be selective.
If you want to see what an approval-gated version looks like, the first task is free, and how we price work that gets rejected explains why binning weak output has to cost the vendor rather than the user.
Last reviewed 18 September 2026.