A few weeks ago I wrote that slop was never an AI problem, and I promised a follow-up on the thing arriving to catch it. Here it is, because the catching has arrived faster than I expected.
In the space of a month: Anthropic started watermarking the text its Claude models generate, invisibly, in a way that survives a copy and paste. The EU’s transparency rules now want AI-generated content to be machine-readable as such. And LinkedIn shipped a button that lets anyone flag a post as “seems like AI slop.” Detection, from three directions at once. The instinct is relief. Finally, we can tell.
Tell what, though. That is the question nobody is asking, and it is the only one that matters.
Because every one of those tools answers exactly the same question: was a machine involved. Not whether the thing is any good. Not whether it is true. Not whether a person had a point and used a tool to make it faster. Just: machine, yes or no.
The mark proves processing, not worth
A watermarked article can be the best thing you read this week. An unwatermarked one can be worthless drivel typed by hand at midnight. The presence or absence of the mark tells you nothing about which is which. We are building, at speed and with the force of law behind it, an elaborate apparatus for detecting the one thing that was never the point.
Slop was always defined by what it does, positioning nothing and moving no one, not by who or what made it. Detection cannot see that. It can see the tool. It is colour-blind to quality.
The detection reflex, and where it goes wrong
Here is the danger, and it is already starting. Once a mark exists, people use it as a verdict. “AI-involved” quietly becomes “bin it.” And the moment that happens, you throw out excellent work that used a tool, and you wave through mediocre work that did not. You optimise your whole content operation around being undetectable rather than being good, which is a spectacular waste of effort, because undetectable was never the assignment.
It is a bit like marking down an essay for using a spell-checker. Or refusing a builder’s quote because they worked it out on a calculator. The tool touched it. So what. The question is whether the wall stays up.
What your audience actually judges
Your buyer does not care whether AI touched your content. I promise you they do not. They care whether it is useful, whether it is true, and whether it was worth the ninety seconds they gave it. That is the only detector that has ever been running, and it runs in their head, not in a plugin.
So optimise for that judge. Not for the watermark, not for the flag, not for Brussels’ preferred file format. Say something that survives the only test that counts, the one no detector can run: does this do a job for the person reading it.
The verdict
Detection is here, and on balance it is probably a good thing. Provenance and transparency are worth having, and I would rather live in a world with them than without. Just do not mistake the mark for a verdict. It tells you a machine was in the room. It cannot tell you whether anyone in the room had a point.
Detectable isn’t disposable. It never was.