People talk about “AI slop” as if it’s some new disease eating the internet alive. Mass-produced, generic content with that unmistakable stench of ChatGPT-or-Claude. Fine. Hate it. Be angry. But let’s not pretend human slop wasn’t everywhere long before LLMs showed up.
You know what I’m talking about. The dead-eyed blog post written to satisfy a keyword quota. The filler article published because someone needed “today’s content.” The LinkedIn post that’s 80 percent clichés and 20 percent self-congratulation. The expert piece that’s basically lifted from Wikipedia, then sprinkled with extra words.
We used to call it bad writing. Now, apparently, it’s “authentic.”
Put someone with poor judgment behind the keyboard and you’ll get poor results. Type “write me an article about whatever,” add no thought, no direction, and no standards, and the model will hand back exactly what you asked for: generic, predictable, lifeless copy.
Someone with taste and judgment doesn’t treat the first output as finished work. They treat it as raw material. They cut it apart, rewrite it, add their own observations, use it to chase ideas, check the facts, and put uncertainty back where it belongs.
Maybe a few AI fingerprints survive. The rhythm is a little too smooth. There are too many tidy transitions. Too many balanced paragraphs. Too many conclusions that politely restate the introduction. The paragraphs sit there in suspiciously perfect symmetry. So what? Those things can be fixed. They aren’t a death sentence.
The smart person keeps working. The idiot hits copy, paste, publish.
Plenty of authors are already doing this. Many of them are probably keeping quiet about it, because there’s a loud crowd of angry people waiting to turn every “discovery” into a moral crusade. Meanwhile, the authors using AI well are breaking through blocks, trying different plot turns, testing dialogue, and cleaning up chaotic drafts. They’re using it as an editor.
That’s the real dividing line. Can the author still control the voice, rhythm, and intention? When AI serves the author, the authorship remains. When the author outsources the judgment as well as the writing, you get slop.
So yes, the internet is filling up with AI slop. But that’s only the newest form of a much older problem: too many people with nothing to say, combined with standards low enough to let them publish it anyway.
Quality concerns the artifact. Authorship concerns agency. Disclosure concerns trust.
A piece of writing can be great and yet it can be almost completely machine generated. That doesn’t make it bad. Misrepresenting how it was made is a different question entirely.
So what is it that people are afraid of? People use the term “AI slop” even when some of the writing they’re describing is more competent than plenty of human-written content that came before it. The problem is a lack of specificity. “AI slop” is sometimes used to mean bad AI-generated writing, and sometimes simply writing made with AI. It comes off as a bad faith argument. That ambiguity makes the term remarkably convenient: it can condemn bad writing and AI involvement at the same time, without having to prove that the two are the same thing.
There’s a legitimate discussion about copyright, loss of trust, homogenization and false expertise to be had. But another part is much older: suspicion toward any technology that lowers the cost of acquiring a skill advantage.
We often confuse the cost of learning something with the value of what’s being produced. Photography once required chemical knowledge. Typography once required craftwork. Research once required libraries. Editing once required scissors. AI is unsettling because it reduces cognitive friction—the kind of friction we’ve traditionally treated as evidence of intelligence, expertise and authorship.
There is also something more uncomfortable going on. If you spent twenty years acquiring a skill, a tool that lets someone close part of that gap in twenty minutes can feel fundamentally unfair. Not because the resulting work is necessarily worse, but because the suffering that once functioned as proof of competence suddenly matters less.
And that is why I think Pangram may turn out to be one of Substack’s worst product decisions. It could do more damage to Substack’s trust model than the AI slop it is supposed to expose.
Pangram attempts to distinguish human-written from AI-generated text by identifying statistical patterns. Yet human authors are already reporting false positives. And once a detector’s verdict is shown to readers, “misclassified” can quickly become “accused.” A false positive doesn’t merely misclassify a text. It plants suspicion between the author and the reader, with Substack itself supplying the accusation.
Substack says it wants authenticity, but has outsourced the adjudication of authenticity to another AI system. The irony is exquisite: in an attempt to restore trust in human authorship, the platform asks a machine to decide who counts as human.