We have spent the last few years obsessing over one question:
“Was this written by AI?”
I think we are asking the wrong question.
The more important question is:
“Is this worth reading?”
AI-generated content gets criticized for being repetitive, generic, inflated, and empty. We call it slop. But there is an uncomfortable problem:
Humans have been producing exactly the same kind of content for centuries.
The difference is that we never built a serious detection layer for it.
A human can write 2,000 words without introducing a single new idea.
A human can repeat the same argument five times using different vocabulary.
A human can write:
“In today's rapidly evolving world…”
and spend an entire paragraph saying almost nothing.
A human can make an extraordinary claim without evidence.
A human can turn one simple observation into ten paragraphs of intellectual decoration.
And because a human wrote it, we rarely call it slop.
That double standard needs to end.
Slop Is Not an AI Problem
I don't think slop should be defined by its author.
AI ≠ slop.
Human ≠ quality.
The correct distinction is:
Signal vs. Slop.
The origin of the text should be irrelevant.
What matters is whether the content creates value.
That leads to a much more interesting engineering problem:
Can we build a content audit system that evaluates the output rather than the author?
The Delete Test
One of the simplest mechanisms I have been experimenting with is what I call the Delete Test.
Take a paragraph and ask:
If I completely delete this section, what do I lose?
Does it contain:
- new information?
- evidence?
- a necessary part of the argument?
- a new perspective?
- a concrete example?
- an actionable insight?
If the answer is consistently “nothing”, then we have a strong signal of content slop.
This is powerful because it ignores vocabulary.
A paragraph can sound sophisticated and still fail the test.
Fancy language isn't information.
Complexity isn't depth.
Length isn't value.
Information Gain > Intellectual Tone
Another useful metric is Information Gain.
Before reading a piece of content, I have some knowledge about the subject.
After reading it, what do I know that I genuinely didn't know before?
If the answer is:
“Nothing. It just confirmed what I already believed.”
we should be suspicious.
This doesn't mean every sentence needs to contain revolutionary knowledge.
It means a long piece of content should justify its existence through some combination of:
new information + evidence + reasoning + perspective + action.
Otherwise, we're potentially just manufacturing reading time.
The Three-Layer Slop Catcher
From an engineering perspective, I wouldn't send everything directly to an expensive LLM.
I'd build a pipeline.
CONTENT
│
▼
┌─────────────────────┐
│ Layer 1 │
│ Cheap Heuristics │
│ Regex / Statistics │
│ $0 │
└──────────┬──────────┘
│
suspicious?
│
▼
┌─────────────────────┐
│ Layer 2 │
│ Semantic Judge │
│ Cheap LLM │
└──────────┬──────────┘
│
high uncertainty
│
▼
┌─────────────────────┐
│ Layer 3 │
│ Delete Test │
│ Deep LLM Analysis │
└──────────┬──────────┘
│
▼
SIGNAL / SLOP
The goal isn't to build another AI detector.
The goal is to build an information-quality filter.
Don't Give Me One “Slop Score”
A single score isn't enough.
I'd rather see something like:
{
"novelty": 82,
"evidence": 71,
"specificity": 91,
"actionability": 64,
"redundancy": 13,
"slop_score": 18
}
Now we know why the content received its assessment.
Maybe the writing is excellent but has no evidence.
Maybe it contains valuable research but repeats itself excessively.
Maybe it introduces a brilliant idea but never explains how to implement it.
A good auditor should expose the weakness instead of simply assigning a label.
And Then There Is Human Thought
The problem becomes even more interesting when we move from articles to our own notes.
How many times do we write the same idea?
How many “new” ideas are actually old ideas wearing new vocabulary?
This is where embeddings become useful.
Imagine comparing every new note against your previous thoughts.
If semantic similarity is high and there is no new information and no action has occurred, we may have detected a cognitive loop.
Not:
“This thought is similar to something I wrote before.”
But:
“I've been thinking about this for three months without producing new information or taking action.”
That's a much more useful signal.
The Bigger Question
This project started as a question about AI slop.
But I think it leads somewhere much bigger.
The internet is rapidly becoming a place where humans and machines produce enormous quantities of text.
The scarce resource is no longer content.
The scarce resource is attention.
If that is true, then the next generation of content infrastructure shouldn't optimize only for:
How much content can we generate?
It should optimize for:
How much noise can we remove without destroying signal?
And that means applying the same standard to everyone.
Human.
AI.
Human + AI.
Doesn't matter.
Audit the output.
Not the author.
The Rule
I would summarize the entire idea with one rule:
Don't detect who wrote it. Detect whether it deserves to be read.
AI doesn't have a monopoly on slop.
Neither do humans.
And if we're serious about building a higher-quality information ecosystem, perhaps the first thing we need isn't another content generator.
Maybe we need a content garbage collector.
The interesting engineering challenge starts there.
What do you think?
Should human-written content be subjected to the same automated quality audit we increasingly demand from AI-generated content?
If you would build this system, what signal would you measure first?
Seyed Alireza Alhosseini Almodarresieh
Top comments (0)