The awkward thing about AI writing is not that it is always bad.
Often it is too good at looking like an article. Clean structure. Positive tone. Smooth transitions. A tidy conclusion. The problem is that after three seconds you can smell it.
You know the rhythm. Start with a broad statement about a changing world. Add a few balanced sections. End with some version of “the future is full of possibilities.” Every claim is safe. Every sentence is polished. Every paragraph seems desperate to prove that it has a structure.
That is not writing. That is the default template.
Two GitHub projects recently hit the trending lists at the same time. taste-skill tries to make AI-built interfaces less generic and more tasteful. stop-slop focuses on removing common AI tells from prose. Hacker News also had a lively discussion around Various LLM Smells.
I think this signal matters. “AI slop” is no longer just a reader complaint. It is turning into tools, workflows, and skill files. In other words, de-slopping is becoming an engineering problem.
What AI slop actually is
AI slop is not one specific word. It is not even one specific punctuation habit.
It is a cluster of habits: summarizing too much, avoiding offense, balancing every argument, forcing ideas into neat groups, and inflating ordinary observations into “major shifts.” It writes uncertain things with too much confidence and turns weak opinions into opinion-shaped paragraphs.
You have seen sentences like these:
In today’s rapidly evolving digital era, AI is transforming the way we create content.
This not only improves efficiency, but also unlocks unprecedented opportunities for individuals and businesses.
Overall, the future is full of endless possibilities.
Nothing is technically wrong there. That is exactly the problem. There is no author in the room.
The issue is not “AI generated.” The issue is “nobody took responsibility.” No specific experience. No point of view. No cuts. No trace of a person who actually thought about the thing.
Readers are getting better at noticing this because they are surrounded by it. Newsletters, social posts, product pages, job descriptions, pitch decks, emails. The same rhythm everywhere. The more common it becomes, the faster people recoil: here we go again.
Why this becomes engineering
A lot of people think de-slopping means asking the model to “write more casually.” That is too shallow.
Real de-slopping means breaking writing into inspectable steps:
- Define the bad smells.
- Turn them into rules the model can follow.
- Use examples to calibrate style.
- Put the rules into the workflow so you are not relying on vibes every time.
That is what makes projects like stop-slop interesting. They do not just say “sound more natural.” They name the failure modes: throat-clearing openings, business jargon, inflated significance, fake profundity, vague attribution, and those perfectly balanced three-part structures that make every paragraph feel assembled.
taste-skill points in the same direction, but for interfaces. Many AI-generated pages now have their own smell: purple gradients, rounded cards, centered hero copy, glowing backgrounds, SaaS sameness everywhere. The issue is not that the pages are unusable. The issue is that they are predictable.
Predictable is the opposite of taste.
Taste sounds subjective, but in a production workflow it has to become rules. Which words do we avoid? Which structures do we delete? When do we keep an imperfect sentence? When should a paragraph be shorter? When should the writer take a sharper position? You cannot just tell a model to “have soul.” That is too mystical. You have to give it counterexamples, boundaries, and review criteria.
That is the engineering part.
De-slopping is not about hiding AI use
There is a bad version of this conversation: make AI text pass detectors, or make it look like you never used AI.
I do not like that framing. It is narrow and a little cowardly.
A better framing is this: AI can help with production, but the final work still needs human judgment. Let the model draft, organize, rewrite, and summarize. Fine. But do not ship the default output unchanged. The default output serves the average case. Your content has to serve a specific audience, a specific account, a specific argument.
AI can save labor. It cannot have taste on your behalf.
That sounds harsh, but it is useful. The valuable part of writing is often not making sentences smooth. It is deciding what not to say, which example to cut, where to make the claim sharper, and where to admit uncertainty.
Models do not naturally do that well. Not because they are doomed forever, but because the default path is safe. Safe writing becomes correct, complete, and boring.
Personal brands will feel this first
If you are writing an internal memo, a little AI smell may not matter. People just need the information.
But if you are building a personal brand, a company narrative, or a product point of view, AI smell is expensive. Those things are not selling information alone. They are selling trust and recognition.
Anyone can say “AI agents are moving into workflows.” The value is in how you say it:
- Which example do you start with?
- Where do you think the trend will land first?
- Which products are mostly hype?
- What judgment are you willing to bet on?
- What have you actually tried yourself?
That is what readers remember.
The biggest risk for creators is not that AI takes away the ability to write. The bigger risk is that everyone uses the same model, the same prompts, and the same pleasant voice until the internet becomes a bowl of warm oatmeal. When that happens, the scarce thing will be the human rough edge: preference, irritation, experience, taste, hesitation, judgment.
A practical de-slopping workflow
If you already use AI for writing, I would change the workflow like this.
First, make AI do the material work.
Ask it to organize sources, extract claims, find counterexamples, and propose structures. Do not ask for a full article too early. Once the model starts drafting, it tends to fall into its article voice.
Second, write your judgment first.
Five sentences are enough. What do you agree with? What do you reject? What should the reader notice? Which example matters most? Where might people misunderstand the topic? Those sentences are the spine of the piece.
Third, let AI expand your judgment, not invent its own.
Be explicit in the prompt: no grand opening, no summary-style ending, no “in this article,” no fake neutrality. Keep concrete examples. Allow short sentences. Allow first person.
Fourth, run a smell pass.
Look for these:
- Does the piece start with a “rapidly changing world” setup?
- Does it lean on mechanical “not X, but Y” contrasts?
- Are there too many perfectly neat groups of three?
- Are there adjectives that add no information?
- Are there paragraphs that sound correct but contain no judgment?
- Is the conclusion just motivational fog?
Fifth, read it out loud.
This is low tech, but it works. AI-heavy writing reveals itself when spoken. It sounds too flat, too smooth, too much like support copy. Human writing does not need to be messy, but it needs to breathe.
The human part gets more important
I do not think de-slopping will remain a niche need. The more common AI generation becomes, the more important editing becomes.
Future content will probably be written by humans and AI together by default. Readers will not care that much whether you used AI. They will care whether the piece has judgment, new information, credible experience, and a specific person behind it.
So de-slopping is not anti-AI.
It is a reminder that generation is only the start. Editing is where the work becomes yours. Tools will keep getting better at writing, but writers need to get better at deleting, shaping, and judging.
A recognizable voice is not something the model gives you.
It is what remains after you keep rejecting the model’s average answer.