--- title: How to not AI Slop author: Henrik Kniberg date: 2026-05-09 excerpt: Illustrated by two extremes: an unedited voice transcript of Henriks sitting on a rock talking, vs an AI-written article --- # How to not AI Slop *By Henrik Kniberg • May 9, 2026* > Illustrated by two extremes: an unedited voice transcript of Henriks sitting on a rock talking, vs an AI-written article ***Note: This is not a normal article. ****It is a raw, unedited voice transcript of me sitting on a rock in the forest reflecting on, well, how to not write AI slop. Followed by an AI-generated version of an article with the same topic. I want to contrast the two extreme ends of the scale: 100% human created with no edits, vs 100% AI created with no edits. Human slop vs AI slop, if you will. Hope you find it interesting! I did at least :)* ——— ### How to not AI Slop (human version) All right, this is a bit of an experiment. I'm currently taking a walk, or more like I'm sitting on a rock in the forest and just recording my voice speaking out this article. What you're reading now is the unedited version. *Edit: Here is the *[*raw voice recording*](https://cdn.sanity.io/files/3uximhcm/production/6e1c94869f896857734da5851fe41e625e1275cb.m4a)*, if you prefer to listen. Crappy sound quality though, just talked into my iphone. * There is also, for contrast, an AI-generated version. I kinda made the mistake of generating that first. I should have generated it after because I might have gotten influenced by what it wrote. But I basically wanna illustrate the extremes of an article completely written by a human without any edits, and an article completely written by an AI without any human edits. However, when I read what it wrote, I was actually positively surprised. I was thinking that article would be a good example of AI slop, but actually it didn't sound very much like slop at all, so I'm almost a little bit disappointed. But here it is. I'll put a link in here so you can compare, and maybe you can tell me which version you like best. *Edit: Scroll down to see the AI version (or *[*listen to it*](https://cdn.sanity.io/files/3uximhcm/production/415ed01b1bee99ac69a206a779b5f969d25340bf.mp3)*).* But anyway, how to not AI slop. So what is slop? Well, you probably see it and feel it when you read it. I'm thinking specifically about LinkedIn now. There are just some recognizable signs, types of phrases bullet points, emojis. What happens is when everyone uses AI, articles just start sounding kind of the same. They start sounding kind of like meh. And it just doesn't feel as fun to read. Another problem is, other than that it's a little bit kind of boring to read something where it doesn't feel like it came from a human, it feels a little bit disrespectful too if somebody has posted something under their own name but it was just AI-generated, and it seems like they didn't even read it themselves or at least haven't put any effort into making it their own tone of voice. At least I feel like that personally, and when I publish stuff, I try to make it clear that this is me. This is my voice. I may have used AI for some edits and some ideation, but the text is mine, and I want people to feel that. So okay, why should you use AI at all then? If AI slop is a problem, shouldn't we all just not use AI at all? Well, I don't think that's the right solution either, because AI is super useful. It saves time on things like if I wanna do research, ideation, proofreading. There's all kinds of scenarios where AI can both save time and help improve the quality of your work without necessarily turning it into AI slop. So by saving time and improving quality, that benefits everybody. Plus, there's people who just don't like to write, who are just really bad at writing or just wouldn't have written anything at all if they didn't get AI help, and that's a bit sad. I think if somebody has something in their heads, has some idea they wanna convey, and they just don't really know how to put it into words, then I would rather hear their AI-generated version than to not hear their voice at all. So how do we find this balance? Well, before going into that, I wanna mention one more thing. I think it's actually fine to produce pure AI-generated content as long as it is labeled as such. For example, I guess I'm doing it right now because I'm giving you another version of this article that was AI-written, but I'm telling you it was AI-written. In my book, [Generative AI In A Nutshell](/ainutshell), I have some chapters written by Egbert, my sarcastic AI sidekick, and also some blurbs in the book here and there. But they're labeled as such. I'm not pretending it's me. It's Egbert. And I think that's fine. If you know that you're reading something that is AI-generated, it doesn't necessarily even need to be bad. It could be really great content, but it feels more honest to be clear that this is not someone... or this is not an AI pretending to be me or someone else. Okay, so what's my recommendation then? I'm not saying don't use AI at all, but I'm also saying maybe don't use one hundred percent AI. Maybe not have AI... maybe don't use AI to write the whole thing. So how do we find that balance? Well, like I mentioned already, there are some ways that you can use AI without taking away from your personal tone. Ideation is perfect, right? I have some ideas. I'm not sure what to write about, or maybe I do know what I wanna write about, but I want some ideas about what... maybe some key points to make. Normally, that's what I would have done before an article like this, except right now, I kind of want to illustrate the extreme cases. So for this article, I did no research. I'm just sitting on a rock in the forest talking freely. If you're writing about a topic like, for example, I have another article I'm working on about universal basic income. I don't know very much about that topic, so using AI to research that is super useful. It basically makes It improves my knowledge so that I can produce a better article, and then everybody wins from that. Feedback is useful, to find out is this article suitable for the intended target audience? So after you've written the article, you can say, "This article is targeted at middle managers at tech companies. Give me feedback from the perspective of that kind of person." That's usually super useful. There are some ways you can use AI that are useful, but a little bit of a slippery slope, so be careful. For example, tweaking the text. Let's say I have a paragraph that is rather long and rambly. Asking AI to shorten it is probably fine. It's still gonna probably stick to my tone of voice. But it's a slippery slope if you do that too much. If you have AI kinda rewrite everything to make the language better or something, then at some point it becomes no longer your voice, and it starts becoming AI slop. In fact, even if it's not AI slop, even if it's a very well-written AI article that doesn't even feel like slop, again, it's not quite your... it's not you anymore. So be careful with that. But like I said, balance, right? So for example, let's say I'm using a metaphor in my article, and I don't think it's a very good metaphor. I can help... AI can help me find a better metaphor. Or maybe I want to find an example of something to illustrate a point.AI can help me come up with that example. A good kind of guideline is start from your text. Don't have AI write the article first and then you tweak it. Do vice versa. So if you speak your article like I'm doing now or just write something randomly, I mean, even if you hate writing, even if you don't think you can express yourself, I'm pretty sure you can talk, I'm pretty sure you're able to write. So just write down a really crappy first draft, unedited version of whatever random thoughts you had in your head. It can be super unstructured, but it is *your* voice. It is you making a point in your way, and then you can use AI to kind of structure that, clean it up a little bit. And like I mentioned, as long as you don't go too far with that editing, then your voice will still be in there somewhere. It's your points being made in your way. So that's a useful approach. In fact, that's how I write a lot of my stuff. I take a walk, I brainstorm, I get a really messy pile of notes. AI helps me structure it into, like, a synopsis, and then I write it myself based on that synopsis using my words, and then back to AI to help me tweak some of the text a little bit. But the key thing is when I read that text afterwards, and I read every word of it, then I need to feel that this is me. These are *my* words.This is *my* voice. So I think that's an overriding key point here. Whatever it is you create, and no matter how much or how little you use AI - at the very least, read what came out, read the final result, and kind of consult your gut feeling. Does this feel like my voice? Does it feel like me? If not, then maybe don't publish it. Or maybe do publish it, but make it clear that this is, you know, me plus my AI friend or something just to make it clear that it's not entirely your own voice. Other tips, use good models. There's a very big difference. I think a lot of kind of LinkedIn AI slop is out there because of using cheaper models with like lazy prompts and not really putting effort into it. And that's when you get these generic articles that all sound the same. And that touches upon another point, by the way. I may be a little bit old-fashioned, but I feel like if I'm reading something, I will appreciate it and respect that text a lot more if I know that the person who wrote it put effort into it. There's some kind of intrinsic value to effort. It's like if I buy a painting and I know the painting was painted by a real person putting time and effort into creating this work of art, that's going to feel more worth it. Same with listening to music. If I know there's humans sitting there tuning the instruments, playing the instruments, doing the mixing, it'll just feel more valuable to me. So I'm not saying there's anything really wrong with AI-generated music or AI-generated art or AI-generated articles, just that as a consumer, it'll feel more valuable if I know the effort was put into it. So in a similar vein, if I'm listening to, if I'm watching, if I'm looking at some AI-generated art, if I know that the person behind put a lot of work into that with prompting, follow-up feedback, using different tools, then it'll just feel more valuable. That doesn't necessarily make any rational sense. Maybe it's just the way we humans work. So keep that in mind, right? If you just spit out content like, you know, a new article every hour without even reading it yourself, chances are your readers won't really feel like they respect that text. It's probably better to produce a little bit less, but put more effort into it. So there are some exceptions. Sometimes it doesn't need to be your voice. Sometimes it doesn't need to be human written. For example, product documentation. In our product at Abundly, we now use AI to generate [documentation](https://docs.abundly.ai/introduction/welcome) and to generate [changelogs](https://docs.abundly.ai/changelog). It doesn't look like AI slop to me when I read that. We put a lot of effort into the prompts, into the context, using good models so that it produces documentation and change log that is really useful and correct and doesn't feel like boring AI slop. But technically, it is 100% generated and it's not pretending to be written by a human. And that's kind of the key point. Also, it did take some effort to get those prompts in place so that it is really good. So yeah, I hope this was useful. Was there anything else I wanted to say? I think at the end of the day, it really boils down to, do you want this piece of content to be yours? Should it feel like you? Can you stand for this? If so, put in that little bit of extra effort and read it yourself honestly, or maybe ask a colleague or friend to read it and just kind of check, does it still feel like mine? Okay, I'm scrolling up and reading my notes. What were the key points I wanted to make here? Yep, I think that was it. So, yep, good luck with your content creation, and I hope you enjoy this this article. ——— Okay, here's me again. I just listened through the recording. I'm still on my rock. And I just wanna add one thing. Don't be afraid to publish content, an article or something else that is kind of raw and unpolished. Maybe there's spelling mistakes in there. Maybe the grammar is a bit wrong. That's great. I honestly think people will appreciate that more. When everything starts sounding like generic, perfectly polished AI text, then raw, messy human content I think will be more and more appreciated actually. Maybe not as extreme as this pure raw recording with no edits at all. You know, a little bit of editing is probably good. But generally speaking, I would say don't be afraid to just get your thoughts out there, even if they're not quite polished. ## AI Generated version *The article below was written using a single prompt with Claude Opus 4.7. No edits. Prompt: Write an article titled ”how to not AI slop” which is about how to use AI help in article writing, without it turning into AI slop.* [Voice narration](https://cdn.sanity.io/files/3uximhcm/production/415ed01b1bee99ac69a206a779b5f969d25340bf.mp3) (using ElevenLabs). ### How to Not AI Slop (AI version) You can usually spot it in the first paragraph. The triple parallel construction. The em-dash where a period would do. A phrase like "navigate the complex landscape of." Hedge after hedge. By the third sentence the reader's eyes have glazed over and they don't even know why. That's AI slop, and it's not actually about AI. It's about outsourcing the parts of writing that are supposed to be hard. This matters because AI is genuinely useful for writing. The problem isn't the tool. The problem is that most people use it backwards: they ask the model to generate prose and then they edit, which is the one workflow guaranteed to produce slop. The fix is to flip it. ### What slop actually is Slop is text that's grammatically fine and informationally empty. It has the surface features of writing — sentences, transitions, conclusions — without the substance. It hedges instead of arguing. It abstracts instead of describing. It balances when it should pick a side. It explains the same point three different ways under three different headers. The reader finishes and can't remember a single specific thing. Some specific tells: - The "it's not just X, it's Y" construction. Once you see it you can't unsee it. - Tricolons everywhere. Lists of three are a real rhetorical device, but AI defaults to them whether or not the content has three things in it. - Words like *delve, tapestry, harness, unleash, navigate, leverage*. Not banned, but if you have three in one piece you've been caught. - The ceremonial conclusion. "In conclusion, X is a complex topic that requires careful consideration." This is throat-clearing in a dinner jacket. - Bullet points that should be sentences. Bullets fragment ideas that have causal or logical flow between them, which is most ideas worth writing about. - The hedge sandwich: every claim wrapped in "while it's true that," "it's important to note," "however," "ultimately." The slop signature is so consistent that a careful reader can usually tell within a paragraph whether AI was driving. ### What AI is actually good at Reframe the model from co-writer to research assistant and editor. Specifically, it's good at: - **Research and synthesis.** Pulling together what a field thinks about something, then telling you who disagrees. You still verify, but the legwork goes from hours to minutes. - **Stress-testing your argument.** Paste your draft and ask: "what's the strongest objection to this? where am I hand-waving?" The model is much better at finding holes than at filling them. - **Line edits with specific instructions.** "Cut every hedge in this paragraph." "Find sentences over 30 words and shorten them." "Where am I using passive voice and shouldn't be?" - **Structural feedback.** "Does my third section earn its place or am I repeating section two?" - **Proofreading.** It catches typos, tense slips, subject-verb mismatches, the comma after the dependent clause. Notice what's not on the list: writing the actual sentences. That's the part where your voice lives. If you outsource it, the piece will sound like nobody. ### A workflow that doesn't produce slop Write the first draft yourself. Badly is fine — badly is even good. The first draft is where you find out what you actually think, which the model cannot do for you because it doesn't know. Once you have a draft, use the AI to interrogate it. Ask for the weakest argument. Ask what a smart skeptic would push back on. Ask whether your thesis appears in the opening or only emerges by paragraph six. Ask which examples are doing real work and which are filler. Then edit yourself, by hand, in your own sentences. If the model suggests rewrites, treat them as diagnostic — "oh, this paragraph is fixable" — not as the fix itself. Pasting the model's rewrite is how voice dies. For polish, use the model as a line editor with narrow instructions. "Find every instance where I use 'really' or 'very' and flag it." "Mark sentences that start with 'It is' or 'There are.'" Specific surgery, not vibes-based polishing. ### The specificity test Here's a quick test for whether something is slop. Take any sentence in your draft and ask: could this sentence appear, unchanged, in an article about a different topic? If yes, it's probably empty. *"AI is transforming the way we work"* — fits anywhere. *"AI is changing how junior associates draft contracts, by surfacing precedent in seconds and getting clause language 70% of the way there"* — only fits one place. Slop fails the specificity test almost everywhere. Good writing passes it almost everywhere. The fastest way to de-slop a draft is to go sentence by sentence and replace anything generic with something concrete — a name, a number, a date, an example, an object you can picture. ### On voice Voice is the thing AI can't give you and won't fake well. Voice is how you sound when you're not trying to sound like anything. It's your weird metaphors, the things you find funny, the arguments you find tedious, the words you use that nobody else uses quite that way. If you let the model write your prose, you'll get its voice, which is the average of every blog post on the internet. The texture goes flat. People can feel this even when they can't articulate it; they just bounce off the piece without knowing why. So: keep your hands on the keyboard for the actual sentences. Use the model for everything around them. The slop test isn't whether you used AI — almost everyone does, eventually. It's whether the piece sounds like a person. ## Appendix - Henrik’s reflections Comparing these two, I think my version is a bit painful to read because it is long and rambly (since it is just a raw voice recording). I normally wouldn’t publish something like that. The AI version is surprisingly good for being a single prompt with no edits. I agree with all the points, but it still isn’t my article, so I wouldn’t put my name under it. And it is missing some points that I consider important. To get the best quality, I would take my original voice transcript, and then get AI help to shorten it and make it a bit less rambly. I would use a mix of manual edits and AI edits, iterating until I’m happy with the text. I would also ask AI to suggest if there are any other important points I might bring up, and then I would write those myself, with minor AI assistance. The final test would be that I could read the whole article myself without feeling bored or losing focus along the way, and still feeling like it is my words. But I won’t do that here, since the whole point of this article is to compare the extremes - raw human content vs raw AI content. --- *Read the full article at [https://www.abundly.ai/blog/how-to-not-ai-slop](https://www.abundly.ai/blog/how-to-not-ai-slop)*