Teaching a Small AI to Drop Its Familiar Writing Habits
Qwen3-4B-Instruct
The original AI model used for the experiment.
slop score
A score that counts selected writing habits.
LoRA
A way to train a small part of an AI model.
AI-generated prose often repeats recognizable habits. It may use em dashes, “not X, but Y,” lists of three, instant rhetorical questions, or signposts such as “here’s the thing.” Erik Markus Kannike tested whether those habits could be reduced inside Qwen3-4B, a small text-writing model. His report describes a released model called Qwen3-4B-deslop v4.
What happened
Kannike started with Qwen3-4B-Instruct and trained it toward plainer prose. The smaller 2.5 GB build can run with Ollama, a tool for running AI models. This was a personal experiment on a MacBook Air laptop, not a company product announcement. The goal was to reduce writing habits without adding a system prompt each time.
How the test worked
The author created a slop score. It is a weighted count of selected habits per 1,000 words. The test used 40 prompts held out from training. They covered ten topics and four genres. Each prompt produced one sampled answer, with no system prompt.
The original model averaged 121. The released model averaged 12.5. Em dashes fell from 20.5 per 1,000 words to zero. Sixty-eight percent of its answers landed in the author’s human-writing range. It also answered 30 of 30 factual questions correctly. It matched the requested genre and length about as often as the original.
Why the training was harder
The process used a teacher model to produce answers. Only answers that passed style and damage checks were kept. A LoRA adapter learned from them. The author also combined the adapter with an edit to four neurons.
An intermediate version exposed the danger of optimizing surface patterns. It scored 9.9, yet opinion pieces avoided taking a position. A city-move story became a family story. A plant article cited an invented study. The author added task-fit checks. They looked for an early opinion, first-person writing, email greetings, reasonable length, and invented studies or percentages. The released model’s task-fit rate returned to 80%.
What the result means
The experiment suggests that a model’s writing style can be changed during training. It does not show that the new style is better. The released voice is plain and sometimes flat. It still invents names and places. It also often runs longer than requested. Removing visible habits can therefore create a different weakness.
What is confirmed, and what is not
The numbers come from the author’s own detector and rules. The detector is a style guide expressed as regular expressions. The training data was filtered with some of the same checks, so the task-fit result may look better than a fresh evaluation. Scores also moved between runs because of sampling.
The study covers only short English prose in four genres. Code, long conversations, other languages, and tool use may behave differently. No blind reader test has yet shown which version people prefer.
The post drew 155 points and 110 comments on Hacker News, a technology news site. Those figures show community attention. They do not prove the experiment is correct.
What to watch next
Kannike plans to use stronger human writing as training material. He also plans versions around 7–8B and 30B, plus a blind comparison where readers choose the better text. The useful question is no longer only whether AI writing looks less artificial. It is whether the model can keep meaning, task fit, and a distinct voice at the same time.
A Small AI Learned to Avoid Some Familiar AI Writing Habits
📰 Full story: Teaching a Small AI to Drop Its Familiar Writing Habits
A Qwen3-4B model was trained to use fewer repeated writing patterns.
Qwen3-4B
The name of an AI model that writes text.
slop score
A number that counts repeated writing habits.
Hacker News
A website where people discuss technology news.
💡 The gist
- A small model learned to avoid several common AI writing habits.
- Its habit score fell from 121 to 12.5 on 40 new prompts.
- It still sounds flat and sometimes invents names and places.
AI writing often repeats certain patterns. It may use long dashes. It may build lists of three. It may use the same sentence shapes.
Erik Markus Kannike, the author, changed Qwen3-4B-Instruct. Qwen3-4B is a model that writes text. The new version is called Qwen3-4B-deslop v4. A 2.5 GB version can run with Ollama, a tool for running models. The experiment was done on a MacBook Air.
The author made a slop score. It counts selected writing habits per 1,000 words. The test used 40 prompts. The model had not trained on those prompts. They covered ten topics and four kinds of writing.
The old model averaged 121. The new model averaged 12.5. Em dashes fell from 20.5 to zero. Sixty-eight percent of answers reached the author’s human-writing range. The model also answered 30 of 30 fact questions correctly.
The work needed more than deleting words. An early version scored 9.9. It still missed the task in important ways. It avoided opinions in opinion pieces. It invented a study in one article. So the author added checks for task fit. These checks asked whether the answer followed the requested form.
This shows an important reason. A model can lose one habit and gain another. A clean-looking answer can still miss the point. The released model uses plain, sometimes flat sentences. It still invents names and places. It also often writes longer than requested.
The test has limits. It covers only short English prose. Code, long chats, other languages, and tools may differ. The score uses rules written by the author. No blind reader test has compared the texts yet.
The post received 155 points and 110 comments on Hacker News, a technology news site. Those numbers show attention. They do not prove the result.
The author plans larger models. He also wants better human writing as training material. A blind reader test is planned next. It will ask whether people prefer the new writing. It will also show whether fewer AI habits mean better writing.
💬 A simpler summary of the model that removes AI-like writing
People liked the goal, but they disagreed about whether the writing became genuinely better. The moving cursors on the page were also distracting to many readers. The performance and detector results are user reports from the comments.
- Some readers liked that the model could be trained with little computing hardware, although the thread did not independently verify this.
- Some people saw a small improvement, while others still found the text clearly AI-like and hard to read. One suggested using shorter words.
- One commenter reported that Pangram rated it 100% AI-generated, while slop-cop rated it more cleanly. Another said the rewritten example was about three times longer, with the same information and no formatting.
- People split over the purpose: hiding AI style might hurt trust in online writing, but cleaner text could help with documents, summaries, assistive communication, and clearer thinking.
- Many disliked the shared cursors. The author said they came from a WebRTC/Trystero experiment and later turned them off by default.
initial digest at 110 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
A Little AI Practiced Using Fewer AI-Like Phrases
📰 Full story: Teaching a Small AI to Drop Its Familiar Writing Habits
A small text-writing AI learned to repeat some phrases less often.
Qwen3-4B
The name of a small AI that writes words.
Hacker News
A website where people talk about computers.
Qwen3-4B is a small AI that writes words. Erik Markus Kannike, the author, changed its writing habits. The AI used some phrases again and again. The author counted those phrases. He tested the AI with 40 new writing tasks. Its score fell from 121 to 12.5. Long dashes fell from 20.5 to zero. The AI still answered 30 of 30 fact questions correctly. But it still made up some names and places. Its voice could sound flat. People have not compared the writing secretly yet. The post got 155 points and 110 comments on Hacker News. Hacker News is a website about computers. Those numbers show attention. They do not show that the test is right. The author wants to try bigger models. He also wants people to choose between two writings.
💬 A story about an AI model with fewer AI habits
This model tried to make AI writing sound less like typical AI writing. People did not all agree that it worked.
- Some people liked that it could be trained with a small amount of computer equipment. Others said it still sounded like AI and was hard to read. One person reported a 100% AI result from one checker, a cleaner result from another, and a rewrite that was about three times longer for the same idea.
- Some people worried that hidden AI writing could make online conversations less trustworthy. Others said it could help make documents, summaries, and communication tools for people who need assistance.
- The page showed many moving cursors. Readers found them distracting, so the author said they were an experiment and hid them by default.
initial digest at 110 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
💬 HN reactions to Qwen3-4B trained to reduce AI writing tics
Commenters liked the idea and the small-model training result, but reports on readability and human-likeness were mixed. The page’s shared cursors also distracted many readers. All performance, defect, and detector claims below are commenters’ self-reports, not independent validation.
initial digest at 110 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.