13. How to Make AI Writing Sound Less Formulaic
Last updated:
Once I started using AI to help me write, I noticed a few patterns.
Pattern 1: The deep opening
It opens with something that sounds profound: "In this era of rapid change, X is redefining how we ..."
Or: "Most people think X. The real problem is Y."
Pattern 2: Announcing the thesis right away
For example: "What really matters is not the tool but the mindset."
There are more examples, and they like a clean binary contrast:
- Not A, but B
- The real problem is not X, it is Y
- The point is not X, it is Y
Every section also follows the same shape: claim, explanation, example, summary.
For example:
AI lowers the cost of coding. That means ... So the real change is ...
Everything is as complete as a textbook. Even teachers do not talk like that.
Pattern 3: Transitions
AI leans hard on transition words, and it shows up most in Chinese:
- However
- At the same time
- More importantly
- In other words
- From this angle
- It is worth noting
- This means
- In the end
(Who writes with that many transitions? They land out of nowhere.)
Pattern 4: Manufactured punchlines
For example:
AI will not make engineers obsolete. It will make engineers who do not use AI obsolete.
The future belongs to those who can steer AI, not those who compete with it.
LLMs are also stubborn about parallelism.
Build faster, test ideas more cheaply, iterate more efficiently. It is not a tool, not a trend, but an entirely new way of working.
These lines look great in a screenshot. Read a whole AI-written article and every paragraph runs on the same beat.
Pattern 5: The forced uplift at the end
The ending reaches for something profound even when nothing needs lifting, just to show off the prose. It is close to the AI version of "let us join hands for a better tomorrow." Funny.
In the end, what we need to consider is not where AI is heading, but who we want to become.
The future is already here. The real question is whether we are ready.
There are ways around it
Plenty of people have noticed this problem and put together fixes.
The first one I found was blader/humanizer. It comes from Wikipedia's Signs of AI writing, the taxonomy of AI writing traits that WikiProject AI Cleanup has maintained for years. Many other skills treat it as a base rule set.
There is a Chinese version too, jiji262/humanizer-chinese, built on blader/humanizer and supplemented with a few papers. It works reasonably well for me and kills a lot of the usual AI article shapes, including every pattern above.
Another Chinese option is op7418/Humanizer-zh.
For English writing, add both petergyang/no-ai-slop and blader/humanizer to your agent's skill set. Write the piece first, then translate it into other languages and tell the agent to translate for meaning. That avoids a lot of slop.
Even with those skills restricting the AI, some AI smell survives.
A test
I ran a test: same prompt, one run with the skill loaded and one without, capped at 200 characters.
Here I used only op7418/Humanizer-zh.
Prompt: Why does AI make writing easier while making a personal style harder to reach?
A | Plain AI
AI makes writing easy because it can open for us, build the frame, and connect scattered thoughts into an article. The difficulty comes with it: once a sentence is already smooth, we cannot be bothered to think about how we would say it. Over time the article reads well and the writer's own traits fade. Personal style hides in concrete choices: which word you prefer, how you tell a small thing, what judgment you are willing to keep. The writer still has to make those choices.
B | After Humanizer
AI can produce a smooth first draft in seconds, which saves the time spent staring at a blank page. The trouble is that ready-made sentences are too easy to accept. You wanted to write "I do not like it," saw it write "there is still room for improvement," thought that was fine, and left it. After a few rounds the attitude fades and the wording cannot be faulted. I think the hardest part of personal style to keep is these small choices. After writing with AI, you still have to go sentence by sentence: is this what I mean, would I say it this way?
Some AI flavor remains, just less of it. The next step is to add back the human parts.