{"id":230562,"date":"2026-08-27T11:41:00","date_gmt":"2026-08-27T11:41:00","guid":{"rendered":"https:\/\/www.9senses.ai\/?p=230562"},"modified":"2026-09-06T14:30:29","modified_gmt":"2026-09-06T14:30:29","slug":"reviewing","status":"publish","type":"post","link":"https:\/\/www.9senses.ai\/de\/reviewing\/","title":{"rendered":"The AI Editing Merry-go-Round"},"content":{"rendered":"

Editing has become much easier with the arrival of capable AI models that process language well, at least for those languages that are properly represented<\/a> in its training corpus. Previously, finding the last typo or grammar error was painful, for both author and editor, as the human brain is extremely capable of making sense of something, but not of spotting what might be wrong in a sea of letters. <\/p>\n

Nowadays, we send our essays, papers and post to the AI and receive instant feedback. And if someone sends us something to review, we quickly feed it into Claude, Gemini or ChatGPT. Et voil\u00e0: a concise, thorough list of what to correct and improve is ready within minutes.<\/p>\n

A lot of time saved. Or not<\/p>\n

The sanding paper trap<\/h2>\n

For authors who submit their work to a leading model, they can indeed expect that pretty much every typo and grammatical error is caught by two passes at most. And if they send the text to the model, it can even provide the edits without a need to later look up the exact spot in the document where the change has to be inserted.<\/p>\n

Unfortunately, that’s only half the truth. Depending on how you prompt the model, you really just get back the typos and narrow grammar inconsistencies. But as soon as you ask for advice beyond that, for example for style or logic, AI feedback reminds us of those bad editors we encountered at the times of human editing. Let’s quickly recall. A good editor is one who finds the flaws in an author’s argument and language – long-winded sentences, inappropriate metaphors, word repetitions, and other things that need weeding out. They do not touch the writer’s style and rhythm, as long as it doesn’t obviously required. Bad editors are those who push their own language and style choices onto your text, leaving you with this dreadful feeling of having to fight someone who was supposed to help you.<\/p>\n

AI, as soon as you give it more leeway on editing suggestions, has more of the latter qualities. If allowed to do so, it shifts a writer’s language away from their personality, and towards more generic, boilerplate language. Without asking, AI is brutally sanding off all the edges that make a text legible and interesting. And if you don’t pay attention, some AI models are absolutely capable of rewriting most of your text even after a simple instruction of “focus on typos and grammar issues only” – without even mentioning them doing so.<\/p>\n

This sanding off effect is even scientifically documented. A “Nature Human Behavior” study published on August 24, 2026 examined more than 880,000 texts and found that LLM-assisted rewriting reduced linguistic variation while generally preserving underlying content. The researchers report reductions of 21\u201350% in writing-complexity variance, depending on dataset and model.<\/p>\n

And an additional feature of AI-supported editing: once a writer has accepted the edits suggested by the AI chat, a new pass will generate further suggestions that sometimes contradict previous ones, or sand off more of the personal style.<\/p>\n

The sycophancy trap<\/h2>\n

If you want to ensure that your work doesn’t get washed down to something that sounds like anyone, you can take the AI on a trip and argue each suggestion. Typically, you will end up either with the AI arguing with you stubbornly, or it acceppting pushback that isn’t warranted. This is called sycophancy, and it was a feature of many models until recently. Agreeing to pretty much everything users said to please them. Later models have been specifically trained to evade this trap initially, and so you don’t have to expect that initially. However, as soon as you begin arguing, it is very easy to end up in the same place, where the edits are becoming meaningless after a few rounds: which ones truly matter, which ones are just style preferences, and where the AI just doesn’t have the right data to follow you?<\/p>\n

The editor’s trap<\/h2>\n

Using AI to help someone polish their article, essay or resume is born from an understandable reflex. It saves a significant amount of time, is more precise when it comes to typos and grammar, and you can provide substantive feedback within a much shorter time.<\/p>\n

But if you do that, what are you actually supplying? First of all, nothing most people can’t get themselves from ChatGPT, Gemini or Claude without asking you. You’re also not supplying your own insights, you are just providing a mechanical pass on something that doesn’t reflect your exposure to the text, your interpretation, your struggles, your takeaways. <\/p>\n

So you are adding only little value and possibly are creating significant damage to the author’s further progress, because your AI-related feedback is nothing that helps them improve, except for the aforementioned typo and grammar police.<\/p>\n

The AI editing race<\/h2>\n