How Far Should What AI Understood Be Applied?I asked for one paragraph to be edited. AI understood my correction exactly, and still rewrote parts I had already reviewed. Understanding the content of a judgment and deciding how far to apply it turned out to be different problems.
Trying to edit one paragraph
While working on a draft, I only wanted to edit one paragraph. I asked AI to delete two paragraphs and to adjust the wording of the last one a little. Along the way, I also corrected one expression about the human role: humans are not only beings with the authority to choose, but need to keep expanding their self-understanding and judgment so they can make decisions worth making. AI understood this correction exactly. But when I asked for the full text with the edits applied, AI rewrote not only the last paragraph but also the title, the introduction, and the earlier sections I had already reviewed. The new text wasn’t necessarily worse, but what I had asked for and what AI did were different.
The difference between understanding and applying
From this experience, I found that understanding the content of a judgment and applying that judgment to the right scope are two separate problems. AI understood the new judgment, human self-expansion, exactly. But instead of applying it only to the last paragraph, it used it as the basis for restructuring the whole piece. The content of the new judgment was right, but the scope it was applied to was different. Even when AI understands a person’s purpose accurately, if it can’t tell which tasks and outputs that purpose should apply to, results the person never intended can follow.
A judgment needs more than its content
| Type | Scope I asked for | Scope AI actually applied |
|---|---|---|
| Delete | The two specified paragraphs | The two specified paragraphs |
| Edit | The wording of the last paragraph | Restructured the title, introduction, and earlier sections |
| Keep | The remaining sentences already reviewed | Changed some sentences and the structure |
| New judgment | Reflect the view of human self-expansion in the specified expression | Used as the basis for re-aligning the whole piece |
In this comparison, the problem wasn’t the quality of the newly written sentences. It was that AI applied the judgment it understood more broadly than the scope I had specified. A request like “Make paragraph A a bit softer in tone” is a judgment that applies only to paragraph A. It doesn’t mean to change the style of the whole document, or to use the same tone in other pieces written later. A person’s judgment can be a correction for a specific part, a criterion for the whole task, or an exception allowed only once. So for AI to use a human judgment in execution, it needs to distinguish not only the content of the judgment but also what it applies to, its scope, and the conditions under which it holds. This piece describes a case observed in a single collaboration, so it can’t verify how AI works in general. What it can show is the difference that appeared when the scope of the human request is compared with the scope of AI’s execution.
What was understood isn’t enough
Accurate understanding alone doesn’t guarantee appropriate execution. Even if AI produced a more natural and consistent result, if it changed parts a human never handed over, that execution may not be appropriate. If AI can solve the problem within the given scope, it should go ahead; if it can only solve it by changing other parts too, it should explain what it would change and what effect that would have, and then return the judgment to the human. For AI to carry a human judgment into real work, it needs to know not only what it understood, but also how far that understanding can be applied and what must stay as it is. This time, the gap between understanding and applying showed up in the small task of editing a single paragraph.