I would like to share a feature idea based on my experience using ChatGPT.
When working on a complex project — for example an image, a CV, a presentation, a spreadsheet or a piece of code — I often make many small changes over time. Once part of the project has been approved, I expect a later change to affect only the element I specify.
However, with generative AI, a request such as:
“Change only the window in the upper-right corner.”
can sometimes result in other parts of the image changing as well.
This made me think about a possible different approach.
My idea
Instead of treating the entire result as a single block, the AI could maintain a structured representation of the project, made up of objects, properties and relationships.
For example, a picture of a house could be represented as:
-front wall
-side walls
-roof
-chimney
-door
-individual windows
-garden
-background
Each object could have its own properties, such as position, size, shape, color and material, as well as relationships with other objects.
If I then say:
“Make the upper-right window larger.”
the system could identify that specific window and modify it, while preserving the rest of the scene.
If the wall immediately surrounding the window also needs to change because of the modification, that could be treated as a necessary dependency. But the roof, chimney, other windows, garden and background should remain unchanged.
I think this principle could be applied much more broadly
The same approach could be useful for:
CVs and documents — change only the title color without changing the layout;
spreadsheets — change only the formula in cell B17;
software projects — modify only the authentication module and its necessary dependencies;
presentations — modify one slide element without changing the rest;
images and video — edit one identified object while preserving everything else;
3D models and web pages — modify a specific component while maintaining the existing structure.
The underlying principle would be:
“Do not modify what the user did not ask you to modify.”
Another important part: project state
I think the AI should also maintain a structured understanding of what has already been approved.
For example, if I have already approved:
-a two-column layout;
-a specific font;
-specific colors;
-the position of a photograph;
-the order of the sections;
then these decisions should become part of the approved state of the project.
If I subsequently ask to change only the heading color, the AI should understand that I am not asking it to redesign the entire document.
Why I think this could be valuable
This approach could potentially provide:
-greater precision;
-greater consistency across many iterations;
-fewer unintended changes;
-better user control;
-better preservation of previous decisions;
-and, where technically possible, less unnecessary computation and energy consumption.
I am not suggesting that this is a complete technical solution. It is an idea that came from a practical limitation I have personally encountered while using ChatGPT.
My thought is that generative AI should not only be able to create content, but also maintain a structured representation of what it has created and make localized, incremental changes to it.
I would be very interested to know whether something similar is already being explored, or whether this could be considered as a possible direction for future ChatGPT editing capabilities.
I have also prepared a more detailed document explaining this concept and can share it if useful.
Puolo Fabio