What is an Ai?
First we need to understand what Ai is and what it does in the process. This is not an endorsement for AI use, it has immense problems discussed here: how-ai-companies-use-copyright-infringement-to-steal
To start off: Ai is not sentient, it does not decide what it does nor does it have a vision or ideas in itself. Ai is deterministic in its outputs meaning what is inputted in: will determine what comes out of it. If you use the same prompt, settings and same seed in the same configuration you will get an identical picture everytime.
In this sense it works identically to a digital camera. A person chooses the settings of the camera and presses a button for the camera to create a picture depending on how the light hits the sensor. Fundamentally identical process. Now with a camera person physically goes to the scene they photograph, with Ai the prompt is the scene.
Ai images from datasets, new images?
Ai is not magic, it cannot create everything, but it still can create images and things that are not in the dataset.
Following image example below that is really simplified:
Dataset contains squares and circles. When you prompt a square the Ai will follow the blue line and end up with a square, with a circle prompt it will follow the yellow line and end up in the circle. When you ask it for a combination of square and a circle it will try to create something that matches both and follows the green line and ends up with something like in the image below:
This is the reason a lot of the Ai images feature mangled up artefacts or limbs etc. when as the Ai tries to do something that doesnt perfectly align with something it has knowledge of it tries to go the middle route.
Now this is really simplified but also the reason an Ai with squares and circles in its training data cannot produce a triangle. Hence some things are impossible for Ai to make, and some things may require strange prompts due to what things need to be combined to produce something the user wants under the limitations of the model.
Creation of Ai image process
With Ai there are a lot of different levels of automation. For example with Dall-E you cannot control seed or most of the other settings either (Dall-E seed is randomized to create an illusion of uniqueness), then with tools like ComfyUI you have a lot of more control over those variables and model that is used as well as hardware if you run it locally.
The prompt, if in text format, is just a user friendly format for inputting the values in the Ai. Words inside the Ai model get transformed into tokens / values that respond to certain dataset / weight shapes. Depending on the Ai tool the prompt can also be used to adjust the settings. Using text based interface for interaction with a tool certainly is abstraction, yet not different from voice-controlling a camera a mouse on a computer screen. Method of controlling doesnt make the Ai sentient.
Who is the author?
Now using Ai is mostly just combining shapes that are already in the model and with the same settings anyone can end up with the same image. With custom setups and models+loras run locally this is unlikely though. Someone else might arrive at similar images but not the exact same. To me it looks more like finding a path to an image within the limitations the model sets.
This ultimately a question that comes down to how much automation you personally agree upon tool usage. If you use a washing machine can you say you’re doing the laundry? In the end it works the same way as an Ai as well, you choose what you throw in, what setting and powder you run and the machine does the rest. Not very interesting analogy though.
In the end Ai is not sentient and the user is the one choosing where it goes, whether they actually know how the controls work or not. So in cases of less automation and clear vision, I’d give more credit for the user, in cases of more automation and shorter prompts I’d not as it usually ends up not being what the user’s idea was in the first place, or the idea was so shaky that anything that even barely resembles it: satisfies the user.