Background Sources
This scenario is an entirely fictional case involving GeoAI, a technology company developing a generative artificial intelligence model capable of producing digital artwork from text prompts. Although the company and characters are fictional, the ethical dilemma is informed by research into the use of copyrighted artwork in training generative AI systems, particularly concerns surrounding artist consent, ownership, compensation, and the reproduction of training data.
Moreover, this case primarily concerns computer science and software engineering, particularly machine learning and generative artificial intelligence. Image generation models are trained on large datasets of images to learn patterns that allow them to generate new visual content. However, these datasets may contain copyrighted artwork collected without the explicit permission of its creators, raising ethical questions about the responsibilities of engineers developing such systems.
A common type of generative AI model used for artwork is a diffusion model. Models like this are trained on large datasets, particularly images, and noise is added to an image so the model can learn to remove it. With this, the model can learn to generate new images through text prompts provided by a user. But, before training, engineers have to collect, process, and modify images to prepare the dataset for training. Collecting images from public websites does not mean they have permission to use them, which makes it important for engineers to understand where training data comes from and the permissions they have to use it.
Another issue with these diffusion models is that they can sometimes memorize images used during training. Carlini et al. (2023) showed that diffusion models can sometimes reproduce images that were used in their training, particularly repeated images. Research has also shown tools that address parts of this problem. For example, Shan et al. (2023) developed a tool called Glaze, which makes small changes to images to interfere with AI models attempting to imitate an artist's style. Dubiński et al. (2025) also created Copyrighted Data Identification (CDI). This tool uses statistical methods to investigate whether images were inside a model’s training data. These tools can help protect artists and investigate training practices, but do not fix every issue about copyright or permission.
Our team's research explored this issue from several perspectives, including artists' concerns about consent and permissions, their right to restrict AI training, and the risks of AI models reproducing training images (Lovato et al., 2024; McIntyre, 2026; Carlini et al., 2023). We also looked at possible technical approaches, such as Glaze, which helps protect artists from style mimicry, and CDI, which helps investigate whether artwork was used in AI training (Shan et al., 2023; Dubiński et al., 2025). While these tools provide possible solutions, they do not replace consent, proper credit, or responsible data collection. Our research highlights the importance of considering artists' rights and engineers' ethical responsibilities throughout the development of generative AI systems.