The Presence of Creativity and Intentions in the Production of AI
As artificial intelligence became present in our society, my question, which was also debated in the class Interaction Design, was about the potential of creativity in artificial intelligence. If artificial intelligence possesses creativity, it will become more than just a tool for us, with outputs becoming influential in a multitude of fields.
Before I jump into the results throughout my hands-on sessions, I would like to clarify the definition of “creativity” in this article. In my personal opinion, creativity refers to an object that retains some intention. In other words, the piece of art is essential to hold some kind of meaning to become defined as creative. In this article, I will be analyzing whether AI is proficient in possessing creativity from the results in the models trained by AI that I have utilized throughout the course.
Results from Models in Final Presentation
In my final presentation, I demonstrated a picture book cover generating an AI model to see whether artificial intelligence is capable of interpreting the user’s intention. In the process of generating the images, I used Runway ML and DALL•E Mini.
Initially, the model from Runway ML was trained by approximately 500 images scraped by a source code made in Google Colabolatory, using Python. (*1) After scraping the images, the datasets were trained on Runway ML with 3000 training steps with an illustration preset training option provided by Runway. As a result, the trained model was inefficient in recognizing the attributes of the cover of a picture book. (*2) As we can see in the figure, although the model understands some of the characteristics of picture books, it could not generate the basic aspects of a picture book, which are the title text of the story and a cover drawing to illustrate that well describes the narrative of the book. Although the model understands that there were texts in each data, the model does not generate texts in the outputs but rather a text-like bizarre shape, that can also look like bubbles. From the abortive outputs, we decided to narrow down the input datasets to scrape from “picture books” into “panda picture books” to facilitate the model in apprehending our intentions. Nevertheless, the outputs(*3) for the “panda picture books” were also unsuccessful in generating the images we were expecting to see, illustrating texts as text-like figures. Moreover, the model failed to identify the illustration as pandas and ended up generating disturbing black and white creatures.
(*1) https://colab.research.google.com/drive/1oUVAXeKMFWWfwFlBllGA-pZ7BR-RwCoA?authuser=1&hl=ja
(*2)
(*3)
Apart from the models made by Runway ML, we also generated picture book images from DALL•E Mini, which is an AI system that can create realistic images and art from any text prompt(*4). Following the Runway ML models’ results, the input we tested was “panda’s journey to the supermarket”. Based on the results from the models we trained in Runway ML, we added a story in the text input to make it comprehensive, as well as make the AI interpret the intentions of the text simply. The outcomes that we achieved from DALL•E Mini were substantially more accurate than the results of the trained models. However, none of the images initiated from DALL•E Mini exceeded the explanation or story that we have provided in the text. As a tool, accuracy is the most essential aspect, but in terms of creativity, the AI is required to have its intention aside from the information we impart.
(*4)
If the model was specialized in generating picture books, the outputs would have ended up becoming more precise. However, if the algorithms for AI are constructed to generate picture book illustrations, the models will be incapable of producing images that exceed our imagination. What we can imply from the sequels of the model, is that AI models are impotent in comprehending the intentions that we expect. Furthermore, even if the AI models accidentally generate something exceptional from the input, there is no intention inside the product, nor it is measurable.
So… is AI creative?
Consequently, throughout Interaction Design and from the results and the outputs of the models Woojin and I have trained, my conclusion was that artificial intelligence does not possess creativity. The outcomes, we obtained through the process of generating images from several amounts of models. Nonetheless, we cannot conclude that AI-constructed models will never subsequently possess creativity. While artwork generated by models with artificial intelligence does not acquire any intention from the AI, we cannot assure that algorithms that admit AI to originate ideas on their own are yet to be found hereinafter. Currently, artificial intelligence is acknowledged as a convenient and practical tool, but someday, AI might become influential just as us.