Creative AI

Computer programs are used to generate texts, images, music, designs, solutions to problems, and so on. It is arguable, however, if and when the programs are creative or not. In this talk, I will discuss the creativity of computer programs themselves. Drawing from the field of computational creativity, I provide conceptual tools for the analysis of different aspects of creativity in computer programs. A key insight is that creativity, as a complex phenomenon, needs to be analyzed from different viewpoints – and with creativity.

Today's Large Language Models are often said to exhibit a certain degree of creativity. However, this claim stands in stark contrast to the way these models are trained. Contemporary language models are strictly task-oriented: a neural network is trained on a vast amount of data, with its parameters optimized to produce the most plausible output based on that training data. In this sense, the model merely imitates human language use, leaving little room for genuine creativity. A simple prompt to write a poem—without any creative input from the user—will typically result in something dull and clichéd. Even if the training data contain a large number of poems, the model's primary objective remains to reproduce that data as faithfully as possible, which is counterintuitive to our idea of creativity. Paradoxically, if we want to elicit a degree of creativity from a language model, we need to impose constraints on its output. By limiting the model's possibilities, we encourage it to search for new ways to express itself. In this talk, we will explore how carefully designed constraints can steer large language models towards more creative outcomes, and we will discuss practical strategies to encourage such constrained creativity.

While computational humor generation has long been considered a challenging task, recent large language models have significantly improved the quality of generated jokes. Evaluating humor quality is difficult, as the exact quality is subjective and dependent on the delivery. Another disparity in evaluation standards between human and computer-generated humor is the difference in writing time between the two. In this study, we evaluate the quality of humor generated by GPT-4 with human-written jokes in an improvisational comedy setting in Dutch in a live performance setting on national TV. We compared the ratings of audience members for the human-written and AI-generated improvised jokes for the same audience suggestion and delivered by the same comedian. We found that humor generated by the AI and the human comedians was about equal, which human-written jokes only performing slightly better. Interestingly, AI jokes received more "best joke" votes, suggesting that AI can create standout humorous content. These results imply that current language models can effectively generate relatively high-quality humor, closely rivaling human comedians when put in an improvisational context.

Autonomous comic book generation remains a distant and unattainable goal for algorithms today. Fortunately so—being a comic book artist is the most enjoyable profession in the world, and we wouldnʼt want to hand it over to machines so easily. The reason lies in the fact that generative models still suffer greatly from a lack of control and consistency. However, the craft of making comics is not always thrilling. Flat coloring, for instance, is a crucial yet monotonous step in large comic series, where artists systematically apply base colors to defined areas in a black-and-white drawing, without adding shading or texture. Itʼs a repetitive task that, at first glance, seems ideal for automation, allowing AI to assist artists while keeping creative control in human hands. At its core, it is a segmentation task, but one that operates on datasets that are often small and bear no resemblance to the photographic images traditionally used to train classical segmentation algorithms.