Can AI Be as Creative as Humans?
Creativity is the kaleidoscope through which humanity views its past, shapes its present, and imagines its future. It’s the engine driving our society forward, pushing the boundaries of science, technology, and the arts. As we sail through the digital age, the horizon of creativity expands with the advent of artificial intelligence (AI). The once-clear line distinguishing human from machine-created work becomes increasingly blurred, raising profound questions about the nature of creativity itself.
M.B.A. Students vs. ChatGPT: Who Comes Up With More Creative Ideas?
In the groundbreaking study [1], Christian Terwiesch and Karl Ulrich explore the realm of innovation by pitting M.B.A. students against AI, specifically ChatGPT, in a competition to generate innovative ideas. To assess who excels in idea generation — humans or machines — they compared 200 ideas from students with 200 ideas generated by ChatGPT4. The results were staggering: ChatGPT easily outperformed humans in terms of quantity, producing 200 ideas in just an hour. For the top 10% of ideas, ChatGPT dominated with 35 out of 40 ideas.
Alan Turing, a pioneer of computing, once speculated on the potential for machines to compete with humans in intellectual fields. This curiosity leads us to explore not just whether AI can be creative, but how we can understand and measure this creativity against human standards.
Recent research [2] proves in theory that AI can be as creative as humans if it can properly fit the data generated by human creators. In the following, we will introduce this new research result about AI creativity.
The Challenge of Defining Creativity
Defining creativity has long been a complex endeavor, with scholars from psychology, philosophy, and cognitive science offering various perspectives. Despite this diversity, a common thread is the recognition of creativity as a blend of novelty and quality. Yet, what constitutes “novel” and “quality” work varies significantly across cultures and disciplines, making a universal definition elusive. This challenge is amplified when we attempt to understand creativity in contemporary AI models, as these systems generate outputs based on their training data, blurring the lines between original creation and sophisticated replication.
A Framework for AI Creativity
To navigate these complexities, we propose a new approach to understanding AI creativity. We introduce the concepts of Relative Creativity and Statistical Creativity, offering a framework that moves beyond the binary question of AI’s creative capacity to a comparative analysis with human creativity.
Relative Creativity: Imagine an AI that creates a piece of art without prior exposure to similar works. If this artwork, discovered decades later, mirrors a piece independently created by a human artist, does this not signify the AI’s creative prowess? This thought experiment underpins the concept of Relative Creativity, where an AI’s output is considered as creative as that of a hypothetical human creator if it is indistinguishable from that of the human creator, according to an evaluator.
Statistical Creativity: Expanding from the concept of Relative Creativity, Statistical Creativity utilizes statistical analyses to benchmark AI-generated creations against the works of a collective of human artists. This approach facilitates a quantitative evaluation of AI’s creative capabilities, determining if AI’s creativity matches that of a specific human creative ability.
To explore the potential for creativity in state-of-the-art Large Language Models (LLM), believed to possess some capacity for creative output, researchers delve into the concept of statistical creativity within autoregressive models, utilizing prompts as contextual framework (termed Prompt-Contextualized Autoregressive Statistical Creativity). The term “Prompt-Contextualized” signifies that the model’s responses or outputs are shaped by the specific prompts or contexts provided; meanwhile, an “Autoregressive” model refers to a model that generates sequences one element at a time, basing each new prediction on its predecessors.
The question arises: Can AI models achieve a level of creativity that rivals human creativity through training, and if so, what processes are involved? In the previous study, researchers defined the creativity standard statistically. Essentially, if the AI can meet the standard in its test performance, it is considered to possess creativity on par with that of humans. Utilizing statistical learning theory, the researchers have quantified the volume of data required to be fitted during the training phase to meet the standard of statistical creativity (Creativity Emergence Conditions). Furthermore, in the journey from training to generalization, a critical insight emerges: the necessity of gathering and fitting data that encapsulates the conditions under which creations are generated.
Theoretical and Practical Implications
1. AI Can be as Creative as Humans
Our journey begins with a groundbreaking theoretical revelation: AI possesses the potential to match human creativity, provided it can assimilate an ample amount of conditional data created by human minds. This finding suggests a future where AI, fueled by its learning capabilities, could generate creative outputs that are novel and potentially indistinguishable from those that might be produced by future human creators. It’s a vision that redefines the boundaries of creativity, highlighting the transformative power of AI in generating unseen, innovative creations.
2. Unveiling Relative and Statistical Creativity
At the heart of this exploration lies the introduction of two novel concepts: Relative Creativity and Statistical Creativity. These concepts serve as the compass guiding us through the complexity of AI creativity. Relative Creativity offers a lens to view AI’s creative outputs in comparison with those of hypothetical human creators, moving beyond absolute definitions to a more nuanced, relative understanding. Statistical Creativity, on the other hand, provides a quantifiable measure to assess the extent of AI’s creativity by comparing its outputs to those of a specific human population. Together, these concepts form a robust framework for exploring the creative capabilities of AI.
3. Embracing the Subjectivity of Creativity
A pivotal realization in our journey is the acknowledgment of creativity’s inherent subjectivity. By integrating this subjectivity into our comparative analysis, much like the Turing Test does for intelligence, we allow the study of AI creativity to maintain a degree of objectivity while recognizing the diverse perspectives on what constitutes creativity. This approach ensures that our exploration remains grounded in the real-world application and understanding of creativity, making it more relevant and accessible.
4. Measuring Creativity in Contemporary AI Models
As we navigate through the realm of AI, we encounter the practical aspect of assessing creativity in contemporary models, particularly autoregressive models. The proposed practical measure of statistical creativity is tailored for these models, ensuring its applicability to the latest Large Language Models (LLMs). This measure not only aids in evaluating the creative output of AI models but also serves as a benchmark for future advancements, enabling a consistent and objective assessment of AI creativity.
5. Guidelines for Cultivating AI’s Creative Abilities
Finally, this work underscores the importance of a nuanced approach to training AI for creative endeavors. Contrary to the prevalent focus on amassing large datasets, this research emphasizes the need to collect data that captures the generative conditions of creations. This approach, which advocates for fitting a substantial amount of conditional data without overlooking the generative conditions, is crucial for unlocking AI’s creative potential. It represents a shift towards a more deliberate and informed method of cultivating creativity in AI, setting the stage for the development of truly innovative and creative AI systems.
References
1. “MBA Students vs. ChatGPT: Innovation at the Crossroads” — The Wall Street Journal, Tech Section. Available at:https://www.wsj.com/tech/ai/mba-students-vs-chatgpt-innovation-679edf3b.
2. “Can AI Be as Creative as Humans? Creativity serves as a cornerstone for societal progress and innovation. With the rise of advanced generative AI models.” — ArXiv. Available at: https://arxiv.org/abs/2401.01623.
3. Project Page: “Can AI Be as Creative as Human?”. For more details and social media description, visit: https://ai-relative-creativity.github.io/.