Can AI Be as Creative as Humans?

Haonan Wang1, James Zou2, Michael Mozer3, Anirudh Goyal3, Alex Lamb4,
Linjun Zhang5, Weijie J. Su6, Zhun Deng7, Michael Xie1, Hannah Brown1, Kenji Kawaguchi1
1National University of Singapore, 2Stanford University, 3Google DeepMind, 4Microsoft Research, 5Rutgers University, 6University of Pennsylvania, 7Columbia University
haonan.wang@u.nus.edu, kenji@comp.nus.edu.sg
YouTube arXiv

Abstract

Creativity serves as a cornerstone for societal progress and innovation. With the rise of advanced generative AI models capable of tasks once reserved for human creativity, the study of AI's creative potential becomes imperative for its responsible development and application. In this paper, we prove in theory that AI can be as creative as humans under the condition that it can properly fit the data generated by human creators. Therefore, the debate on AI's creativity is reduced into the question of its ability to fit a sufficient amount of data. To arrive at this conclusion, this paper first addresses the complexities in defining creativity by introducing a new concept called Relative Creativity. Rather than attempting to define creativity universally, we shift the focus to whether AI can match the creative abilities of a hypothetical human. This perspective draws inspiration from the Turing Test, expanding upon it to address the challenges and subjectivities inherent in assessing creativity. The methodological shift leads to a statistically quantifiable assessment of AI's creativity, term Statistical Creativity. This concept, statistically comparing the creative abilities of AI with those of specific human groups, facilitates theoretical exploration of AI's creative potential. Our analysis of the AI training process reveals that by fitting extensive conditional data, including artworks along with their creation conditions and processes, without marginalizing out the generative conditions, AI can emerge as a hypothetical new creator. The creator, though non-existent, possesses the same creative abilities on par with the human creators it was trained on. Building on theoretical findings, we discuss the application in prompt-conditioned autoregressive models, providing a practical means for evaluating creative abilities of generative AI models, such as Large Language Models (LLMs). Additionally, this study provides an actionable training guideline, bridging the theoretical quantification of creativity with practical model training. Through these multifaceted contributions, the paper establishes a framework for understanding, evaluating and fostering creativity in AI models.

Overview

aicreativity

Figure 1: Overview of the Relative Creativity (a) data and Statistical Creativity (b).

AI Creativity Q&A

Question 1: What is the main focus of the paper?

Answer: The paper focuses on establishing a concrete framework for understanding, evaluating, and fostering creativity in AI. It introduces the concepts of Relative Creativity and Statistical Creativity to define AI creativity and theoretically proves that AI can be as creative as humans, provided it can fit a sufficient amount of data generated by human creators. This theoretical analysis, in turn, provides practical guidelines for evaluating and fostering creativity.

Question 2: What is Relative Creativity and Statistical Creativity?

Answer: Relative Creativity is a concept that assesses AI's creativity by comparing its outputs to those of a hypothetical yet realistic human creator. An AI model is considered "relatively creative" if it can produce works that are indistinguishable from that creator, as determined by an evaluator. Statistical Creativity is a methodology that makes assessing Relative Creativity feasible and realistic. It involves comparing creations of AI to those of actual human creators and uses a distribution distance metric to determine if an AI model can replicate the creative abilities of a specific group of humans.

Question 3: What are the key theoretical results of this paper?

Answer: This paper theoretically show that by fitting extensive conditional data, including artworks along with their creation conditions and processes, without marginalizing out the generative conditions, AI can emerge as a hypothetical new creator. The creator, though non-existent, possesses the same creative abilities on par with the human creators it was trained on. Therefore, the debate on whether AI can be as creative as human is reduced into the question of its ability to fit a sufficient amount of data.

Question 4: What are the practical guidelines of the theoretical findings?

Answer: The analysis of statistical creativity in autoregressive models with a prompting setup offers a practical measure for Large Language Models (LLMs). Furthermore, the examination of the training process highlights the significance of collecting data on the generative conditions of creation, as opposed to the widespread method of merely amassing large datasets. It also suggests that fitting a significant amount of conditional data, without disregarding the generative conditions, is crucial for the emergence of creative abilities.

Question 5: Why is the subjectivity of creativity important in the context of AI, and how does the paper address it?

Answer: The subjectivity of creativity is crucial because what is considered creative can vary greatly across cultures, disciplines, and individuals. The paper addresses this by incorporating subjectivity into the assessment of AI creativity from a relative view, where the AI's creative output is compared to that of a specifically chosen human anchor. The different perspectives and standards of creativity are crystallized in the choice of the anchor human. This allows the study of AI creativity to maintain a degree of objectivity.

Question 6: How does this concept of creativity differ from previous works in computer science and cognitive science, and how does it mirror the philosophy of Turing Test?

Answer: This concept of creativity diverges from traditional approaches in computer science and cognitive science by adopting a relative measure, instead of striving to define creativity in an absolute sense. Mirroring the Turing Test, which evaluates intelligence by comparing machine behavior to human responses rather than adhering to fixed definitions, Relative Creativity assesses AI by contrasting its outputs with those of a hypothetical human creator. By employing this comparative method, it effectively bypasses the complexities associated with establishing a universal definition of creativity, concentrating on the more tangible goal of determining whether AI can match human creative abilities in comparable scenarios.

Question 7: In the context of AI creativity, how can different degrees of creativity be reflected and measured?

Answer: Different degrees of creativity in AI are reflected by choosing the target distribution of humans (the specific group of humans) with varying levels of creativity. For example, comparing AI to children versus PhD researchers as the anchor human group would yield different assessment result. AI can range from being derivative, replicating existing data distributions, to mimicking everyday humans, and even replicating behaviors of highly creative individuals. The theory provides one way for controlling the desired levels by selecting the appropriate population of human beings.

BibTeX

@article{wang2024can,
  title={Can AI Be as Creative as Humans?},
  author={Wang, Haonan and Zou, James and Mozer, Michael and Zhang, Linjun and Goyal, Anirudh and Lamb, Alex and Deng, Zhun and Xie, Michael Qizhe and Brown, Hannah and Kawaguchi, Kenji},
  journal={arXiv preprint arXiv:2401.01623},
  year={2024}
}