Creativity in AI: Progresses and Challenges (2410.17218v4)

Abstract: Creativity is the ability to produce novel, useful, and surprising ideas, and has been widely studied as a crucial aspect of human cognition. Machine creativity on the other hand has been a long-standing challenge. With the rise of advanced generative AI, there has been renewed interest and debate regarding AI's creative capabilities. Therefore, it is imperative to revisit the state of creativity in AI and identify key progresses and remaining challenges. In this work, we survey leading works studying the creative capabilities of AI systems, focusing on creative problem-solving, linguistic, artistic, and scientific creativity. Our review suggests that while the latest AI models are largely capable of producing linguistically and artistically creative outputs such as poems, images, and musical pieces, they struggle with tasks that require creative problem-solving, abstract thinking and compositionality and their generations suffer from a lack of diversity, originality, long-range incoherence and hallucinations. We also discuss key questions concerning copyright and authorship issues with generative models. Furthermore, we highlight the need for a comprehensive evaluation of creativity that is process-driven and considers several dimensions of creativity. Finally, we propose future research directions to improve the creativity of AI outputs, drawing inspiration from cognitive science and psychology.

Creativity in AI: Progresses and Challenges

The paper "Creativity in AI: Progresses and Challenges," authored by researchers from EPFL and other Swiss institutions, presents a comprehensive review of the state of creativity in AI. It identifies the current capabilities and limitations of AI systems in emulating various aspects of human creativity, categorizing these into linguistic, artistic, scientific creativity, and creative problem-solving. The analysis is highly relevant for researchers focused on advancing AI to achieve more human-like creative capabilities.

Overview of AI Creativity

The authors highlight that while cutting-edge AI models, especially those based on Transformers, have achieved significant progress in generating creative outputs like poems, visual arts, and music, they still fall short in areas like creative problem-solving, abstract reasoning, and maintaining long-range coherence. The outputs often suffer from lack of diversity and originality, which raises questions about the true nature of AI creativity versus mere interpolation and memorization.

Key Findings

• Linguistic Creativity: AI models are adept at generating humor, figurative language, and lexical innovations. However, humorous content often lacks diversity and coherence, revealing the limitations of current LLMs in fully understanding and generating high-quality humor.
• Creative Problem-Solving: AI struggles with tasks that involve convergent and divergent thinking, essential for creative problem-solving. The models demonstrate cognitive biases like functional fixedness, limiting their ability to think innovatively or adapt to novel scenarios.
• Artistic Creativity: AI has advanced significantly in generating high-quality visual and musical content through models like Diffusion and GANs. Nonetheless, these models exhibit errors in compositionality and commonsense reasoning, indicating that they still lack a deep understanding necessary for complex artistic expression.
• Scientific Creativity: The potential for AI in scientific creativity is illustrated through models like AlphaFold. However, fully automating the scientific process remains challenging, with existing methods often lacking novelty and comprising repetitive or poorly motivated outputs.

Implications and Future Directions

This paper's insights have implications for both theoretical and practical advancements in AI. Notably, it underscores the importance of evaluating AI creativity not just by the output but also by the process. This involves understanding how AI "creates" and whether it can achieve spontaneity and intentionality akin to human processes.

The research identifies several promising directions for future work:

1. Evaluating Creative Processes: Analyzing AI systems' creative processes, beyond mere outputs, could lead to deeper insights into their potential to replicate or assist human-like creativity.
2. Developing Creative Architectures: Innovative AI architectures that actively diverge from training data and incorporate spontaneity and agency could enhance the creative outputs of AI systems.
3. Novel Evaluation Metrics: A multidimensional evaluation of creativity is necessary, incorporating dimensions such as novelty, usefulness, surprise, agency, and spontaneity.

Overall, this work aims to equip AI researchers with a clearer understanding of current capabilities and challenges, encouraging exploration into more sophisticated models that approach human-like creativity. By presenting a lucid and comprehensive state-of-the-art summary, the authors set a foundation for the next wave of research dedicated to bridging the gap between AI's current outputs and authentically creative processes.