As AI-generated art and stories impress, researchers highlight critical gaps in creativity, urging a shift from output-focused measures to understanding the cognitive processes behind true innovation.
Research: Creativity in AI: Progresses and Challenges. Image Credit: Collagery / Shutterstock
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.
In an article submitted to the arXiv preprint* server, researchers in Switzerland surveyed the creative capabilities of artificial intelligence (AI) systems, focusing on creative problem-solving, artistic outputs, and linguistic innovation.
"Computers can’t create anything. For creation requires, minimally, originating something. But computers originate nothing; they merely do that which we order them, via programs, to do." Ada Lovelace
While advanced AI models could generate linguistically and artistically creative content, they still struggled with abstract thinking, logical coherence, and exhibited issues like lack of diversity and hallucinations.
The authors emphasized the importance of addressing copyright and authorship concerns and called for a comprehensive, multi-dimensional, process-driven evaluation of creativity. This evaluation would consider both the outputs and the underlying mechanisms that contribute to creative processes.
Inspired by insights from cognitive science and psychology, they proposed future research directions to enhance AI creativity, suggesting the integration of theories from combinatorial, exploratory, and transformational creativity.
Related Work
Past work has explored the concept of creativity as a hallmark of human intelligence, raising questions about machines' ability to think and create.
While advancements in transformer-based models have demonstrated impressive generative capabilities across various domains, issues such as lack of diversity, originality, and real-world reasoning persist.
Additionally, these models struggle with tasks requiring more complex forms of creativity, such as abstract thinking and compositionality, often producing incoherent outputs over longer narratives.
Exploring Creativity in AI-Language Models
Linguistic creativity has been debated regarding its alignment with everyday definitions of creativity. While Chomsky's theory emphasizes the fixed rules of grammar, others, like Sampson, differentiate between fixed (F-creativity) and extending (E-creativity) forms of creativity. Recent studies have focused on the E-creativity of AI systems, including humor, figurative language, and linguistic innovation.
Although large language models exhibit capabilities in generating humor and metaphorical expressions, challenges remain, such as producing clichéd or incoherent content and struggling with novel word formation.
Overall, the linguistic creativity of AI models is often questioned, as it may heavily rely on existing human-generated text rather than showcasing truly original innovation.
A summary of domains, dimensions, types and processes of creativity covered in this survey.
Innovative Cognitive Integration
Creative problem-solving involves innovative solutions to problems and requires creativity and commonsense reasoning.
It encompasses convergent thinking, which aims for a single optimal solution, and divergent thinking, which explores multiple ideas.
While language models perform well on divergent tasks, they struggle with tasks needing lateral thinking and deep abstraction.
Despite some success in analogy-making, models often require more robustness and generality of human reasoning in complex situations.
The challenge lies in developing AI systems that effectively integrate these cognitive abilities, allowing them to move beyond mere pattern recognition to genuine innovation.
Advancements in Artistic Creativity
Artistic creativity involves producing original works, including storytelling, poetry, visual arts, and music.
AI advancements have enabled automatic creative content generation, but challenges persist, such as ensuring long-term coherence and capturing the depth of human expression.
While machine learning (ML) has improved narrative generation, factual inconsistency and lack of creativity remain in storytelling.
Poetry generation has evolved from template-based approaches to neural networks, yet AI-generated poetry often lacks emotional depth and thematic consistency.
AI also benefits visual and musical creativity, but challenges like compositional errors and subjective quality evaluations continue to exist.
Automating Scientific Creativity
Scientific creativity entails generating innovative ideas and solutions within science, leading to discoveries and technologies.
Since the 1970s, AI research has focused on automating scientific discovery, with early efforts centered on automated equation discovery and symbolic regression.
Recent advancements utilize Bayesian statistics and neural networks to enhance various scientific processes, including idea generation and hypothesis formulation.
Notable developments like the AlphaFold model have revolutionized biological research by accurately predicting complex 3D protein structures. Despite these advancements, challenges persist regarding the reliability of large language models (LLMs), including issues like hallucinations and poor planning.
The rise of generative deep learning techniques prompts critical questions about copyright and authorship in machine-generated works.
In the United States, copyright law permits the reproduction of protected works under the fair use doctrine, but applying this to generative models can be complex.
In the European Union, legally accessible data for training is allowed if specific conditions are met, though verifying these conditions is often difficult.
Additionally, the authorship of AI-generated content remains contentious; if human involvement is minimal, determining ownership complicates matters.
Various perspectives exist on whether the developer, user, or AI should hold authorship, underscoring the need for clearer legal frameworks surrounding intellectual property rights.
Creative Process Evaluation
Cognitive scientists have developed frameworks to evaluate creativity, focusing on input, process, and output, strongly emphasizing the creative process.
Although AI research often analyzes creativity from an output perspective, understanding the mechanisms behind creativity is equally crucial.
The creative process in computational creativity can be categorized into combinatorial, exploratory, and transformational types, with theories from Boden and Wallas outlining how ideas are developed.
The research highlighted that current AI systems fall short in mimicking human-like creative processes, as they lack the spontaneity and intentionality associated with genuine creativity. Enhancing AI creativity involves architectural innovations, prompt engineering, and decoding strategies that better mimic human-like creativity.
Conclusion
To sum up, the rapid advancements in AI, mainly through models like large language and diffusion models, showcased impressive creative capabilities, yet the question of genuine machine creativity remained unresolved.
This survey explored linguistic, creative problem-solving, artistic, and scientific creativity while addressing copyright and authorship issues in generative artworks.
It highlighted major challenges facing current AI systems and proposed potential research directions for evaluating and enhancing their creativity.
The authors suggested that adopting a process-based evaluation could help future research assess whether machines could achieve a human-like creative process, enriching the understanding of AI and its capabilities.
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.
Journal reference:
- Preliminary scientific report.
Ismayilzada, M., et al. (2024). Creativity in AI: Progresses and Challenges. ArXiv. DOI:10.48550/arXiv.2410.17218, https://arxiv.org/abs/2410.17218