California Management Review
California Management Review is a premier academic management journal published at UC Berkeley
by Anirban Mukherjee and Hannah H. Chang
Image Credit | Alan de la Cruz
Creativity, a defining trait of human intelligence, is a subject of extensive study and ongoing discussion (Amabile 1996)3, (Boden et al. 2004)4. Its manifestations, such as divergent thinking that generates novel ideas and convergent thinking that refines these ideas to meet specific goals, have fueled numerous theories about its essence and underlying processes. As artificial intelligence (AI) technologies continue to advance, there is increasing interest in mirroring or simulating this sophisticated, uniquely human form of creativity within generative AI models. But why is this interest in AI creativity so significant? Is it merely about making AI more human-like, or does embedding creativity have pragmatic benefits for society?
The importance of creativity in AI goes beyond making machines mimic human-like traits. AI models endowed with creative capabilities can unlock a plethora of practical applications, from creating engaging content in the entertainment industry to devising innovative solutions in science, medicine, and business. Creativity in AI can generate unforeseen solutions to complex problems, opening up new opportunities and pathways that might not have been visible or conceivable through human thought alone.
Indeed, early neural networks took cues from the intricate workings of biological systems. Similarly, more advanced AI and machine learning systems can gain from drawing parallels with social systems. These systems, typified by dynamic interactions, collaboration, and innovation, provide a rich source of inspiration for pushing AI technologies forward. Human creativity, in particular—with its nuanced interplay of novelty and usefulness—provides a valuable blueprint for developing AI models capable of both innovating and delivering practical solutions (Simonton 2010)11.
We bring to focus the vital elements of human creativity—novelty and usefulness—in this paper. Novelty refers to generating original content, previously unencountered, while usefulness pertains to crafting content that is relevant, valuable, and practically applicable in a given context. Striking this balance in AI models is a challenging endeavor. Unlike traditional rule-based systems, AI models learn from examples they encounter during their training phase. They do not have access to an explicitly codified set of rules that separate fact from fiction, leading to the generation of creative content that may either deviate too far from practical constraints, known as hallucination, or stay too rigidly within the confines of existing data, a phenomenon known as memorization.
To tackle this challenge, we propose an approach inspired by human creative processes. By incorporating principles like divergent and convergent thinking, domain-specific creativity, evaluation, and explanation, we aim to promote the development of AI models that generate innovative content, minimize inaccuracies, and address ethical considerations.
In the sections that follow, we will explore the conceptual definition of creativity, delve into the challenge of balancing novelty and usefulness in AI, and outline our proposed framework for the development of generative AI models that are creative, useful, and ethical. By drawing connections between human creativity and AI, we aim to stimulate further discussion and research in the broader scientific community on AI creativity.
Creativity is characterized as the capacity to generate novel, yet valuable ideas or solutions to complex, open-ended tasks (Amabile 1983)2. It requires producing outcomes that are original, and at the same time, relevant, feasible, or aligned with a specific goal. Creative tasks can be heuristic, algorithmic, or a mixture of both. Heuristic tasks lack a direct solution and call for innovative approaches, whereas algorithmic tasks have a clear path to a single correct solution. For instance, designing a fresh advertising campaign for a product is a heuristic task since there is no one correct answer or predefined path to success; instead, a range of creative solutions could potentially be effective. On the other hand, solving a mathematical equation is algorithmic due to its pre-defined resolution path.
CRISPR-Cas9’s introduction illustrates creativity within the biotechnology sphere, aligning well with Amabile’s model. The novelty of CRISPR is clear: it repurposed a bacterial defense mechanism into a tool for precise genetic editing, a groundbreaking approach. But novelty alone doesn’t justify CRISPR’s broad success and influence; its usefulness is just as vital. CRISPR-Cas9 has been used in the treatment of genetic diseases such as sickle cell disease and beta thalassemia. It also has potential applications in enhancing crop yield and resistance and furthering genetic research. In simple terms, CRISPR-Cas9 not only introduces a novel method, but it also effectively tackles critical challenges across diverse fields.
Google’s introduction of the PageRank algorithm brought a completely novel approach to web search. Traditional search engines ranked results based on the frequency of a search term appearing on a webpage. Google’s founders, Larry Page and Sergey Brin, proposed a new method, ranking webpages based on their interconnectedness or link structure, essentially viewing a webpage as important if other important pages linked to it.
The usefulness of this novel approach became apparent as it resulted in more accurate search results, making it easier for users to find pertinent information. Google’s novel approach to search transformed the way people interacted with the internet, and it cemented Google’s position as a global tech giant. The introduction of Google’s PageRank algorithm brought a wholly novel approach to web search, significantly enhancing its usefulness.
Generative artificial intelligence (AI) poses a challenge: balancing novelty and usefulness, two crucial factors determining the success and applicability of produced content. To navigate this spectrum more effectively, AI models must carefully apply divergent and convergent thinking and address hurdles associated with implicit learning and the ambiguous information boundary.
A significant challenge that arises in achieving the optimal balance between novelty and usefulness in generative AI is due to the inherent tension between these two aspects of creativity. Unlike traditional rule-based systems, which explicitly encode information, generative AI models engage in implicit learning. This means they analyze vast amounts of data points in their training datasets to capture patterns and relationships, thereby understanding domain constraints and principles (Radford et al. 2019)11.
Implicit learning could be likened to learning a language through immersion, rather than formal education. An AI model, like a person living in a foreign country, picks up on linguistic patterns, rules, and idiosyncrasies through exposure to a wealth of data points (conversations, written text, etc.), rather than being taught the rules of grammar explicitly.
However, this form of learning also creates an ambiguous information boundary, where it’s challenging to distinguish between established facts and potentially new information. This could be compared to our foreign language learner beginning to invent words or phrases that sound plausible but aren’t actually part of the language. This blurred boundary complicates the balance between novelty and usefulness, as the model navigates a vast space of possibilities without explicit guidelines on accuracy, truth, or appropriateness.
For instance, consider a language-based AI model trained on a vast dataset from various internet sources, including scientific articles, novels, news, and forums. When asked to generate an article about a futuristic concept such as “underwater cities,” the AI model might end up generating a highly creative and elaborate narrative, filled with details about the city’s architecture, transportation, and lifeforms.
However, because these details have been ‘imagined’ by the model (which essentially means the details are extrapolations from its training data and not based on real, factual information), some of the content might be scientifically inaccurate or purely speculative. For instance, the AI model might suggest that these underwater cities are built from a non-existent, corrosion-resistant material or inhabited by fictional marine species. This is an example of hallucination, where the AI’s quest for novelty overrides the factual constraints of reality.
Similarly, suppose we have a generative AI model tasked with creating a unique story. However, instead of generating a novel narrative, the model produces a passage that is remarkably similar to a famous opening of a well-known novel. For instance, it might generate a story beginning with “It was the best of times, it was the worst of times,” closely mirroring the opening line of Charles Dickens’s “A Tale of Two Cities.”
This instance demonstrates memorization, where the AI model, in its pursuit of creating useful (or contextually relevant) content, ends up generating outputs that lack originality and essentially echo fragments from its training data. In such cases, the AI model’s creativity is stifled, and the balance between novelty and usefulness is skewed towards the latter.
Balancing novelty and usefulness poses a significant challenge for generative AI models. If these systems veer too far towards novelty, they may override domain facts, principles, or boundaries, thereby creating content that appears creative but significantly deviates from the domain’s foundational logic or history. This deviation could lead to the compromise of its usefulness.
A phenomenon termed hallucination comes into play here. It involves AI responses that contain random inaccuracies or falsehoods expressed with unjustifiable confidence. A notable example of this is large-scale language models like ChatGPT, which might unintentionally generate outputs that blur the lines between reality and imagination, thus stretching the boundaries of innovation but risking the generation of misleading or fabricated information (Brown et al. 2020)5.
Hallucinations can be viewed as an emergent property of the creative process. They represent not a failure of the model, but a misdirection of emphasis. When users require factual information, the concept of appropriateness takes precedence, and the model should adhere to the truth reflected in the training examples rather than creating its own reality. Hallucinations crop up when the model fails to achieve the right balance.
Conversely, when AI systems lean heavily towards usefulness, they may become overly fixated on generating content that strictly adheres to real-world constraints and principles. This focus can lead to a phenomenon known as memorization, where AI models reproduce content verbatim from their training data. As a result, the generated outputs may be useful but lacking in originality, potentially constraining their creative potential.
Fundamental to the process of creativity is the fine equilibrium struck between divergent and convergent thinking (Boden et al. 2004)4. Divergent thinking sets the stage for the generation of an array of ideas or possible solutions, while convergent thinking shines in the curation and selection of the best-fit idea or solution. Ample research on creativity underpins the relative efficacy of these thinking styles, establishing their pertinence to the context and stage of problem-solving—divergent thinking gains prominence during ideation, whereas convergent thinking finds utility in the evaluation and selection phase.
Analogously, the pursuit of equilibrium between novelty and usefulness in generative AI can be guided by principles of creative problem-solving. This involves the identification of problems, the birth of alternative solutions via divergent thinking, and the selection of the most promising solution using convergent thinking. Imbuing generative AI models with these creative problem-solving techniques can engender structured and target-driven methodologies for content generation.
Collaboration, a cornerstone of creativity literature, can bolster both types of thinking in AI development. Techniques such as ensemble learning or multi-agent systems can instill a collaborative spirit, further optimizing the balance between novelty and usefulness in AI-generated content. By fostering cooperation among AI models, a broad spectrum of ideas can be generated, with another specialized set of models focusing on their refinement in terms of usefulness.
Cognitive flexibility, defined as the capacity to alternate between different modes of thinking, is underscored in creativity literature as an indispensable factor in creative problem-solving (Davis 2009)6. In the AI sphere, this might translate to the development of models capable of dynamically modulating their emphasis between novelty and usefulness, contingent upon task requirements or user preferences. Such flexible AI models could deliver a more tailored and effective balance between divergent and convergent thinking, ultimately enhancing the quality of content.
A pragmatic approach to integrate divergent and convergent thinking into generative AI models is a two-step process. Initially, models emulate divergent thinking by promoting diversity and generating a broad set of potential outputs. Subsequently, through convergent thinking, models apply constraints or evaluation metrics to refine and select the most fitting output. By architecturally incorporating these steps into generative AI models, developers can facilitate the desired balance between the generation of novel outputs and the usefulness and relevance of the final content.
Optimally balancing novelty and usefulness in generative AI models necessitates adaptation to unique contexts or domains. Given the varying degrees of novelty and usefulness across domains, AI models can be custom-fit to generate content that is both imaginative and pertinent. The creativity literature provides rich insights into this trade-off, crucial for tailoring AI models.
One strategy involves incorporating domain-specific datasets during the training phase or employing transfer learning techniques (i.e., adapting pre-trained AI models for specific tasks) (Pan and Yang 2010)9. This equips AI models with a robust understanding of domain-specific language, norms, and constraints, enabling them to cater to the diverse requirements of various domains, such as healthcare or art.
A nuanced understanding of the specific needs for novelty and usefulness within a context is pivotal for fine-tuning AI models. This can be achieved by analyzing creativity literature to glean domain-specific heuristics or evaluation criteria that reflect the ideal balance between novelty and usefulness. These criteria, when integrated into the AI model’s objective function or fine-tuning process, may enable general-purpose AI models like ChatGPT to adapt more effectively. This approach mirrors the development of neural networks inspired by biological systems, anchoring AI models in human creativity systems.
Enhancing AI models’ adaptability across contexts can be further achieved by allowing users to specify their output preferences. This can be realized through user interfaces that enable users to tweak parameters related to novelty and usefulness, or by integrating user feedback during the fine-tuning process of the AI model. By giving users a voice to express their preferences, the balance between novelty and usefulness can be tuned to their needs, generating AI content that meets the specific requirements of different domains.
The choice of evaluation metrics to assess generative AI models critically influences the balance between novelty and usefulness. Prioritizing grammatical correctness and fluency metrics may lead AI models to produce coherent, yet unimaginative outputs. Conversely, emphasizing uniqueness might lead to novel but grammatically incorrect or difficult-to-understand outputs.
Researchers can cultivate balanced generative AI models by learning from creativity literature and adopting evaluation techniques that consider both novelty and usefulness. One such technique is the Consensual Assessment Technique (CAT), which involves expert judgments to evaluate the creativity of AI-generated outputs (Amabile 1982)1. CAT offers a comprehensive evaluation by amalgamating various aspects, including fluency, coherence, and originality, into the assessment process (Eastwood and Williams 2018)7.
Moreover, researchers can devise custom evaluation metrics that ensure a balance between novelty and usefulness by amalgamating existing metrics. For instance, they could blend metrics like BLEU, which measures the similarity between the generated text and a reference text, with metrics such as Self-BLEU, which evaluates the diversity of generated outputs by comparing them to each other (Zhang et al. 2018)12. Such a combination nudges AI models to generate content that strikes a balance between coherence and originality. AI explainability and interpretability, crucial for comprehending AI models’ underlying mechanisms and decision-making processes, can provide further support for researchers in adjusting models and comprehending the context-specific trade-offs between novelty and usefulness (Gilpin et al. 2018)8.
We propose a multifaceted framework to construct AI systems that strike an optimal balance between novelty and usefulness. This approach amalgamates domain-specific knowledge, user preferences, and cooperative techniques. Our proposed model includes the following pillars:
Domain-Specific Analysis: To tailor AI models to cater to the unique requirements of different contexts, a comprehensive understanding of the domain-specific characteristics and constraints is required. Through a deep dive into the relevant creativity literature and consultation with domain experts, we can establish the desired equilibrium between novelty and usefulness. This ensures the AI-generated content resonates with the specifics of the domain while maintaining its innovative edge.
Domain-Specific Data and Transfer Learning: The application of domain-specific datasets during the training phase, coupled with transfer learning methodologies, can fine-tune AI models to suit the target domain. This nuanced understanding of the domain-specific language, norms, and constraints results in more relevant and novel content generation.
User Preferences and Customization: Introducing a user interface or feedback mechanism that allows users to express their novelty-usefulness preferences or adjust AI model parameters enhances the personalization of the experience. This facilitates the adaptation of AI-generated content to meet the varying needs and specifications of different domains.
Custom Evaluation Metrics: The development of bespoke evaluation metrics that encapsulate both novelty and usefulness can enhance the effectiveness of AI models. These metrics, inspired by existing ones and human judgement techniques such as the Consensual Assessment Technique, allow the models to generate content that harmonizes originality and applicability.
Collaboration Mechanisms: The implementation of collaborative mechanisms, such as ensemble learning or multi-agent systems, refines the balance between novelty and usefulness in AI-generated content. By encouraging diverse idea generation and refinement based on utility through a collective of AI models, the end content is more likely to fulfill the dual objectives of creativity.
We study challenges faced by generative AI models in striking an optimal balance between novelty and usefulness. Inspired by human creativity, we propose a comprehensive approach that leverages creative problem-solving methods to enhance these models’ capabilities.
The balance between novelty and usefulness isn’t just crucial for AI creativity—it’s also key to addressing ethical concerns. By carefully managing this tradeoff, we can guide AI models to produce creative outputs that not only respect real-world constraints and principles but are also self-aware.
Navigating this balance is vital to aligning generative AI models with user expectations and ensuring they make positive contributions to society. As the field of AI continues to evolve, we hope our research serves as a valuable link between human and AI creativity.
Amabile TM (1982) Social psychology of creativity: A consensual assessment technique. Journal of personality and social psychology 43(5):997.
Amabile TM (1983) The social psychology of creativity: A componential conceptualization. Journal of personality and social psychology 45(2):357.
Amabile TM (1996) Creativity and innovation in organizations (Harvard Business School Boston).
Boden MA et al. (2004) The creative mind: Myths and mechanisms (Psychology Press).
Brown T, Mann B, Ryder N, Subbiah M, Kaplan JD, Dhariwal P, Neelakantan A, et al. (2020) Language models are few-shot learners. Advances in neural information processing systems 33:1877–1901.
Davis MA (2009) Understanding the relationship between mood and creativity: A meta-analysis. Organizational behavior and human decision processes 108(1):25–38.
Eastwood C, Williams CK (2018) A framework for the quantitative evaluation of disentangled representations. International conference on learning representations.
Gilpin LH, Bau D, Yuan BZ, Bajwa A, Specter M, Kagal L (2018) Explaining explanations: An overview of interpretability of machine learning. 2018 IEEE 5th international conference on data science and advanced analytics (DSAA). (IEEE), 80–89.
Pan SJ, Yang Q (2010) A survey on transfer learning. IEEE Transactions on knowledge and data engineering 22(10):1345–1359.
Radford A, Wu J, Child R, Luan D, Amodei D, Sutskever I, et al. (2019) Language models are unsupervised multitask learners. OpenAI blog 1(8):9.
Simonton DK (2010) Creativity in highly eminent individuals. The Cambridge handbook of creativity:174–188.
Zhang BH, Lemoine B, Mitchell M (2018) Mitigating unwanted biases with adversarial learning. Proceedings of the 2018 AAAI/ACM conference on AI, ethics, and society. 335–340.
