The Synergy of Human Creativity and AI

It's natural to question the role of human creativity in the face of AI advancements. However, humans possess unique qualities that set them apart from AI, particularly in terms of flexible thinking and the ability to generate diverse and tangible responses.

Humans are more flexible thinkers than AI, even though AI excels in abstract, intangible concepts. Humans generate fewer but responses are more diverse and tangible, grounded in the real world. This contrast is fascinating and deserves further study: humans think in observable, concrete terms, while AI often operates in abstract, conceptual ways using complex language. I find it compelling to view "flexible thinking" as a human trait, showcasing our creativity in ways machines can't quite replicate.

This also suggests refined possibilities for collaboration: we can leverage AI to explore the intangible and ungraspable realms of creativity, and then translate those ideas into practical, real-world applications.

Human creativity can get stuck, and AI tools are invaluable for getting us unstuck. Human originality increases over time as our semantic networks expand, meaning creativity takes a while to warm up. Sometimes we attach too early to an idea (anchoring bias) or feel too stressed to think broadly. We may fixate on more obvious responses at first, and AI doesn't. AI doesn't need this "ramp up time" and can spark new thought paths instantly, making it a natural partner for overcoming creative blocks.

One application of this is in architecture. Architecture is a highly creative field that has always been an interesting case study for technology applications. It requires not only high creativity but also precision, math, and engineering—tasks that don't always feel creative. So, how can an architect master both divergence in the design space and convergence in practical manifestation? Midjourney, as an ideation partner, is one way.

One architect found that Midjourney was true enough to the intuitive physics to inspire design. By generating realistic combinatorial concepts, it got them out of their creative rut and enabled them to imagine different designs faster, enhancing their overall creativity. It seems that time matters in human divergent thinking, where tasks tend to follow a diminishing returns curve before reaching a plateau for more original responses. AI helps us reach that plateau faster.

Another fascinating application is in analogical reasoning. Analogical reasoning, the process of identifying similarities between distinct concepts, is a powerful tool in discovery and innovation. It enables us to bridge disparate fields and transfer insights to uncover new possibilities. Its power lies in revealing hidden connections, facilitating breakthroughs that drive scientific advancement and technological innovation.

A recent MIT paper highlights the sophistication of LLMs in analyzing knowledge and creating new research pathways. It shows how an LLM can help construct knowledge graphs from diverse subjects and identify deep structural similarities between domains. For example, the paper draws analogies between Beethoven's 9th Symphony and bioinspired materials. Using generative AI like Claude Sonnet and GPT-4, researchers converted a large dataset of scientific papers into an ontological knowledge graph. They then performed structural analysis to uncover relationships and common structures.

Researchers identified critical nodes within the graph using measures like betweenness centrality and degree centrality. These nodes act as key bridges in the network, facilitating the discovery of new knowledge and maintaining overall connectivity. In Beethoven's graph, nodes with high betweenness centrality could connect different musical ideas or sections. Similarly, in the bioinspired materials graph, these nodes could link various functional parts of a material. Understanding these bridging nodes helps draw analogies between how music transitions and how material properties interact.

Today, there's a common bias that people want AI to handle mundane tasks or speed up their work for greater "productivity." While this is understandable, it shows a failure of imagination. What people actually want from technology often needs to be discovered, and new designs built to support those discoveries.

Take, for example, an experiment from Google where rapper Lupe Fiasco teamed up with a creative technologist to use a large language model (LLM) in rap. The technologist assumed the rapper would want the LLM to write lyrics. "Turns out this isn't what he wanted at all," said the techie. What Lupe wanted was a way to explore the "universe of possibilities" for any given word. Not a series of lyrics, not a song, not even a verse—just one single word. And an LLM can do that in a way no human ever could. He even broke it down to a single syllable, moving from "expressway" to "way" to "whey" to "Wei" to "weigh." The team then built an app that can take any word and "explode" it into whatever the LLM can find that's even remotely related. That's the kind of wild, turbo-charged version of the Divergent Associations Test we should demand! Watch the video and see how the space of possibilities for creativity, design, and new applications could be truly vast.

Finally, one of the recent papers regarding humans being outdone by AI on divergent creativity suggests an even bigger role for LLMs in human creativity: mapping our minds. Rather than merely comparing the creative outputs of humans and large language models (LLMs), the study delves deeper, posing intriguing questions: Can LLMs build individualized models of creative thinking? Can these models help us understand the human creative process on a profound level?

This human-machine synergy offers a holistic view of creativity, blending the subjective experience of a creative "aha" moment with the objective precision of computational models. Future research could transform this blend into a seamless fusion, where generative models serve as bridges—'generative passages'—connecting our personal, first-person creative experiences with the third-person, data-driven world of neural processes. Picture this as a two-way street where our lived experiences inform AI's algorithms, and in turn, AI offers us new perspectives on our creativity.

Of course, AI innovation doesn't exist in a vacuum; domain knowledge is essential to sort the wheat from the chaff. However, the sheer variety and weirdness of some creations that multi-modal models can generate is exciting. This weirdness is a resource we can tap into in a "keep humans weird" kind of way.

AI's tendency to "hallucinate" or generate unexpected responses is often seen as a drawback, especially in tasks requiring precise, factual answers. However, in divergent thinking, this trait can be an asset. Divergent thinking values the generation of numerous, varied ideas, and hallucination fits well within this framework.

Current models, prone to hallucination, may excel in benchmarks based on divergent thinking. In these tasks, hallucinations can produce valuable and imaginative responses, such as creative uses for common objects. Conversely, in convergent thinking tasks, where a single correct answer is needed, hallucinations can lead to wrong outcomes and undermine reliability.

Real-world creativity remains the domain of humans for now because creativity is fundamentally linked with human agency and our expressive selves. Perhaps AI's biggest impact here will be in the realm of search. Searching for information and exploring design spaces for inspiration is now facilitated by generative technologies across multiple modes. Whether you're looking for a new structure, concept, or code, generative AI can be a powerful extension of our minds, making search more efficient and expansive.

It's possible to imagine a bifurcation in search: creative exploration versus traditional retrieval. In many ways this would mark a natural evolution in how we handle information. Embracing both modes and leveraging their strengths unlocks new potentials in knowledge discovery, creativity, and innovation. This dual approach not only boosts our search capabilities but might also enrich our ability to solve complex problems and drive progress across various fields.