Recent research from the University of Washington and Allen Institute for AI has sparked discussion about artificial intelligence’s creative capabilities; or rather, generative AI LLMs’ apparent limitations.

Their paper, “AI as Humanity’s Salieri” — a nod to W. A. Mozart’s technical if less creative contemporary — presents compelling evidence that large language models like ChatGPT are significantly less “creative” than human writers when measured by their tendency to recombine existing content rather than generate wholly original material.

While fascinating, this research presupposes a curious thing: that AI should necessarily be more creative than humans simply because it is technically superior. But must it?

According to the mythology of Star Trek, the answer is a definitive “no.” This is precisely the challenge faced by the show’s beloved AI-powered android Lt. Comm. Data who always “endeavored to become more human,” to quote one of his lines from the 1994 film Star Trek: Generations. Indeed, this was a narrative thread that weaved the character’s entire story arc through through some 35-odd years of storytelling.

In fact, Data’s relationship with creativity wasn’t portrayed as a deficiency but as a fundamental aspect of his artificial nature. His pursuit of artistic expression — futile though it was, from painting to poetry to playing violin and acting, and even dreaming — wasn’t necessarily aimed at matching human creativity, but at least for understanding it.

When AI models are fine-tuned to be more “human-like” through reinforcement learning, their creativity scores actually drop.

Ultimately, such pursuits were concluded in vain, as the character conceded that his “growth as an artificial life form [had] reached an impasse” (Id.) and in order to fully realize his creative capacities, he would need to overcome the very essence of what made him “artificial,” namely, the lack of feelings and emotions, for which, the show’s writers simply hacked in the concept of an “emotion chip,” because of course.

Data continued his somber, almost melancholy delivery that “for 34 years, I have endeavored to become more human, to grow beyond my original programming. Still, I am unable to grasp such a basic concept as humor. This emotion chip may be the only answer.” (Id.)

Viewed through the lens of today’s LLMs, the writers’ solution to further evolve Data’s character seems timely, while also suggesting the futility and senselessness of not only striving to imbue LLMs with human levels of creativity, but curiously, to presuppose that they should possess such creativity in the first instance.

With almost clairevoyant prescience, the showrunners’ approach to Data’s AI creativity often involved his analysis of vast databases of human art, combining techniques in novel ways, and producing works that were neither purely imitative nor wholly original. This mirrors precisely how modern language models approach creative tasks, and that’s not necessarily a problem. (Interestingly, the issue of copyright infringement was never considered on the show.)

The University of Washington research introduces a “Creativity Index” that measures how much of a text can be reconstructed from existing web content. Their finding that AI systems score 66.2% lower than human writers might initially seem disappointing. However, this measurement might better be understood as documenting a type of creative process that differs wholly in type, and not merely in degree.

When AI systems combine and transform existing content, they’re engaging in what we might call “combinatorial creativity.” This process — identifying patterns, making connections, and synthesizing new combinations — is actually a crucial form of innovation. Many human breakthroughs come not from creating something entirely new but from seeing novel connections between existing ideas.

This push — perhaps an instrinsicly human desire, perhaps even an imperative — to make AI systems “more creative” like humans, more in our own image, might, however, be misguided. Just as Data’s character demonstrated that artificial intelligence could be both different from and complementary to human intelligence, technically superior without necessarily being holistically better, modern AI systems might be most valuable when they maintain their distinctive characteristics rather than attempting to mimic human processes; after all, the question remains: why must an AI possess human levels of creative fidelity or artistic virtuoso?

AI creativity should be an accidental byproduct of its actions, and not its purpose.

Interestingly — perhaps ironically, and almost comically — the research shows that when AI models are fine-tuned to be more “human-like” through reinforcement learning, their creativity scores actually drop by 30.1%. This suggests that pushing AI toward human-like behavior might compromise its native strengths rather than enhance them. Incredibly, Star Trek’s writers anticipated as much with Data’s character, too: numerous times his attempts at imitating art or humor were met with catastrophic failure; often hilarious, sometimes dangerous.

This perspective has important implications for how we develop and deploy AI systems. Instead of trying to make AI more “creative” in human terms, we might instead focus on:

1. Leveraging AI’s ability to process and recombine vast amounts of information in novel ways

2. Developing tools that complement rather than replace human creativity

3. Creating interfaces that make AI’s unique creative processes more accessible and useful to humans

Put another way, AI creativity should be an accidental byproduct of its actions, and not its purpose.

Star Trek’s Data never became human-like in his creativity, but he contributed unique perspectives precisely because of his different way of thinking. Similarly, modern AI systems might be most valuable not when they perfectly mimic, let alone exceed, human creativity, but when they offer complementary capabilities that expand the boundaries of what’s possible.

Human creativity, as we understand it, likely stems from our ability to feel emotion. It is the spontaneous expression of our inner lives — manifested in tangible, audible, or visual forms — and fundamentally different from generating novel output based on some corpus of pre-training data. This distinction explains Data’s creative limitations as well: his lack of emotion inherently precluded the kind of spontaneous creativity we associate with humanity.

The University of Washington research helps us understand AI’s creative nature, and crucially, its limitations, which seem to bounce against an asymptotic upper bound, doomed like some Sisyphean climb into inexorable futility.

Like Data, AI systems represent a different kind of mind with different kinds of capabilities. As we continue developing these systems, perhaps we should focus less on making them more human-like and more on understanding and leveraging their unique characteristics.

Rather than dismiss this fascinating research, therefore, it should be applauded for showing us not what AI cannot become, but rather, what it — and we humans — should embrace.