Beyond the AI paradox: Are Machines Capable of Genuine Creativity?

Andrei Mihai

There is a paradox at the heart of artificial intelligence. Machines excel at things that are nearly impossible for us – analyzing vast datasets, predicting complex patterns, and playing games like chess or Go at extremely high levels. Meanwhile, they struggle with tasks that are common sense even for children or toddlers, like helping someone in need, understanding a conversation, or even just interacting with the environment.

But what does this mean for what is arguably the most human trait of all – creativity? Can a machine that struggles to understand even the most basic of human processes truly create art or write a novel? Can it develop innovative gameplay? Or is it just a highly efficient pattern-matcher, mimicking creativity without truly understanding it?

These questions came into the global spotlight in 2016 when Google DeepMind’s AlphaGo stunned the world by defeating Go grandmaster Lee Sedol with a move that seemed almost … creative.

Move 37

At the 11th Heidelberg Laureate Forum (HLF) in 2024, this moment was also highlighted. A panel at the HLF looked at the paradox of AI, with moderator Bennie Mols kicking off the discussion with two videos that showcased, on one hand, AI’s inability to manage simple tasks, and, on the other, its ability to surprise us with new concepts. In particular, Mols focused on the famous “move 37.”

AlphaGo’s victory over one of the world’s best Go players in history, Lee Sedol, was stunning in itself. Go is one of the most complex games ever created, with more possible board positions than there are atoms in the observable universe. Mastering it has always been considered an achievement beyond the reach of computers because it is so complex that it requires a certain level of intuition and “feel” for the game.

When AlphaGo was pitted against Lee Sedol in a 5-game series, few people (if any) expected it to win.

However, it was not just the victory that was so striking, but the way in which the AI played. It brought several surprises, including one particular move, the 37th move in game two. It was a move that no human Go player would have made, a concept that was never before explored by a human player.

AI had previously shown glimpses of intuition and creativity. In chess, for instance, AlphaZero often prefers pushing its lateral pawns, even without calculation to show tangible rewards. Yet, this Go move was something else. It stunned both commentators and Sedol. The game’s organizers had to double-check that everything was in order with the system. Professional players initially mocked it but then noted that it was a “creative” and “unique” idea. The more they analyzed the move, the more they realized how brilliant it was.

Was this creativity? Or was it simply the result of vast computational power running through an enormous database of possible moves? You could argue that creativity in humans often arises from pushing beyond conventional thinking, which AlphaGo did. But of course, you can also argue that it was simply an unexpected move generated by an algorithm designed to maximize its chances of winning.

Stochastic Parroting

Although Go is a very popular game in some countries, the world at large was not truly faced with the possibility of AI being creative until ChatGPT came along.

Suddenly, anyone with internet access could access this tool that not only generates human-sounding text and reacts to what you are typing but can even apparently produce things like poems or literature. The key word here is apparently.

One of the central arguments against AI creativity is the concept of stochastic parroting. Coined by Emily M. Bender and her colleagues in a seminal 2021 paper on large language models, the term refers to how AI systems like GPT-3 can appear to generate creative or intelligent responses, but in reality, they are just statistically echoing patterns from the data they were trained on. These models generate text by predicting the next word based on probabilities derived from millions of previous sentences, making them expert pattern matchers, but not originators of truly new ideas.

For example, when you ask a large language model to write a poem or a story, it can produce something that seems fresh and original. But it is simply remixing pieces of literature it has encountered in its training. Much like a parrot might repeat human speech without understanding the meaning, AI models can “parrot” creative forms without grasping the underlying intent.

Simply put, AI can shuffle the data in new ways but it does not experience the same processes that we do, facing a limitation that keeps it tethered to imitation rather than innovation.

Something similar appeared to happen with algorithms like DeepArt or DALL-E, which generate impressive visual art based on user prompts. These systems seem to generate new ideas (you can easily create images of a cat astronaut, for instance), but are merely rearranging elements existing from photos they were trained on.

Interpolation vs Innovation

Alexei Efros, one of the pioneers in computer vision who laid the foundation of image-generating AIs, discussed whether AI can truly create art or whether it is more interpolating existing data in different ways. Efros received the 2016 ACM Prize in Computing and held a press conference at the 11th HLF, 2024, expressing his views on whether AI is innovating or just interpolating data in different ways.

“I lean more on interpolation. I think the reason why there are some novel moves in chess and other games is because it is fundamentally a very closed system and a simple system. There are a lot of possibilities, but the rules are very well-defined and very simple. When you’re searching this big space, maybe the computer can find something that humans haven’t found yet, but I would be surprised if we will have something like this for literature or art or music, because there, the search space is, in fact, infinite.”

Efros went on to explain that since society even lacks a precise definition of what exactly art is, algorithms cannot truly create art. This is in stark contrast to games like chess and Go, where we have clear definitions of what the moves are and what the objective is.

“So it’s an open-ended search space, as opposed to chess, where [there are] a fixed number of squares with a fixed number of pieces […] it’s a closed universe. It can be practically infinite, but technically, it’s closed. Art is not closed. We don’t even know what’s going to be art in 100 years. So I think we are not going to see, you know, a computer artist, anywhere, anytime soon.”

Can an AI Be Truly Creative?

One possible avenue for future AI creativity lies in unsupervised learning, where there are no predefined outcomes and machines instead explore new possibilities on their own. In theory, this could allow AI to develop entirely novel ideas, free from human-imposed constraints.

However, even in unsupervised learning, machines are not capable of experiencing curiosity, frustration, or inspiration – all of which are integral to the human creative process. At the heart of this discussion is the concept of intent. Human creativity is often linked with a need to communicate, express, or address problems. AI, on the other hand, lacks intention. It does not create with purpose or meaning. Even when AI-generated works are impressive or beautiful, they lack the emotional and cognitive context that drives human creativity.

How crucial these elements are to the process of creativity is very much an open question. We cannot even strive to solve this problem without truly defining what human creativity is. AI is already passing some human creativity tests, but that just goes to show that our tests are not perfect and creativity can be mimicked. Creativity can be regarded as a social construct, and how exactly that construct is defined may end up deciding whether AI is creative or not.

A quick browse through published scientific literature finds conflicting results. It is not difficult to find some papers claim that AI can be creative. “We prove in theory that AI can be as creative as humans under the condition that it can properly fit the data generated by human creators. Therefore, the debate on AI’s creativity is reduced into the question of its ability to fit a sufficient amount of data,” one 2024 study reads. Meanwhile, another paper from 2023 concludes that “Artificial creativity may be original and effective but it lacks several things that characterize human creativity.”

Whether or not AI can replicate the processes behind human creativity, it seems able to exhibit some pseudocreativity, or an ability to mimic creative outputs (even though the processes behind these outputs are not necessarily linked to creativity).

AI as a creativity partner?

While this debate will likely continue for some time, a more pragmatic approach might be to regard AI as a partner to augment human creativity rather than a competitor.

AI is first and foremost a tool, and rather than replacing human creators, it can enhance our own ability to innovate. Several studies and projects have highlighted the ability of AI to augment human creativity rather than replace it. For instance, writers can use language models to overcome writer’s block or find alternative ways to continue a plot. Musicians employ AI to compose melodies or experiment with chord progressions that they might not have conceived on their own. Visual artists utilize generative algorithms to create complex patterns and designs, blending computational precision with human aesthetic judgment.

Yet, here too some thorny questions emerge. For instance, one recent study found that AI can boost writers’ creativity, but tends to produce more similar stories. This could mean that in the long term, AI actually has a detrimental effect on creativity, but this has not been sufficiently explored yet, as generative text models are still relatively new.

There are also valid concerns that this could lead to a future where human artists, writers, and innovators are displaced by algorithms. We are already seeing this happen around us. AI-driven platforms can generate high-quality images and artworks based on simple text prompts that can be used instead of human work. AI writing tools are being used to produce articles, marketing copy, and even entire books, reducing the need for human writers. Additionally, AI-generated music and deepfake technology are impacting musicians and voice actors, as algorithms can now compose original scores or mimic human voices with remarkable accuracy.

This shift is not just sparking a philosophical debate about how we value creativity, it is affecting the creative job market.

Ultimately, whether or not AI can ever be creative will likely remain an open question for some time. While the machines may not “feel” their creations, they are undoubtedly changing the way we define and experience creativity, as well as the creative market.

The future of creativity may hinge not on whether AI can replicate human ingenuity, but on how we choose to accept and integrate working with these algorithms.

The post Beyond the AI paradox: Are Machines Capable of Genuine Creativity? originally appeared on the HLFF SciLogs blog.