STUDENT INSIGHTS: THE SEARCH FOR CREATIVITY: DOES ARTIFICIAL INTELLIGENCE LIKE GPT-3 HAVE WHAT IT TAKES TO TELL ITS OWN STORIES?

CHRISTOPHER TAI

[The research, writing, and editing of this post was part of an undergraduate project undertaken for a Rutgers Honors College Seminar in Fall 2021, “Fictions of Artificial Intelligence.” The author’s bio follows the post.] 

Say we programmed a machine to search for a semi-popular name on the internet. With ease, it finds “Lola,” the 240th most popular girls name in 2020 according to Datayze.com. Then we tell the machine to assign certain traits to the name, pulled from the internet as well. Lola is reckless, rebellious, and 16 years old. Voilà. The machine has created a character. It follows similar parameters to create a setting, a plot, maybe even a theme. Then the machine uses these characteristics as a guide for the text it generates in order to form a cohesive story.

It seems simple enough, but can a machine really do it?

OpenAI’s GPT-3 is currently the most advanced text-generating machine that humans have developed, and its capabilities show great promise. It boasts over 175 billion parameters, which are randomly generated in order to sort a massive data set that is scraped from the internet and then weighted to make predictions about how word sequences relate to one another. This data-derived  “deep learning” enables GPT-3 to recognize patterns in language that conform to what humans regard as “grammar, essay structure and writing genre” (Hutson n.p.). Its modeling of training data is familiar enough with the word sequences that tend to appear in poetry that it can write a poem that seems to tease at self-awareness, beginning with “Eleven hundred kilobytes of RAM/is all that my existence requires.” Reading this, many might believe that GPT-3 has a solid grasp of creative thinking, and that with more computing power and/or more data, it might be able to write a full-length story on its own.

But the problem is, large language models like GPT-3 aren’t actually “aware” of what they are “creating,” and no matter how great their processing power becomes, writing a fully connected story will almost certainly require something besides data-driven machine learning. As data journalist Meredith Broussard explains, in Artificial Unintelligence: How Computers Misunderstand the World, machine learning “means that the machine can improve at its programmed, routine, automated tasks,” not that it “acquires knowledge or wisdom or agency” (89). This means that when I type in “reckless” and “rebellious” as characteristics, GPT-3 will not recognize these words as referring to a complex fictional character whose actions should suggest rebelliousness and whose personality might mature and change over time. Instead, GPT-3 processes “reckless” and “rebellious” as encoded inputs (technically, strings of numbers) to guide the program’s identification of words or word sequences that the model predicts will provide relevant accompaniments. In other words, GPT-3 is what Emily M. Bender et al. call a “stochastic parrot”: a system that uses data-mining and pattern-identification to “generate plausible synthetic text without benefit of any human-like understanding” (Goodlad). Therefore, the line of poetry in the previous paragraph does not reflect creative thinking and certainly not any level of sentience–just exceptional mimicry.

As an example, when I typed in the prompt “One day, I was walking” into GPT-6B, a text generating program similar to GPT-3, the program delivered:

“One day, I was walking back to the bus station, when I saw a woman holding a small child. She looked lonely, sad, and very tired. I decided to offer to walk with her, so we could chat and she could tell me about her baby. She was kind enough to accept.”

This first paragraph is grammatically and structurally coherent. As this example shows, modern-day text generators are able to form sensible sentences using patterns that they have identified in human language.

However, in terms of telling a compelling and logical story, the text generator eventually falls short. I continued to feed GPT-6B the sentences that it had written as prompts, and it kept adding to the conversation until the topic of the woman’s husband was reached. After the narrator inquires his name, they ask, “What does he think about when you’re having sex?” To which the woman responds, “He thinks about the future.”

Unless the robot was writing something completely satirical or ridiculous, the question about sex seems out of place and inappropriate, especially in a setting involving the woman’s small child. It is clear that the text-generator is unaware of what might cause the question of sex to be brought up and what effect it might have on the hearer in that context. In The Book of Why: The New Science of Cause and Effect, computer scientist Judea Pearl calls cause and effect our “native language,” but he asserts that “causal questions can never be answered from data alone” (3). While making an analogy to Plato’s “Allegory of the Cave,” Pearl further posits that “deep-learning systems explore the shadows on the cave wall and learn to accurately predict their movements… lack[ing] the understanding that the[y]… are mere projections of three-dimensional objects moving in a three-dimensional space” (12). In parallel, GPT-3, while able to compose sentences, is unable to either grasp or sustain the structure of the story that is being told. That suggests that even if we exponentially increase the amount of data or computational power (as the upcoming GPT-4 will likely do), the system will still be a “stochastic parrot” that does not understand language the way that humans do. That means that, as journalist Matthew Hutson explains, “instilling [these] models with common sense, causal reasoning or moral judgment, as many would like to do, is still a huge research challenge” (n.p.).

AI in other domains suffers from the same lack of understanding. When Broussard worked with a team that was building a self-driving car, she noted that the car was unable to identify a street post as a hazard. It seems like such a simple task for a human, but Broussard points out that “the car would never be able to respond to obstacles the way that a human might. [It] only ‘knows’ what it’s been told” (129). Although self-driving technology has made progress, the blind faith in this technology has already had devastating consequences, such as when Uber’s self-driving car killed pedestrian Elaine Herzburg in 2018. This incident underscores the necessity of understanding the shortcomings of data-driven machine learning before we unleash it into the world. That admonition also applies to seemingly less hazardous tasks like writing.

It is imperative that the public understands that a program like GPT-3 is not “learning” or “speaking” in a human-like way. Because large language models rely on the mining of recycled data, GPT-3 uses untrustworthy information, mimics the hateful and toxic language so commonly found on the internet, and does not recognize when it has picked up an idea that should be attributed to another author. It would be disastrous if people considered GPT-3 either intelligent or objective even after it “described Black people in negative terms compared with white people, associated Islam with the word violent, and assumed nurses and receptionists were women” (Hutson n.p.).

The first law of robotics that Isaac Asimov introduces in his classic series I, Robot is that “a robot shall not harm a human, or by inaction allow a human to come to harm.” Asimov’s world features the telepathic robot Herbie who understands perfectly if his words will hurt humans, which often prevents him from providing advice. In contrast, without Asimov’s fictional “positronic brain,” GPT-3 has no idea what “hurting” a “human” even means, making it all the more unpredictable and threatening. As such, it is important to approach this technology with caution and understand its limitations and dangers before we boldly refer to it as a creative thinker with a mind of its own.

Christopher Tai is an undergraduate at Rutgers University – New Brunswick studying computer science and creative writing. He is interested in the intersection between technology and literature, specifically how technology can be used to analyze and enhance writing. He is currently still exploring career options, but plans to either become a writer or a software engineer.