Good Readers and Good Writers and AI

Good Readers and Good Writers and AI

Vladimir Nabokov famously stated his position on the purpose of art in the afterword to Lolita: “For me a work of fiction exists only insofar as it affords me what I shall bluntly call aesthetic bliss, that is a sense of being somehow, somewhere, connected with other states of being where art (curiosity, tenderness, kindness, ecstasy) is the norm.”^[1] Here, he identifies ‘tenderness’ as a crucial component of art, and it is tenderness that comes to define his understanding of art both in its frequency in his texts (the word itself appears on his pages more often than any of the others) and in the space he gives to the all-consuming impact of love. For Nabokov, tenderness is love at its most generous, and writing provided him with a model for how to grieve its loss. Fiction, when attuned to tenderness and kindness, as much as curiosity or ecstasy, could transform loss: memory and lost time resurrected through artifice and ‘the real’ could be shared—with the loved one, with the reader—and it is in this sharing that consciousness attains a measure of immortality.

Nabokov believed that good writers (and readers) must see the world as the potentiality of fiction.^[2] And he conceived of an interpersonal relationship between writers and readers that spoke to something interpersonal in the act of writing itself. Describing his awe of sunsets in his boyhood, Nabokov wrote that “I did not know then (as I know perfectly well now) what to do with such things—how to get rid of them, how to transform them into something that can be turned over to the reader in printed characters to have him cope with the blessed shiver—and this inability enhanced my oppression.”^[3] Like these ephemeral sunsets, Nabokov saw that writing could be a dissolution of the self, a way to transform individual memory into collective experience.

Perhaps there is no art form as collective as writing. Writing is inherently a social act. The writer interacts with the world and those who populate it through the potentiality of fiction.

But where does AI-generated narrative fit within a paradigm that exalts tenderness and intentionality as necessary precursors to creativity and artistic production? And how does the possibility of wholly autonomous, AI-generated creativity in the form of a poem, short story, or novel fit within a legal framework predicated on “human authorship?”^[4]

What is Natural Language Processing? 

Natural language processing (“NLP”) is the branch of artificial intelligence (“AI”) concerned with giving computers the ability to understand text and spoken words.^[5] NLP combines computational linguistics-rule-based modeling of human language-with statistical, machine learning, and deep learning models, which enable computers to process human language.^[6] This includes the ability to process language in the form of text or voice data and to understand its semantics, including the writer’s intent and sentiment.^[7]

The Rise of GPT-3

Advanced language models like GPT-3 raise compelling questions about the limits of the current copyright regime to address AI-generated creativity where attribution and ownership cannot be readily identified in a human source. Created by OpenAI, a research laboratory co-founded by Elon Musk, GPT-3 is the most recent and currently the most advanced of these language generation models.^[8] GPT-3 (“Generative Pre-trained Transformer”) is a third-generation, autoregressive language model that uses deep-learning to produce uncannily human-like text.^[9] From a source input (called the prompt), GPT-3 generates sequences of words, code, or other data.^[10] The latest iteration of this language model, GPT-3 uses 175 billion parameters (i.e., the values the neural network tries to optimize during training).^[11]

Unlike previous models that might have required fine-tuning through additional supervised learning training them on the exact task of interest, GPT-3 can perform novel tasks with just a prompt. Given a human language description or a few examples of the desired task, GPT-3 can perform many tasks for which they were never trained.^[12]

Capable of writing automatically and autonomously to produce texts on demand, GPT-3 is remarkably easy to use. Analogous to how a search engine like Google operates, GPT-3 writes a text continuing the sequence of words fed to it (the prompt) without any understanding. It does so for the length of the text specified, regardless of whether the task is simple or difficult, meaningful or meaningless.^[13]

Provided with the starting text and without supervision, input, or training concerning the “right” or “correct” text that should follow the prompt, GPT-3 produces the text that is a statistically good fit.^[14]

In 2017, Ross Goodwin published an experimental, AI-generated novel called 1 the Road.^[15] The Guardian famously released an article written by AI in 2020 that caused a public stir from its human readership.^[16] That same year, the New York Times had GPT-3 write a series of pieces for Modern Love, the paper’s column about relationships and feeling.^[17] Indie novelists have also taken to using AI programs to help them churn out content faster.^[18]

Much of the conversation around AI and creative writing has centered on AI as writing assistants (tools that might aid writers without supplanting them), collaborators that could push writers in new and unexpected directions, or an impending death knell for human writers who might see their roles replaced by AI once it advances to a point where it can write novels and other creative works that are indistinguishable from that of human authors.^[19]

CoAuthor is the latest experiment in Human-AI Collaborative writing and provides some insight into how AI is shaping the process of creative writing. Capturing detailed interactions between 63 writers and four instances of GPT-3 across 1,445 writing sessions in English, CoAuthor is an interface that records writing sessions at a keystroke level, curating a large interaction dataset. In their experiment, the researchers engaged over 60 people to write more than 1,400 stories and essays, each assisted by CoAuthor. As the writer begins to type, they have the option to press the “tab” key. The system then presents five suggestions generated by GPT-3, and the writer can accept or reject the suggestions based on their own sensibilities, modify them, or disregard them entirely.^[20]

CoAuthor tracks all of these interactions between the writers and the model, including text insertion/deletion, cursor movement, and suggestion selection. From this data, researchers can analyze when a writer requests suggestions, how often they accept suggestions, which suggestions get accepted, how they were edited, and how they influenced the subsequent writing.^[21]

CoAuthor can also determine how helpful the accepted suggestions are to the human writer as well as interpret rejected suggestions to improve its suggestions for future language models.^[22]

At each session, the writers took surveys about their overall satisfaction with the collaboration and their own sense of productivity and ownership in the resulting work. While suggestions were sometimes disregarded because they took the writer in a different direction or were too repetitive or vague, the commentary also suggests that the ideas proposed by CoAuthor were often new and useful. CoAuthor’s creators found that the use of large language models increased writer productivity as measured by the number of words produced and the amount of time spent writing. While acknowledging technical hurdles (large language models are prone to generating biased and toxic language), CoAuthor suggests that the best collaborations between human writers and models occur when humans use their own creativity to evaluate suggestions and AI offers suggestions that can guide the writing process along.^[23]

While the possibility of wholly autonomous, AI-generated creativity in the form of a novel is still nascent, advancements of AI in producing human-like text raises important questions of protectability under the present copyright legal framework. Legal personality of machines is currently unavailable under the present legal framework. The US legal system also seems unlikely to recognize mechanical authors. While the US Copyright Act does not have an express statutory definition of authorship, textual references and case law exclude the possibility of construing non-human agents as authors under the statute: “In particular, Section 101 of the Copyright Act defines anonymous works as ‘ones where no natural person is identified as an author,’ thus pointing at natural persons as potential authors. Further, there is a long-lasting understanding that the constitutional history of the word ‘copyright’ would dispose in favour of only humans as ‘authors.’”^[24] Moreover, originality as a condition for copyright protection also prevents protection of AI-generated creativity. Originality is widely defined in most jurisdictions under personality theory as a representation of the personality of the author.^[25] Defining originality in this way suggests that a machine would be incapable of producing something truly original.

Ultimately, the advancements of language-learning models like GPT-3 suggest that AI may one day be capable of producing texts indistinguishable from that of human authors. Whether it will ever be able to write with the mastery or the precision of a Nabokov remains to be seen, given that AI has yet to capture the human sentiment that so often distinguishes great writing. But the developments already underway suggest that the current copyright regime may need to be retooled to recognize AI-generated creativity––whether as assistant, collaborator, or author.