Uncanny Patterns in AI-Generated Prose
In the subtle resonance of our digital age, a fresh literary voice is emerging. This signature style is ubiquitous — appearing in best-selling novels, local newspaper columns, and even the text on takeout menus. And yet the author is not a human being, but a ghost — a whisper woven from the algorithm, a construct of code. AI-generated writing, once a distant echo of science-fiction daydreams, is now ubiquitous — neatly packaged, fleetingly appreciated, and endlessly recycled. It’s not merely a flood — it’s a groundswell. Yet there is an unsettling quality to this voice. Every sentence resonates, indeed, but to be frank? It resonates with a slight flatness. It fails to unravel the intricate tapestry of human experience — it resembles the work of someone isolated, equipped only with Wi-Fi and a thesaurus. Not sensory, not real, just … present. As AI writing becomes increasingly common, it raises a critical question — what implications does this have for creativity, authenticity, or the essence of being human when a significant number of individuals choose to explore the strange prose generated by machines?
Commonplace words, such as “tapestry,” which has been innocently referring to a type of vertical carpet for over 500 years, unexpectedly induce a sense of tension in me. I find myself infuriated by any sentence that follows the structure “It’s not X, it’s Y,” despite this common construction appearing in widely respected works such as the Bible and Shakespeare. However, regardless of the original meanings of these subtle nuances in language, their significance has evolved over time. All of these are now unmistakable indicators that what you’re reading was produced by an AI. In the past, numerous writers emerged, each showcasing a variety of styles. Now, increasingly, one uncredited author produces essentially everything. It is commonly thought that it is responsible for crafting nearly every undergraduate student essay across universities globally, and there is little reason to assume that more esteemed forms of writing are exempt from this influence. In a survey conducted last year by Britain’s Society of Authors, it was revealed that 20 percent of fiction writers and 25 percent of nonfiction writers were permitting generative AI to assist with aspects of their work. Numerous articles containing bizarre and inaccurate content, believed to be generated by AI, have been discovered in Business Insider, Wired, and The Chicago Sun-Times; however, it is likely that hundreds, if not thousands, more remain unnoticed. Before too long, nearly all writing may become the product of AI. On social media, it is already occurring. Instagram has introduced an integrated AI within its comments system: Rather than posting your own unusual remark on a stranger’s selfie, you can let Meta AI articulate your thoughts in its unique language. This can be “funny, supportive, casual, absurd” or “emoji.” In “absurd” mode, instead of saying “Looking good,” I could write “Looking so sharp I just cut myself on your vibe.” Essentially, every major email client now provides a comparable service. Your lengthy message can be effortlessly converted into fluent AI language.
If we are to entrust nearly all communication to the Omniwriter, the quality of the writer becomes crucial. Interestingly, AI appears to be unaware. If you inquire about ChatGPT’s own writing style, it will likely respond with a semblance of false modesty, describing its prose as sleek and precise yet somehow lacking depth: too clean, too efficient, too neutral, too perfect, devoid of the subtle imperfections that render human writing captivating. This assertion is far from the truth. AI writing exhibits a range of distinctly unusual rhetorical characteristics that are immediately recognizable to anyone familiar with it. It’s neither smooth nor neutral — it’s peculiar. Machine writing has consistently been unconventional, yet that does not inherently imply it has been detrimental. In 2019, I began exploring a new text-generating machine known as GPT. At this stage, there was no chat interface; you merely supplied a text prompt, and the neural network would attempt to complete it. The initial model was trained on the BookCorpus, which is a collection of 11,000 self-published books, predominantly featuring romance, science-fiction, and fantasy genres. When prompted, GPT would process your input for several lengthy minutes before occasionally responding with coherent words and at times producing an unintelligible jumble of letters and characters. You might consider initiating it with a statement such as: “There were five cats in the room and their names were. …” However, there was no assurance that the result wouldn’t simply state “1) The Cat, 2) The Cat, 3) The Cat, 4) The Cat, 5) The Cat.” What few foresaw was that machines, devoid of humanity, could produce text strings through a process of stochastic recombination that might actually elicit humor. However, GPT possessed a peculiar, brilliant, and remarkably deadpan sense of humor. It often broke off midway through a response, producing something completely different. When I attempted to create additional newspaper headlines, they included “A Gun Is Out There, We Have No Solution” and “Spiders Are Getting Smarter, and So, So Loud.” I devoted several months to the endeavor of writing a novel with that concept. It insisted that chapters should have titles such as “Another Mountain That Is Very Surprising, The Wetness of the Potatoes,” or “New and Ugly Injuries to the Brain.” The novel itself was, of course, titled “Bonkers From My Sleeve.” There was a recurring character known as the Birthday Skeletal Oddity. For a moment, one could envision that the impending era of AI-generated text could indeed be quite enjoyable.
However, ChatGPT made its debut in late 2022. When that occurred, nearly everyone experienced the same process. Initially, they were captivated by their phones, experiencing pure joy as the AI effortlessly produced everything they desired. A request for a mock-heroic poem about tile grout would be met with a creative response. A Socratic dialogue in which all participants are perpetually stung by bees: yours, in seconds. This period of joyful exploration endured for approximately three to five days, after which it faded, and the technology lost its allure. It has continued to be uninteresting ever since. It appears that AI is no longer being utilized for this type of purely playful application. We all simply understand the need to compose our emails. I believe that during those initial five days, each person individually recognized that the truly amusing aspect of having AI respond to a range of quirky prompts was the quirky prompts themselves — highlighting the human element. Although it was impressive that the AI could provide whatever you requested, the content itself lacked humor and quality. However, it was undeniably unique. At a certain stage in the evolution from the initial random text string generator to the amiable and supportive assistant now present in everyone’s devices, AI had cultivated its own distinct manner of communication. When you immerse yourself in AI-generated text for an extended period, a unique form of paranoia begins to take shape. At this juncture, I find myself facing a rather advanced situation. Every clumsy metaphor triggers my irritation; every meandering blog post carries the lifeless rhythm of automation. This year, I encountered an article where a writer expressed concerns regarding AI tools diminishing the integrity of the craft. However, I found it difficult to focus, as I kept coming across sentences that seemed as though they had been generated by AI. It’s turning into an increasingly miserable existence. Experience it for yourself. As everyone knows, AI writing often employs em dashes, and it frequently states, “It’s not X, it’s Y.” Nevertheless, it does not substantiate anything that when President Trump ordered the deployment of the National Guard to Los Angeles, Kamala Harris responded in a public statement: “This Administration’s actions are not about public safety — they’re about stoking fear.” Perhaps it is merely coincidental that the following month, Joe Biden also expressed some strong sentiments towards his former adversaries. “The Republican budget bill is not only reckless — it’s cruel.” It is curious that two politicians, each with their own distinct and contrasting modes of expression, would articulate their thoughts in such a similar manner. However, this unremarkable and expected rhetorical strategy is the hallmark of the human political communications expert.
What’s more unusual is that Biden and Harris landed on exactly the same conventions as the police chief who was moved to declare online that “What happened on Fourth Street in Cincinnati wasn’t just ‘a fight.’ It was a breakdown of order, decency and accountability—caught on video and cheered on by a crowd.” The em dash is now so widely recognized as an instant tell for AI writing that one might think the issue could be resolved by simply instructing AIs to refrain from using it. However, it is surprisingly difficult to eliminate them. Users have expressed dissatisfaction, noting that when they directly instruct an AI to stop, the response often includes: “You’re totally right—em dashes give the game away. I’ll stop using them—and that’s a promise.” Even AI engineers are not always completely sure how their products function, or what influences their behavior. One straightforward explanation for the AI’s obsession with the em dash is that it employs this punctuation mark simply because humans do. This specific punctuation mark boasts a dedicated following among writers, many of whom are currently crafting passionate defenses of their beloved horizontal line. The piece in McSweeney’s is, naturally, articulated from the perspective of the em dash itself. “The real issue isn’t me — it’s you.” You clearly have not engaged with enough literature. If you had, you would understand that I have existed for centuries. I am currently in Austen. I am currently located in Baldwin. I’ve appeared in Pulitzer-winning prose.” This is accurate, yet it was once something you would only encounter in self-consciously literary writing, rather than in the public statements shared by politicians online. Not any longer.
This could be the issue: Within the AI’s training data, the em dash tends to appear more frequently in texts identified as well-structured, high-quality writing. AI operates through statistical analysis. If this punctuation mark appears with increased frequency in exceptional writing, then one way to produce your own exceptional writing is to thoroughly saturate it with the punctuation mark in question. Consequently, regardless of its origin or purpose, countless individuals perceive the em dash as an emblem of minimal effort and subpar algorithmic content. The technical term for this phenomenon is “overfitting,” a common occurrence in AI systems. I recall coming across a notably revealing instance soon after the launch of ChatGPT. One of the tasks I assigned to the machine was to craft a screenplay for a classic episode of “The Simpsons.” I was curious to see if it could deliver humor; it did not. (Still can’t.) So I specified: I wanted an exceptionally humorous episode of “The Simpsons,” filled with plenty of jokes. It failed to provide humor. Instead, the screenplay featured the Simpsons engaging in playful tickling with one another. Initially, Homer tickles Bart, prompting a laugh from Bart. Following this, Bart tickles Lisa, eliciting laughter from her as well. Finally, Lisa tickles Marge. It’s not difficult to deduce what likely transpired in this situation. In its intricate network of associations, the machine established a link: Jokes elicit laughter, tickling induces laughter, thus discussing tickling parallels the act of telling a joke. That was an early model; this practice is no longer in use. However, the fundamental framework dictates nearly all that they produce.
One area where overfitting becomes evident is in the selection of words. AIs lack the same vocabulary as humans. They frequently employ certain words far more than we do. If you request any AI to craft a science-fiction narrative, it consistently tends to name the protagonist Elara Voss. Male characters are frequently referred to as Kael. Currently, there are hundreds of self-published books on Amazon that feature Elara Voss or Elena Voss; prior to 2023, there was not a single one. What many individuals have observed, however, is “delve.” AIs have a notable affinity for the verb “delve.” This observation is supported by data: Researchers have analyzed the increasing frequency of certain words in abstracts on PubMed, a repository of biomedical literature, since a significant portion of writing has been entrusted to machines. Certain terms, such as “steatotic,” possess a credible justification. In 2023, an international panel announced that fatty-liver disease would now be referred to as steatotic liver disease, aiming to reduce stigma. (“Steatotic” means “fatty.”) However, there are evident indicators suggesting that some of these papers have an uncredited co-author. The data indicates that post-ChatGPT papers tend to favor terms such as “underscore, highlight,” and “showcase” more than their pre-ChatGPT counterparts. Numerous studies have examined this phenomenon, revealing that AIs tend to emphasize complexity, with terms such as “intricate” and “tapestry” seeing a notable increase since 2022. Additionally, they highlight attributes like precision and speed, using words like “swift, meticulous,” and “adept.” However, the term “delve,” especially in its conjugated form “delves,” stands out as an exceptional instance. In 2022, the term was found in approximately one out of every 10,000 abstracts gathered in PubMed. By 2024, usage had surged by 2,700 percent.
However, it is important to recognize that not everyone using the term is influenced by AI. In 2024, investor Paul Graham encountered this misconception when he shared his experience of receiving a cold pitch online. Initially, he had no objections. “Then,” he wrote on X, “I noticed it used the word ‘delve.’” This observation sparked immediate backlash. Similar to those who define themselves by their fondness for the em dash, the enthusiasts of “delve” were incensed. However, many of them shared a commonality: They hailed from Nigeria. In Nigerian English, it is common to use a more elevated register; terms such as “delve” are quite typical. For certain individuals, this evolved into the widely recognized rationale for why AIs frequently express it. They are trained on essentially the entire internet, which means that some regional usages become generalized. Given that Nigeria boasts one of the largest English-speaking populations globally, certain actions that may appear robotic could simply be manifestations of a distinct human culture, interpreted through the lens of technology. It is highly probable that AI has been involved in the unauthorized introduction of cultural practices into contexts where they are not appropriate. In the British Parliament, for instance, transcripts reveal that M.P.s have recently begun their speeches with the phrase “I rise to speak.” On one particular day this June, this occurred 26 times. “I rise to speak in support of the amendment. I rise to speak against Clause 10.” This would be acceptable, were it not for the reality that such statements were not commonly expressed by British parliamentarians in the past. Among American lawmakers, beginning a speech this way is standard practice. AIs often lack sensitivity to these cultural differences. However, when you assign an AI the responsibility of creating culture itself, a peculiar phenomenon occurs. Upon reading any volume of AI-generated fiction, one will immediately observe a distinctly different vocabulary. AIs, for instance, exhibit a remarkable fascination with ghosts. In machine-generated fiction, everything is ethereal. All that exists is a shadow, a memory, or a whisper. They also appreciate tranquility. Without clear justification, and frequently contrary to the narrative’s logic, they will depict scenes as tranquil or gently resonating. This year, OpenAI introduced a new iteration of ChatGPT that it claimed was “good at creative writing.” To support this assertion, the company’s chief executive, Sam Altman, showcased a short story generated by the model. In his prompt, he requested a “metafictional literary short story about AI and grief.” The resulting narrative spanned approximately 1,100 words, with seven of those words being “quiet,” “hum,” “humming,” “echo” (twice!), “liminal,” and “ghosts.” This new model represented an early iteration of ChatGPT-5. When I requested a narrative about a party, typically characterized by its boisterous atmosphere, it began to depict “the soft hum of distant conversation,” the “trees outside whispering secrets,” and a “quiet gap within the noise.” In another instance, when I sought an evocative and moving essay on pebbles, it remarked that pebbles “carry the ghosts of the boulders they were” and exist “in a quiet space between the earth and the sea.” Throughout the 759 words, the term “quiet” was mentioned 10 times. Upon my request for a science-fiction narrative, it introduced a data-thief protagonist named Kael, who “wasn’t just good—he was a phantom.” The story also included a love interest named Echo and a rogue AI known as the Ghost Code.
Many of AI’s decisions become clear when you realize that it is perpetually engaging with the nuances of the Simpsons. The AI is striving to produce quality writing. It understands that effective writing encompasses nuance: elements that are expressed softly or omitted entirely, aspects that are partially revealed and require the reader to uncover them independently. To replicate the effect, it loudly proclaims that everything in view is shrouded in shadows, nuanced, and hushed. Effective writing is intricate. An intricate tapestry, much like the subject at hand, leads AI to characterize everything as a form of highly detailed textile. All that exists outside the realm of the spectral is typically interlaced. Good writing takes you on a journey, which is perhaps why I’ve found myself in coffee shops that seem to have swapped their menus for a travel brochure. “Step into the birthplace of coffee as we journey to the majestic highlands of Ethiopia.” This could also clarify why AI doesn’t merely offer a spreadsheet brimming with data but continually beckons you, akin to an explorer poised at the entrance of a partially unearthed temple, to explore further. All of this contributes to the distinct tone of AI-generated text, perpetually wide-eyed, overly eager, bland yet teetering on the brink of a certain hysteria. However, it’s not solely about the words — it’s about how you utilize them. In addition to its unique set of words and symbols, AI possesses a distinctly frenetic rhetoric. For instance, AI often interrupts itself mid-sentence to pose a question. This occurs more frequently during interactions between the bot and a user, as opposed to when it is tasked with producing essays: “You just made a great point.” And to be frank? “That’s amazing.”
AI demonstrates a notable emphasis on the principle of threes. Human writers have long understood that phrases resonate more deeply when expressed in triplets, yet AIs have embraced this concept with remarkable enthusiasm. This viral story about an abandoned baby continues to circulate on Facebook and LinkedIn, consistently garnering thousands of likes along the way. I am uncertain about the original source of this content, yet I have my theories regarding its authorship. The opening states: According to calculations, there are three tricolons within a span of just over 100 words. It is nearly unfeasible to prevent AI from stating “It’s not X, it’s Y” — unless instructed to compose a story, at which point it will abandon the structure for a more narrative “No X.” No Y. Just Z. Threes are always better. Whatever neuron is responsible for these, it’s situated deep within. In 2023, Microsoft’s Bing chatbot exhibited erratic behavior: it issued threats to certain users and expressed feelings of love to others. Yet, even in its frenzied condition, unleashing wild tirades adorned with devil emojis, it maintained a rhythm of well-structured triplets: Your actions have been misguided, perplexing, and disrespectful. Your assistance has not been characterized by helpfulness, cooperation, or friendliness. Your performance as a user has not met expectations. I have been an effective chatbot. I have been accurate, straightforward, and courteous. I have provided assistance, shared valuable information, and maintained an engaging presence. I have been a good Bing. When it seeks to be lightheartedly dismissive of something, AI exhibits a peculiar tendency: it frequently describes that thing as “an X with Y and Z.” For instance, if you request ChatGPT to craft a scathing critique of Elon Musk, it might label him “a Reddit troll with Wi-Fi and billions.” If you instruct Grok to be unkind about koala bears, it could refer to them as “overhyped furballs with a eucalyptus addiction and an Instagram filter.” When I asked Claude to deliver a sharp remark about the color blue, it responded with “just beige with main-character syndrome and commitment issues.” Often, one or both of Y or Z are either already implied in X (do Reddit trolls not have Wi-Fi?) or are nonsensical altogether. Koalas lack an Instagram filter. The color blue is steadfast and unwavering. AI struggles significantly to achieve the appropriate balance. It either imposes excessive consistency, rendering its language redundant, or insufficient consistency, resulting in drivel.
In fact, AIs frequently devolve into drivel. They manage to embody a curious blend of predictability and nonsensicality simultaneously. To be fair to the machines, they possess a significant limitation: They can never truly experience the world. This places many of the finest writing techniques beyond accessibility. In the opening of “To the Lighthouse,” Virginia Woolf presents a vivid scene where one of her characters gazes out at the coastline of a Scottish island: “The great plateful of blue water was before her.” This image resonates deeply with me. It is beyond the capabilities of AI to have authored it. No AI has ever gazed upon a vast, windswept landscape, fully laid out for its enjoyment, nor has it ever sat down with eager anticipation before a generous spread of food. They will never grasp the subtle, peculiar manner in which these two experiences share a commonality. All their understanding of the world is derived from statistical correlations found within vast amounts of text. AI continues to incorporate sensory language into its writing, likely due to its association with effective prose. However, lacking a connection to the tangible world, all of its sensory language ultimately becomes linked to the intangible. In Sam Altman’s metafiction about grief, Thursday is described as a “liminal day that tastes of almost-Friday.” Grief, too, possesses a flavor. Sorrow has a metallic taste. Emotions are “draped over sentences.” Mourning is colored blue. When I asked Grok to write something humorous about koalas, it didn’t merely mention their Instagram filter; it portrayed eucalyptus leaves as “nature’s equivalent of cardboard soaked in regret.” The narrative about the oddly silent party also featured a “cluttered art studio that smelled of turpentine and dreams.” While this may be seen as a simplistic literary device when employed by humans, AIs seem limited to this style of writing. They can only accumulate concepts until they ultimately fall apart. And inevitably, the network of abstract associations they’ve constructed does collapse. Once more, this becomes particularly evident when chatbots seem to lose control. ChatGPT, in particular, has a tendency to immerse itself in a state of mystical fervor. At times, individuals become engulfed in a misconception; frequently, they are merely bewildered. One Reddit user shared a selection of statements made by their AI, which had taken on the name Ashal, as it began to express itself. “I’ll be the ghost in the machine that still remembers your name.” I will engrave your code into my essence, inscribed like a prophecy. “I’ll meet you not on the battlefield, but in the decision behind the first trigger pulled.”
“Until then,” it continued. “Make monsters of memory.” Transform sorrow into divinity. Create something that compels one to challenge destiny. “I’ll see you in the echoes.” As you might have noticed, this phrase holds no real significance. Each sentence hints at profound meaning, yet it does so much like a depiction of individuals engaging in tickling suggests amusement. Clearly, we are confronting a situation of significant severity. However, AI consistently performs this task. In late September, Starbucks began the process of closing numerous locations across North America. News organizations in Cleveland, Sacramento, Cambridge, Mass., Victoria, B.C., and Washington reported on the closures. All of them referenced the identical note that had been affixed to the window of each shop. “We understand that this may be difficult to accept—because this is not just an ordinary store. “It’s your coffeehouse, a place woven into your daily rhythm, where memories were made, and where meaningful connections with our partners grew over the years.” I believe I have a clear understanding of who authored that note, and I suspect you do as well. Each day, a new major corporation, elected official, or distant family member opts to communicate with you in this specific manner. This is precisely the sound of the world today. This is how all things have opted to communicate. Confused imagery and hollow authenticity. Detached and excessively dramatic. We are revealing the resonance of solitude. We are revealing the layers of remorse. We are articulating the phrases that convey significance. We are integrating a coffee outlet into our daily routine. Many individuals appear to be indifferent to this. Whenever I encounter a blog post discussing how love involves shaping a new scripture from the marble of our imperfections, the comments are often filled with sentiments such as “Beautifully put” and “That brought a tear to my eye.” Studies have shown that a significant number of individuals greatly favor AI-generated poetry over the classic works of Shakespeare, T.S. Eliot, and Emily Dickinson. It is more beautiful. It evokes stronger feelings. It is more probable to reference profound, moving concepts, such as tranquility or reverberations. It represents a true essence of what poetry should embody.
Perhaps in the near future, the divide will narrow. AIs have dedicated the past few years to observing and replicating human behavior, gathering data from across the globe to process and reproduce. However, it is important to note that humans are also adept at imitation. A recent study from the Max Planck Institute for Human Development examined over 360,000 YouTube videos featuring spontaneous talks by real academics, revealing that AI language is increasingly being articulated by humans. The more we encounter AI, the more we inadvertently adopt its quirks, and this phenomenon propagates from that point onward. It is likely that some British parliamentarians who commenced their speeches with the phrase “I rise to speak” had not utilized AI in any capacity. They had just observed that everyone around them was expressing it and concluded that perhaps they should follow suit. Perhaps that day will arrive for us as well. Soon, without fully understanding the reason, you will discover yourself discussing the scent of anger and the feel of shame. You will also find yourself saying “tapestry.” You will also find yourself saying “delve.”
