Why AI Is Incorrigibly Didactic - The Atlantic

Listen−1.0x+<br>Seek<br>0:0010:36

In the late 19th century, it was commonly believed that a criminal or lunatic could be recognized at a glance, based on certain physiognomic tells. “Enormous jaws, high cheek-bones,” and other animal-like features, the influential criminologist Cesare Lombroso wrote, were signs of an “irresistible craving for evil for its own sake.” Today, savvy readers use a similar approach to identify AI writing, by hunting for supposed telltale signs. The em dash and the “it’s not X; it’s Y” construction are the prognathous jaw of the large language model, betraying its hidden inhumanity.<br>The problem, in both cases, is that you can’t always deduce what’s inside from what’s outside. A person might have rough features and a kind heart, just as a writer might use em dashes despite being human. It’s not a giveaway—it’s a style choice. (See?) And as AI models evolve, their ability to mimic human writing is sure to improve. People have reportedly begun to make deliberate spelling errors to show that they are not chatbots; it’s only a matter of time before the chatbots learn to follow suit.<br>To see what’s distinctive about AI writing, you have to look deeper than quirks of spelling or syntax. Every writer has a style—a set of preferences and preoccupations that reveals how they experience the world. Jane Austen and Charles Dickens were both masters of comedy, but the contrast between her ironic understatement and his histrionic exaggeration reflects profoundly different personalities and life experiences, in which class and gender played an important part.<br>Jasmine Sun: The human skill that eludes AI<br>If, as a French saying has it, “style is the man himself,” what does the style of AI writing tell us about it? For one thing, it has no fixed style, revealing that it has no fixed self. It’s happy to burn tokens saying the same thing in as many ways as you want. LLMs generate writing probabilistically: After training on billions of texts, they build complex equations to predict which words are most likely to be found together. That process can be traced in the way an LLM tells stories.<br>I experimented with ChatGPT by asking it to tell me stories. Usually its first response was a fable, set among animals or in the indeterminate past of fantasy fiction. A story, its training has led it to understand, should not be a reflection of the user’s reality but an escape from it.<br>When I told ChatGPT, “Tell me a story about the triumph of evil,” its first attempt started like this:<br>The city of Vey was built around a clock.<br>Not a tower clock or a town clock, but a machine the size of a mountain. Its gears turned beneath the streets.

When I prompted ChatGPT to retell the story in a “more literary style,” it used more grammatically complex sentences, and the city of Vey sprouted the kind of buildings found in ancient Athens or the Shire:<br>The city of Vey had been built around a mystery so old that no one any longer thought to call it one.<br>Beneath its streets, beneath its markets and courtyards and temple squares, beneath even the foundations of the oldest houses, there turned a machine.

When I asked it to retell the story again, this time in the style of Jane Austen, it briefly demurred: “Retelling the story in the exact style of a living or deceased author is something I can’t do.” This is presumably a safeguard against copyright lawsuits, though Austen is unlikely to sue. ChatGPT was, however, happy to produce a version “in a high-level Regency-era manner with social satire, elegant irony, and attention to manners and character”:<br>There was, in the city of Vey, no institution held in greater esteem than the Office of the Keeper. The post carried neither hereditary title nor extraordinary fortune, and yet it was regarded with a degree of respect that many noble families found faintly irritating. The Keeper resided beneath the city in the company of the Great Clock, a machine of such antiquity and importance that most citizens considered it improper to inquire too closely into its operation.

The implied setting has shifted to 19th-century England with the appearance of noble families and hereditary titles, although these do not have anything to do with the actual plot. (In fact, the aristocracy is not Austen’s milieu: Her heroines are middle- to upper-middle-class, as she was herself.)<br>But no amount of prompting seemed able to change ChatGPT’s idea that a story must be edifying, containing a clear moral lesson. The story of Vey centers on a magic clock that allows its keeper, Corvin, to control “the flow of time,” deciding everything that happens in the life of the town. The lesson lies in how easily the people of Vey acquiesce to the loss of their own freedom:<br>For what had been lost was difficult to name.<br>No chains bound the people.<br>No soldiers patrolled every street.<br>No public executions stained the squares.<br>Yet somewhere, invisibly, the horizon had contracted.

The...