When a literary magazine published a Commonwealth Prize-winning story that readers immediately flagged as AI-generated, it sparked predictable chaos in writing communities. But beneath the outrage lay a more unsettling question: how close are we to AI producing prose we'd actually want to read? A New Yorker investigation by Ted Friends tried to find out—and what he discovered is both reassuring and deeply weird.

Building the Detector

Friends vibe-coded a simple game using Claude that presented roughly 200-word passages mimicking classic authors like George Eliot, James Joyce, Ernest Hemingway, and Arthur Conan Doyle. He then asked players to distinguish real from fake. Initial results were laughably easy—AI prose screamed its origins through 'tells' like excessive em dashes, tortured metaphors, conspicuous verbs like 'delve,' and characters who constantly fidgeted: running fingers along table edges, adjusting collars. The most damning tell was abstract and almost spooky—the AI scenes had characters, but they mostly did nothing.

Refining the Fakes

Friends spent days iterating with multiple AI agents checking each other's work. He created a system where one agent scanned for mistakes while another generated style-specific 'cue cards'—rules for mimicking each author. For Hemingway, Claude instructed itself to avoid Latinate vocabulary and subordinate clause stacking. For George Eliot, it mandated long architecturally-balanced sentences with embedded clauses. The similes went away. But when Friends pushed for more action, every fake passage filled with characters hopping on horses or delivering packages—the stunted AI prose kicked back in. 'I told Claude that it wasn't doing a very good job of unlearning its bad habits,' Friends writes.

The Results Are In

Within five days, over 30,000 people took the test. Players could identify real versus fake only about 52% of the time—essentially random guessing. But roughly 10% of players consistently outperformed the baseline, suggesting some humans can still spot tells that escape most readers. The robot Bram Stoker was the most convincing: only 17% of players detected it as fake. Friends admits this is 'definitely a better facsimile' than earlier attempts—but notes the prose still describes absence and stasis, with narrators walking empty corridors hearing nothing but wind.

The Core Problem

Despite improvements in style matching, AI-generated passages consistently feature characters who don't act—they touch furniture, look at fireplaces, and exist in scenes where 'nothing happens.' Friends connects this to literary critic James Wood's concept of 'a camera's easy swipe,' arguing that authorial choices always push through the surface. AI makes choices too, but by lifting from everything ever written rather than drawing on personal reveries about a Paris street at dusk. When asked why robots needed other robots to check their work, a UC Berkeley machine learning professor replied: 'One hundred percent serious answer: No one knows.'

The Chess Parallel

Friends ends on an optimistic note that reframes the entire anxiety around AI writing. Grand masters haven't beaten chess computers for two decades, yet hundreds of thousands of kids follow chess influencers on TikTok. We still value the human process—how the game makes our brains move. 'The superiority of the machines is irrelevant when it comes to why we play.' The discomfort reading AI-generated text isn't dread that our usefulness is ending; it's the same disappointment you'd feel discovering your chess opponent used a bot.

Key Takeaways

  • AI can now mimic famous authors well enough to fool ~80% of readers in controlled tests
  • The persistent tell: AI characters don't act—they fidget, observe, but rarely do anything consequential
  • Even with multiple agents checking each other for mistakes, the stasis problem won't resolve
  • Humans still value the process of writing and reading, not just the output

The Bottom Line

The robots can fake Hemingway's style. They can't fake Hemingway's iceberg. And as long as that gap exists, AI will remain a ghost in the machine—technically impressive, fundamentally hollow, and unable to replicate why we write in the first place.