AI Still Can't Write the Fiction We Actually Want to Read
A recent test using Claude to generate passages in the style of classic authors-George Eliot, Henry Fielding, Bram Stoker, Ernest Hemingway-revealed a persistent weakness: the AI produces technically competent prose that fools most readers, but the scenes it creates remain fundamentally inert.
The test presented roughly 200-word samples to players and asked them to identify human-written versus AI-generated text. Real passages came from Project Gutenberg; AI versions mimicked the style of public-domain authors. Early results showed clear tells: formatting problems, tortured metaphors, characters who fidgeted constantly but did little else.
When given feedback, Claude improved. It eliminated obvious tics like overused similes and vague words. It generated secondary AI agents to check its own work for mistakes. By the end of the week, the system fooled more than half of players-with one Bram Stoker passage deceiving 83 percent of respondents.
Yet even the best fakes shared a structural flaw. Characters existed in a state of stasis. They walked empty corridors, heard only wind, reflected on absence. When asked to inject more action, Claude reverted to awkward, easily identifiable prose-characters suddenly running or delivering packages with no narrative justification.
The Problem Isn't Style-It's Agency
The AI had learned to approximate authorial voice. It understood Hemingway's short declarative sentences, Eliot's architecturally balanced clauses, Stoker's epistolary restraint. But it couldn't reliably place characters in scenes where meaningful things happened.
This matters because fiction depends on choice-the author's decision about what happens, why it matters, and how it reflects something true about human experience. Claude draws from every word ever written, not from lived observation or genuine uncertainty about how a scene should unfold.
The result reads like competent pastiche. It satisfies a technical checklist. It doesn't compel.
Why This Isn't the End of Writing
The discomfort readers feel encountering AI-generated prose isn't dread about human obsolescence. It's the same unease you'd feel discovering your chess opponent used a bot to plan moves. The medium itself-whether email, legal brief, or short story-matters less than knowing a human made deliberate choices.
Writing, fundamentally, is the process of converting understanding into words that communicate with other people. Skipping that process always feels like cheating, regardless of whether AI could technically produce the same output.
For AI for Writers, the practical implication is straightforward: Generative AI and LLM tools excel at producing functional prose-emails, summaries, first drafts. They remain inadequate at generating literature readers actually want to encounter. That gap will likely persist because it reflects something deeper than a technical limitation waiting to be solved.
The value of human writing isn't about speed or volume. It's about the thinking that precedes the words.
Your membership also unlocks:
