Everyone’s worried AI will replace authors. So I decided to test it. I fed Claude Sonnet 4.5 nearly 100,000 words of my YA space opera—the complete novel, 5,000 words of a prequel I’d already written, character guides, alien speech patterns, explicit instructions about my protagonist’s psychology. Then I asked it to write the next scene. The result? Competent genre prose that lost my protagonist’s voice entirely. It could analyze what made her voice work, explain it back to me perfectly, then defaulted to templates anyway when asked to generate prose. Grok 4.1 failed the same experiment. This isn’t about whether AI will improve. It’s about understanding what AI fundamentally can’t do—and what that means for writers.