Kevin Roose Wrote His AI Book the Old Way , And That's the Story

What actually happened

Kevin Roose spent two years writing The AGI Chronicles without touching a single LLM for prose generation. He told Wired's Lauren Goode the book "was written by a very tired, overworked, under-slept human being." This isn't a stunt , Roose covers AI daily for the Times and has early access to every major model. He could have used GPT-6, Claude 5.5 Sonnet, or Gemini 3.1 Pro. He chose not to.

Why the irony matters

The guy who explains AI to millions decided the technology he covers wasn't good enough to write his own book. That's a stronger signal than any benchmark. Roose isn't a luddite , he uses AI for research, transcription, and brainstorming. But for the actual sentences? He didn't trust the output. When the industry's most plugged-in journalist passes on the tool, we should ask what he sees that the demos hide.

The technical reality nobody admits

Current models still choke on book-length coherence. They lose track of characters, contradict earlier chapters, and flatten voice into generic "helpful assistant" tone. Roose would have spent more time fixing the model's drift than writing fresh. We've seen this with every author who tried: the first 3,000 words look great. By chapter four the model hallucinates plot points it invented three chapters ago. Context windows don't solve narrative memory.

Compare this to the Amazon spam flood

Last year Amazon had to cap self-published books at three per day because AI-generated sludge overwhelmed the platform. Those books were written in hours by people who never read them. Roose took two years. The contrast tells you everything about where the quality bar actually sits. The market is splitting: infinite cheap garbage on one side, scarce human craft on the other. Middle ground is vanishing.

The trust signal is the product now

Roose's publisher didn't hide the human-only process , they led with it. "Written by a human" is becoming a premium label like "organic" or "handmade." We're three months from seeing "No AI Used" badges on book covers and Substack mastheads. This isn't sentimentality. Readers who pay for nonfiction want to know the author actually thought through the arguments. AI provenance is a liability, not a feature.

What Roose's exhaustion reveals

He described himself as tired, overworked, under-slept. That's the reality of writing a real book , it costs you something. The AI pitch promises to remove the cost. Roose's refusal admits the cost is the point. The struggle to structure an argument across 300 pages is where the thinking happens. Outsource the struggle and you outsource the insight. He knew that. The industry still pretends otherwise.

The next six months will prove him right

Watch for the first major "AI-written" bestseller that isn't a stunt. It won't happen. What we'll see instead: more authors explicitly marketing human authorship, more contracts with "no AI" clauses, more readers asking for provenance. The hype cycle said AI would replace writers by 2025. Roose just showed us the replacement still can't hold a narrative thread for 80,000 words. The timeline just moved again.

This is the adoption curve's honest moment

Every transformative technology has a phase where the boosters claim it's ready for everything and the practitioners quietly keep using the old tools. We're in that phase for long-form writing. Roose didn't make a speech , he just did the work the hard way. That's the most credible benchmark we have. When the New York Times' AI columnist writes his AI book without AI, the conversation shifts from "when" to "whether."