Kevin Roose says his new book about the people building advanced artificial intelligence was written by people, not generated by a model. But in an interview published by WIRED on October 6, he also described an extensive AI-assisted workflow for organizing research, checking facts and challenging his conclusions. The distinction he draws is becoming a consequential one for journalism: automation can support the work without replacing the authorial act.

Roose told WIRED global editorial director Katie Drummond that The AGI Chronicles emerged from more than 150 interviews and roughly a year of reporting and writing. The book follows OpenAI, Anthropic and Google, tracing how people who once worked together came to lead intensely competitive organizations. Roose said he wanted to leave a durable historical record of a period whose internal discussions can otherwise disappear into private messages and auto-deleting workplace channels.

The book's prose, Roose said, came from him, with research help from Jasmine Sun and editing from people at publisher FSG. He described himself as tired and overworked by the end of the project. That claim of human authorship sits beside a detailed disclosure of how he used AI around the manuscript, a transparency choice he said was important enough to address near the beginning of the book.

Abstract AI threads organize a large archive of transcripts, papers and reporting notes.
NotebookLM served as an expanded search and organization system for a research archive built from more than 150 interviews.

For research, Roose loaded interview transcripts, articles, academic papers and other material into a large NotebookLM notebook. He used it as an expanded search system, asking it to surface stories or details scattered across his source collection. He also said the tools helped organize notes, identify people connected to reported decisions and support fact-checking. A human fact-checker reviewed the book as well, while an AI-based checking process caught some issues that both the author and the human reviewer had missed, according to Roose.

He also used a group of Claude personas that he calls a “Council of Claudes” to critique drafts from different positions. One persona approached the work as a committed skeptic of large language models, flagging places where the text might anthropomorphize a system. Another took a more futurist view and pushed for a broader vision of what the technology could become. Roose said some of the output was unhelpful and may have consumed more time than it saved, but some criticism prompted revisions.

That workflow illustrates the boundary Roose is defending. The systems retrieved, sorted, checked and argued with material; he says they did not compose the book. He told WIRED that he has also used AI in recent months for research, editing and improving his journalism, and argued that news organizations should experiment with tools that make reporters more capable instead of using them simply to fill websites with low-quality material.

Abstract AI critics surround a manuscript while a human editor marks revisions in the foreground.
Specialized Claude personas challenged the draft from skeptical and futurist perspectives, while the author retained editorial control.

Roose connected that practical view to a broader position he and journalist Casey Newton call AI realism. Their newly announced company, Machine Gods Media, and its podcast partnership with NPR are intended to occupy a middle ground between claims that the AI boom is entirely fake and coverage that treats every new capability as an unquestioned breakthrough. Roose argued that AI systems can be powerful, impressive and dangerous at the same time, and that coverage can explain those tensions without becoming promotional.

His reporting also left him uneasy about who shaped the current moment. Roose told WIRED that a small group of people in San Francisco made many of the relevant decisions and that they are not representative of the wider public. He said the leaders he studied are fallible, their motives are mixed and no single person should be trusted with the future of AI. At the same time, he contrasted them with earlier social-media leaders by arguing that today's AI executives have long acknowledged that their technology could be dangerous, even while racing to develop it.

The interview closes on the value audiences place on human effort. Roose said people may rate AI-generated writing highly when they do not know its origin, yet react negatively after learning that a machine produced it. His conclusion is that writers and other creators will need to show more of their process so audiences can see the labor and point of view behind the finished work. For journalism, that makes disclosure more than a footnote: it is part of the evidence that a human remained responsible for the story.