Kevin Roose’s human-written book about the AI decade, and what leaders should steal from his process
Kevin Roose remembers the moment he committed to a book about generative AI: “early last year, 2025, and I was in the car on the Bay Bridge stuck in traffic, ” he told WIRED reporter Katie Drummond (WIRED, Oct 6, 2026). The result is The AGI Chronicles, built from roughly a year of reporting and “more than 150 interviews.”
That mix of deep reporting, institutional portraits of OpenAI, Anthropic, and Google, and a practical use of AI tools during production makes Roose’s work valuable to business leaders. His process offers a pragmatic playbook for integrating AI into serious, high-stakes work without ceding final judgment.
How he actually used AI: clear, concrete, human-led
Roose insists the book’s prose was written by “a very tired, overworked, under-slept human being” (WIRED). He used AI deliberately at multiple stages. A giant NotebookLM (Google’s experimental AI research notebook) acted as a searchable repository for interview transcripts, articles, and papers. An “AI swarm” helped with fact-checking alongside human fact-checkers, and a “Council of Claudes” (multiple instances of Anthropic’s Claude model) offered varied editorial feedback. He also credits human collaborators like researcher Jasmine Sun and editors at FSG.
Put simply, machines amplified research and iteration, and humans retained authorship and final editorial control. That hybrid, combining aggregation, multi-model verification, and human sign-off, is a concrete model leaders can adopt today.
Money, media, and a new venture
Roose and longtime collaborator Casey Newton left the New York Times’ Hard Fork feed to launch Machine Gods Media and a podcast called Machine Gods, with a distribution partnership announced with NPR (Roose praised Nadine Zylstra, NPR’s chief content officer, during the conversation). Bloomberg reported offers “of up to $5 million” for the show. Roose replied, “It’s not $5 million, ” later adding, “So much more, Katie. No, look, they have made us a good offer. We could have gotten more money elsewhere” (WIRED, Bloomberg coverage referenced in the interview).
The dollar figures matter, but the strategic signal matters more: legacy platforms see demand for sustained, even skeptical coverage of AI that reaches beyond the tech echo chamber. Roose frames the editorial mission succinctly: the venture “takes AI progress seriously, is clear-eyed about the capabilities and risks of powerful AI systems, and tries to empower and entertain people in the face of radical uncertainty.”
People shape defaults, and the industry’s concentration matters
The AGI Chronicles follows three labs and their overlapping personnel histories. Roose highlights Dario Amodei’s departure from OpenAI to found Anthropic (taking colleagues with him), and he names Amanda Askell, a philosopher and researcher at Anthropic, as a key figure shaping Claude’s behavior (Roose colorfully calls her the “Claude mother” in conversation). Roose’s shorthand: “50 people in San Francisco, give or take, made all of the relevant decisions, and they are not a representative sample.”
That’s an important point for executives. The personal biases, incentives, and default priorities of early teams ripple into product design, safety investments, and governance choices. Roose’s posture toward leadership is practical: “I don’t trust any single person.” For organizations that translates into designing institutional checks rather than relying on individual virtue.
On technical credit and careful phrasing
Roose recounts a scene about the DeepMind protein-folding breakthrough and a parent writing to researchers asking whether the work could help her child. The precise record is this: the researchers behind AlphaFold2, Demis Hassabis and John Jumper, and researcher David Baker, for complementary computational design work, were awarded the 2024 Nobel Prize in Chemistry for advances in protein-structure prediction that used AI methods. The Nobel committee credited the human researchers and their methods. It’s more accurate to say the researchers won the prize for techniques that employed AI than to say an AI system itself “won” the Nobel Prize.
Likewise, modern generative models rest on the transformer architecture (from the 2017 paper “Attention Is All You Need”), a technical lineage that matters when you’re making decisions about product risk, model choice, or IP.
Roose’s stance on safety: guarded optimism
Roose records industry actors who say the work is inevitable, “Some of them, including Dario and Sam, genuinely believe that this technology is inevitable.” Roose himself is skeptical of inevitability and reports feeling “quite hopeful right now relative to where I was a few months ago, and it’s largely because we are now having this conversation.”
For boards and executives, the practical takeaway is that the debate has moved into mainstream institutions. That creates an opening for governance, disclosure, and product design choices that reduce downstream harms, if companies take those openings seriously.
Actionable takeaways for business leaders
- Design explicit human approval gates. Require named human sign-off on any customer-facing model output. Maintain an immutable audit trail of prompts, model versions, and human approvals so you can trace decisions if something goes wrong.
- Run multi-model cross-checks and a quarterly red-team cadence. Use two or more models (or model families) to surface inconsistencies, then schedule formal red-team reviews at least quarterly. Treat model disagreement as a trigger for human investigation, not a technicality to ignore.
- Assign a single accountable owner for AI governance. Give one executive responsibility for policy, compliance, and incident response, and require that role to publish a short annual disclosure about model uses, safety tests, and remediation steps.
- Preserve domain expertise, and let AI augment, not replace it. Use tools like NotebookLM-style aggregation to concentrate subject-matter knowledge; keep experts responsible for interpretation, not just validation.
- Adopt a simple public disclosure line. Example template: “This product uses AI assistance for [task]. Human review and oversight are used for decisions that affect customers.” Route such disclosures through legal/compliance before publication.
Roose’s process is a useful case study because it shows how to combine scale and rigor. AI can accelerate research, surface threads across hundreds of interviews, and generate fast editorial critique, but final narrative decisions, responsibility for accuracy, and public accountability remained human. That balance matters for any company shipping AI features or using AI in regulated contexts.
Questions leaders are likely asking: concise answers
- Did Kevin Roose use AI to write The AGI Chronicles?
He says the prose was human-written, “a very tired, overworked, under-slept human being”, but he used AI extensively for research, organization, fact-checking, and editorial feedback (NotebookLM, an “AI swarm, ” and a “Council of Claudes”), alongside human editors and researcher Jasmine Sun (WIRED, Oct 6, 2026).
- How much reporting went into the book?
Roose reports conducting “more than 150 interviews” and spending “about a year” on reporting and writing (WIRED, Oct 6, 2026).
- Is the Machine Gods podcast deal worth $5 million?
Bloomberg reported offers “of up to $5 million.” Roose disputed that specific figure in the WIRED conversation: “It’s not $5 million, ” and later said they received “a good offer” while suggesting figures were negotiable. Exact deal terms were not disclosed in the interview.
- Should my company copy Roose’s workflow?
Copy the patterns, not the exact tools: centralize research, run multi-model checks, keep a named human in the approval loop, and require auditable records. Those steps scale to enterprise use-cases and help manage both regulatory and reputational risk.
- Did an AI win a Nobel Prize?
The 2024 Nobel Prize in Chemistry honored human researchers, including Demis Hassabis and John Jumper (jointly) and David Baker (separately), for breakthroughs in protein-structure prediction that used AI methods. The prize recognized the researchers and their methods, not an autonomous AI as a recipient.
Final thought
Roose offers a practical example of responsible hybrid work: use AI to make human experts faster and broader, but keep accountability, authorship, and final judgment clearly human. For leaders, that’s a governance blueprint you can implement this quarter: reduce single-point trust in individuals, require auditable approval processes, and treat AI as an instrument that widens human reach, not as a substitute for human responsibility. For traceability, maintain an audit trail of model versions, prompts, and approvals so incidents can be investigated and remediated.