Can A.I. Help Write History?
During my 25 years as a magazine editor, I’ve relished helping writers shape their stories. So I watched with some humility as one of my longtime collaborators, Steven Johnson — a tech journalist and historian — received story guidance not from me, but from an artificial intelligence in a Google cafeteria in Manhattan.
Johnson, known for books on pirates, policing, and public health, was contemplating a new book on the California gold rush but lacked a specific angle. “What’s my twist?” he asked. To explore it, he turned to NotebookLM, an A.I. app he helped develop at Google, which synthesizes research from sources the user uploads, unlike traditional chatbots that draw from massive pre-trained data sets.
Since its global launch, NotebookLM has been marketed for tasks ranging from meeting notes to student research. Among its flashier features is an auto-generated podcast simulating a conversation about the user’s research. But Johnson’s focus was more profound: how A.I. could assist in the process of writing history — quickly analyzing sources, identifying themes, and even shaping narrative structure.
To begin, Johnson uploaded excerpts from H.W. Brands’s The Age of Gold, Lafayette Bunnell’s Discovery of the Yosemite, and two Indigenous perspectives: The Ahwahneechees and Indians of the Yosemite Valley and Vicinity. Then, identifying himself as the author, he queried the A.I. about what the Indigenous sources included that the others lacked.
NotebookLM highlighted how The Ahwahneechees humanized Native individuals through short biographies, unlike the broader tribal portrayals in the other texts. One name stood out: Maria Lebrado, granddaughter of Chief Tenaya. The A.I. pulled together a biography noting her displacement by the Mariposa Battalion in 1851, her marriage, and her late-life return to the valley.
Johnson was intrigued. He imagined opening the book with Lebrado’s emotional return at nearly 90 years old, then flashing back to her youth during the violent 1850s. The A.I. had surfaced a compelling narrative possibility in just 30 minutes of work. “It delivered a concept I could absolutely use,” Johnson remarked.
As a writer myself, the rise of large-language models evokes both curiosity and dread. Letting A.I. do the reading — especially in an era when digitized information overwhelms even the most diligent researchers — feels both seductive and unsettling. With millions of historical documents now a search away, the challenge is no longer access but absorption.
A.I. offers a solution: let it process the material and highlight themes. Historians are already using it. Stanford’s Fred Turner used ChatGPT to shape the structure of a book on New York’s 1970s art scene. It gave his scholarly work a “middlebrow voice” and surfaced useful links in his outline — essentially simulating a test audience’s feedback.
At Wilfrid Laurier University, Mark Humphries used A.I. to analyze thousands of fur-trading records, identifying trading partner networks in seconds that would take graduate students weeks. For him, A.I. accelerates what’s already possible — helping historians see interconnections faster.
Still, not all are sold. Princeton’s Ada Ferrer used ChatGPT to brainstorm book titles but found none quite right. Others, like Pulitzer winner Jefferson Cowie, are uneasy about using tools they caution students against. The tension between ethical caution and creative utility runs deep among academics.
A key concern is accuracy. Charles C. Mann, author of 1491, found that A.I. surfaced good leads but sometimes fabricated details. Unlike a human editor who asks tough questions, A.I. lacks a “bullshit detector,” and its hallucination problem — inventing facts or references — remains unresolved.
Recent tests confirm the issue. OpenAI’s latest model, for instance, returned inaccurate answers one-third of the time. Johnson argues that NotebookLM’s narrower input reduces such errors, since it only works from uploaded, curated sources — but even then, it can still misquote or misinterpret.
To Johnson, the most promising use is with active oversight. He imagines A.I. not as a ghostwriter but as a research assistant or perceptive editor — someone who helps identify the strongest version of your own ideas. For him, it’s a tool of inspiration, not substitution.
But not all writers agree. Biographer Stacy Schiff, when asked about using A.I. for structure, likened it to “enlisting someone to eat your hot fudge sundae for you.” For her and many others, the creative struggle is part of the pleasure — and the point.
