Steven Johnson's Advice on Writing with AI: The Practical Optimist

Steven Johnson wrote Where Good Ideas Come From, How We Got to Now, The Ghost Map, and a dozen other books about innovation and the history of ideas. He co-created NotebookLM at Google. He has spent thirty years thinking about how tools shape thinking, and he is now one of the most credible voices on what AI means for writers.

His writing advice comes from his conversation with David Perell on How I Write, his Substack Adjacent Possible, and his Google Workspace interview about writing his latest book, The Infernal Machine.

The Slow Hunch

Johnson's most famous idea applies directly to his own writing process. Good ideas are not lightning bolts. They are slow hunches that develop over months, years, sometimes decades.

His book Enemy of All Mankind proves the point. The idea for the book's structure came years before its content. Johnson wanted an hourglass shape: a single compressed event in the middle, with all the forces that produced it flowing in from one side and all the consequences flowing out the other. He wrote that structure idea down and sat on it for three or four years. Then he decided the event should be a crime - crimes happen fast, in seconds or minutes. Another year passed before he stumbled across the pirate Henry Every, who pulled off arguably the largest heist in history in 1695. Structure preceded content by half a decade.

The danger is that slow hunches are fragile. They get crowded out by urgent tasks, forgotten in the churn of daily work. Johnny Ive said at Steve Jobs' memorial that young ideas are like baby doves - hold them too roughly and they die. Johnson agrees, but adds a twist: you do not even need to nurture them actively. Just capture them and resist the urge to criticize them immediately. Give them time. That alone is often enough.

The Spark File

Johnson's system is deliberately simple. He maintains a single document he calls the Spark File. Whenever he has an idea, a question, a half-formed connection, he adds it chronologically. No tags. No folders. No cross-references. Just a running list.

Every six months or so, he scrolls through the entire file. This review is where the magic happens. An idea from two years ago collides with something he wrote last month. A pattern emerges. A book takes shape.

The spark file works because it creates what Johnson calls collisions between ideas that would never meet inside the architecture of a well-organized system. Folders separate things. A chronological list forces everything into proximity. The randomness is the feature.

This mirrors his concept of the adjacent possible from Where Good Ideas Come From. Innovation happens at the boundary between what exists and what is newly reachable. The spark file keeps your thinking at that boundary. Each review session expands the adjacent possible of your own mind.

AI as Research Partner, Not Writing Partner

Johnson draws a sharp line. AI is extraordinary for research. It is dangerous for drafting.

In the early 2000s, he used DevonThink to organize thousands of research quotes and experienced a glimpse of what was coming. He searched for material on waste recycling while writing The Ghost Map, and the software surfaced a seemingly unrelated quote about how our skeletons evolved from calcium, a waste product of cell metabolism. Cells were expelling calcium; evolution repurposed it into a backbone. He turned the connection into seven pages of the book. When he stepped back, he asked: who had that idea? "It felt like a duet. Two different kinds of intelligence working together."

NotebookLM fulfills that promise at scale. He uploads research and asks questions in natural language. The real gain is not speed - it is staying in flow. Every time you leave the manuscript to hunt for a fact, you lose the thread. Johnson's goal is to write at the speed of thought, and the bottleneck has always been research interruptions.

But the prose is his. The structure is his. The argument is his. AI can create the illusion of understanding. "If you want to create the illusion of understanding something, you can indeed ask NotebookLM a question with the right sources and just copy and paste the answer." That is not writing. That is sophisticated plagiarism of your own research. Real understanding comes from wrestling with the material. AI accelerates how quickly you reach the wrestling match. It cannot do the wrestling for you.

Surprise as Structure

Johnson's books succeed partly because they are surprising. The Ghost Map turns a cholera epidemic into a detective story. How We Got to Now connects the invention of glass to the Renaissance. Where Good Ideas Come From finds innovation patterns in coral reefs.

He treats surprise as a structural element, not just a rhetorical trick. The brain is constantly predicting what comes next. When you are bored, you know what is coming. When you are leaning forward, you are surprised by what is coming. When you are truly surprised, you are confused. The writer has to hit the sweet spot: the reader does not quite know what is happening, but is intrigued rather than lost.

His New York Times Magazine piece on GPT demonstrated this at the sentence level. The first paragraph deliberately omits its final word. The reader fills it in - "paragraph" - and in doing so experiences the exact prediction mechanism that large language models use. Johnson conceived the trick in a San Francisco hotel room, sent the opening to his editor with the subject line "Are you gonna let me get away with this?" The editor said he might. The structural surprise became the launching pad for 10,000 words.

For The Infernal Machine, he wanted the preface to plant a bomb in the reader's mind - literally. He wrote a sequence where the prose pulls the reader out of a bustling NYPD identification bureau, down the stairs, to a briefcase left suspiciously in a doorway, to the sound of ticking inside. The cinematic movement happens in five or six sentences, but the structural work it does supports the entire book's architecture. The reader needs that detonation to be memorable enough to tolerate the long, multi-threaded journey before the story circles back.

Never Reread While Writing

Johnson battles his own familiarity with the text. A book is a song you have heard too many times. Read a chapter twenty times and it loses its appeal. So much of writing is projecting yourself into the reader's experience - are they confused, excited, bored? - and that projection becomes impossible when every sentence is memorized.

His technique: actively resist rereading chapters as he writes them. He uses Hemingway's trick of leaving a little nugget at the end of each session - a few words of instruction about where to go next. He writes forward, finishes the chapter, does one pass edit, then puts it away. For a hundred-thousand-word book like The Infernal Machine, this means whole sections are forgotten by the time he sits down to read the complete manuscript. He has even discovered that he duplicated entire riffs because he forgot he had already written them. But the trade-off is worth it: he can read his own book with something approaching a reader's fresh eyes.

He changes the typeface, reads on an iPad instead of at the desk, sits in a chair instead of at the writing station. Every shift fights the deadening effect of overfamiliarity.

Structure That Disappears

Johnson's proudest structural achievement is one nobody has ever noticed. The Ghost Map follows a cholera outbreak in 1854 London over about two weeks. Each chapter covers a single day. But he also assigned each chapter a thematic riff that mapped naturally to the events of that day. Day one introduces the world of scavengers and waste recycling because the scene is being set. Day two introduces public health because that is where the narrative logically turns. The structure was meticulous and deliberate and invisible. The book was extensively reviewed with wonderful notices. No reviewer ever mentioned the day-by-chapter architecture.

"When structure works, you don't notice it." It is like door design: a flat plate tells you to push, a handle tells you to pull. When the affordance is right, you move through without thinking. When it is wrong - like the Norman door at his hotel that he was increasingly angry about after twenty-five encounters - you fight it every time.

Finding structure is the hardest part. For Where Good Ideas Come From, Johnson initially planned chapters that scaled up: brain science, then workspace, then office, then city. But each chapter got siloed. The brain chapter was just neuroscience, the cities chapter was just Jane Jacobs. He finally used Scrivener's notecard interface to discover the patterns - slow hunches, liquid networks, adjacent possible, serendipity - that became the book's backbone.

The Paragraph Thesaurus

Johnson confesses that his natural skill is architecture, not sentences. His default setting produces intelligible prose, clear at the sentence level, adventurous at the structural level. He sometimes finds his sentences flat - especially in books like Enemy of All Mankind and The Infernal Machine where he wants a novelistic feel.

He built a tool he calls the paragraph thesaurus: feed in a paragraph, keep the meaning the same, but rewrite with different metaphors and rhythms. The point is not to use the output directly. It is to escape habitual patterns. He tested it on a bland paragraph about Hawaii's climate. The AI returned something absurd: "The climate of Hawaii is like a symphony conducted by Mother Nature." He typed back: "Dude, that is a little over the top." The model apologized and produced a version that was perfect. The fact that "dude, that is a little over the top" is a legitimate instruction to a computer still strikes him as crazy.

Key Takeaways

• Good ideas are slow hunches. Keep a spark file and review it regularly to let fragments collide.
• Structure can precede content by years. Write down architectural ideas even when you have nothing to put inside them.
• Use AI for research, not for drafting. Speed up the hunt for facts so you spend more time thinking.
• Never reread chapters while writing them. Overfamiliarity destroys your ability to read like a reader.
• Surprise is structural. Plan your surprises before you write, at both the sentence and book level.
• When structure works, nobody notices it. That invisibility is the proof of success.
• Let structure emerge from material. Cluster fragments until piles form, then recognize the piles as chapters.

Johnson is the rare writer who builds the tools he advocates. NotebookLM is not a product he endorses. It is a product he designed based on thirty years of thinking about how writers think. That makes his advice unusually credible: he has tested it against his own practice.

His deepest theme across all his books, narrative and idea-driven alike, is what he calls thinking across disciplines and thinking across scales. To understand anarchist bombers, you need the chemistry of dynamite, the economics of dynamite, the political philosophy of anarchism, and the specific actors in their specific world. You need to travel sixty years back to Alfred Nobel and four hundred years back to understand Indian wealth. When the page-turner surprise and the deep-forces surprise arrive simultaneously, that is Johnson's sweet spot.

This post draws from Johnson's appearance on How I Write, his Adjacent Possible Substack, and his Google Workspace interview. Athens is an AI writing editor that keeps you in flow - you write the draft, and AI suggests edits as inline diffs you can accept or reject.