5 Rules of AI Writing
1: Assume your AI usage will be discovered.
2: Test whether you’re going too far by asking: Would it be wrong to accept this help from a human without crediting them?
3: Know that, when AI writes for you, it deletes your future wisdom.
4: Never expect anyone to read your outsourced AI answers.
5: If you—or the reader—would care who the named author is, don’t generate.
It’s an intellectual vanity to claim you can always identify AI writing—that chatbots spit out nothing but sanitized blah. Yes, the humans who talk most with machines are better at detecting their patter. But proclaiming “You can just tell!” is a coping mechanism with an expiration date.
Actually, you can tell. Because AI writes better than nearly every human alive, explaining with clarity, and allowing the curious to bypass the impenetrable prose with which the self-important sandbag outsiders. In many respects, AI writing is a boon to human understanding.
Weird effects abound too.
Bureaucrats must wade through floods of oddly articulate petitions, legal writs, and complaints, as the public’s literary servants overwhelm the public servants. Police can spam the spammers, able to shovel AI slop into criminal-hackers’ chatrooms, much to the ire of the crooks. Meantime, the best man’s tipsy speech at the wedding and his sober words at a funeral—they’re so emotive you know a machine wrote them.
That’s not to mention the high-profile wins for literary AI, such as the prize-winning short story, cited for its author’s “melodic voice,” which earned a different review from the detection software: “100% AI-generated.” Or the (somewhat) non-fiction book The Future of Truth, about how AI undermines factuality, and included AI-generated quotes. Not to mention social media, glutted with even more vapid junk than when humans alone monopolized that genre.
In the 19th century, theologians spoke of “God of the gaps”: that the holy spirit reigned where science faltered. Only, science kept expanding, forcing the supernatural into retreat. The philosopher of technology Benjamin Bratton says we’re now anxiously defending a “humanism of the gaps,” defining our species by whatever machines cannot do quite yet.
But AI capabilities arrow upward in so many domains. Why should writing escape it? From journalism to fiction, more than a few prose pros are already leaning on AI writing, with its tantalizing offer to solve the blank page. Such usage remains taboo, so few admit to it. But a Nobel Prize winner just did, the Polish novelist Olga Tokarczuk, who seeks creative pointers from AI.
“I often throw an idea to the machine for analysis, asking, ‘Honey, how could we develop this beautifully?’ ” she said at a recent onstage event.
“At the same time, I feel a poignant, very human sorrow for an era that is disappearing forever. My heart aches for the passing of traditional literature, written over months in solitude, a work of life crafted in the mind of a fully conscious, single individual.”
Personally, I have devoted my adult life to writing, publishing five volumes of literary fiction, a nonfiction book that I ghostwrote, and scores of newspaper articles—years of piling up sentences such that an editor might buy them on a computer file. But it dawned on me: the computer, not my files, is what enchanted the world. So I pivoted to the topic of the times, AI policy, the many-handed effort to push an elephant.
Which is to say, I worry about the fate of human writing. And yet something odd keeps happening.
The instant I discover that what I’m reading—perhaps rather interesting—is AI-generated, I suddenly don’t care. The words dim before me. It’s an off-switch.
Why?
THE PAINTER MADE OF PIXELS
A conceptual artist named SHL0MS posted a picture on X, and asked readers to say why the image was worse than a real Monet.
Nobody spared the machine’s feelings.
“I’m disappointed I have to even point it out,” one person commented. “The background lilypad-algae amalgam is egregiously vague, like most AI art.”
Another added, “Doesn’t look anywhere near like a Monet. Looks exactly like somebody trying to replicate the style and achieving like 20% of it.”
In fact, the painting was a Monet, and the post was a prank that, seemingly, exposed the pretensions of people who flock to galleries, claiming to admire the art but really admiring status, entering each room and hurrying to the wall text: Who is this? Should I care?
That seems shallow. But the same drive could save human writing.
After all, you should care more about a real Monet than a dupe. Not because one is objectively better, but because its meaning exists in a matrix of social beliefs about beauty, about value, about the shared tales of human civilization: that an irascible Frenchman once swiped hog bristle across canvas, his perceptions and drives filtered into a decorative depiction of a pond, conveyed from dealer to collector, rising the ranks of culture, finding a lonely museum wall, venerated there for years, as the outside world transformed, the artist died, and pixels reproduced his perceptions and drives, connecting your sensing brain to the sensing brain of a particular man who will see nothing ever again.
All that is only in our heads. But where else is meaning but in heads? And writing is humanity’s most sophisticated technology for visiting another’s head.
The problem is that writing is hard. Words elude you; paper doesn’t care. George Orwell equated writing a book to suffering from a long illness. Developers dream that AI will someday cure all illnesses. Why not this one?
WHY (A)I WRITE
In his essay “Why I Write,” Orwell listed four main motives: 1) sheer egoism; 2) aesthetic pleasure; 3) the urge to say what is; and 4) the urge to change what is. Others cite a fifth reason: to know what you think.
Those who don’t write presume the process involves hatching an idea, then putting it into words. It’s commonly the reverse: put down words to hatch the ideas within them.
This points to a flashing peril from humans impersonating themselves via AI writers: cognitive surrender, that we hand over the wearisome word toil, and atrophy in genteel luxury, never enduring the mental frictions that kindle into wisdom. We wouldn’t detect the decline, I suspect, half-believing ourselves the authors still.
As a leading AI thinker told me in private, “I write so I can keep writing in the future, so I can keep thinking in the future.”
But there’s another reason that people write, a motive far more powerful than humankind’s dwindling determination to remain the brightest species. It’s ego. As Joan Didion observed, “Writing is the act of saying I, of imposing oneself upon other people, of saying listen to me, see it my way, change your mind.”
Every person craves attention, and the nerds buy it with words. Readers possess a natural resource—attention—that is everywhere and nowhere to be found. A transaction ensues: reader agrees to spend attention on words in exchange for information (nonfiction) and/or experiences (fiction). The writers’ art is to smuggle themselves into both.
So, when reading human, you’re absorbing the meaning with one eye while focusing the other on its maker, ever trying to reconcile the two, much as the visual cortex creates 3-D from the parallax of two eyes synthesizing one object at dueling angles. If you see that the author was AI, one eye closes. The page becomes markings on a flat surface.
The reader’s presumption of a human behind the words makes secret AI authorship worse than betrayal. It’s trespassing and fraud at once, entering another’s brain on false premises and altering the contents for gain, whether it’s AI-generated reviews that trick you into feeding at someone’s restaurant, or an AI-generated marriage proposal that tricks you into feeding someone’s children. It’s the sale of a shared reality, from an entity that has no reality to share.
That said, much text is purely for dispersing information, and the author is irrelevant. This is true for most legal documents, press releases, and boilerplate emails, not to mention explanations that you prompt a chatbot to conjure. AI composition can be helpful and harmless, provided it’s honest.
THE UNWRITTEN RULES
But nobody knows the ethics yet. Hidden usage seems wrong because deceit is wrong. But must you specify authorship in every case, even if it’s a casual note that you autocomplete? And if you’re candid about it, does that make AI writing fine?
News publications ban it, educators resist it, and academic journals flounder. Yet social norms could be shifting subtly toward acceptance. Increasingly, you hear the view that authorship is immaterial; it’s about making something worth reading. “Honest to god,” wrote Tobias Schneider, a research fellow at the Global Public Policy Institute in Berlin, “what do I care how the text for an academic paper comes about. It could be a monkey throwing darts at a board, if the resulting combination of letters conveys original insight, it should be published.”
As with most AI anxiety, it’s a matter of speed: this is moving faster than the culture, leaving us in a perma-panic as one train after another hurtles through the station, heading to destinations unknown with more of our belongings aboard.
Already, large-language-model vocab is springing more frequently from our mouths, according to one study. Authors, in dread of false accusation, are altering their prose to avoid seeming LLMish. I know because I’m doing so, tense about my appreciation for bullet points and curbing the dashes that I learned to overuse in old-fashioned newsrooms—for melodramatic punch.
Errors have become fashionable too. Once, clean copy was a hallmark of professionalism. But “the house style in most newsrooms is extremely LLMable,” the Atlantic magazine contributor Jasmine Sun noted. “What stands out (besides reporting) is a distinct and authentic first-person voice, even if that means the occasional typo.”
Not for nothing is wabi-sabi—the Japanese tradition of charming imperfections—trending. To err is human. To not err is AI.
For those who venerate human writing, the nightmare is that AI authorship becomes so commonplace that few care about origin anymore. Nobody would need to lie about it anymore. Looking at recent history, technological tools for knowledge work have tended to draw contempt at first, only to gain wide adoption. Many researchers shunned internet sources back in the 1990s, while journalists of the early 2000s scorned Wikipedia. But the technology improved, and norms updated.
If AI writing resolves its quirks, will the same acquiescence follow?
3 FACTORS THAT COULD SAVE HUMAN WRITING
Here’s what may decide whether most people still care about people writing: 1) institutions clarifying the rules and norms of AI usage; 2) reliable detection technology becoming widespread; 3) people retaining the “off-switch” aversion to AI authorship:
Establishing Rules
When it comes to AI-generated text, the closest norms regard plagiarism, which also involves the presentation of others’ work as one’s own. But the concept of plagiarism arose to protect the intellectual property of a fellow human. With AI writing, who is the victim?
If writing is dishonestly presented, one victim is plain: whoever is duped into reading it. This suggests that openly AI-generated text is fine. But victims may exist in those cases too.
Among the harmed could be people who fail to write, and therefore never develop their minds. Also, there are the human authors who’d need to vie against hordes of fluent bots hogging public attention. Most of all, society could suffer from the decline of collective cognition, if writing is no longer a push-and-pull involving thinkers living and dead, but just effortless insta-words materializing onscreen.
Yet AI writing is immensely useful, and stimulating, and will stir human creativity. We must set rules and norms for concrete cases, not just shout, “Don’t!”
Is it wrong for an AI to write an action plan for a business client? What about a write-up of the family vacation to share with loved ones? How about assigning AI agents to write your social-media feed?
The debate over AI writing often misses an even more likely outcome: assisted writing. What if you come up with all the ideas, but an AI turns them into clearer prose than you could, which you then tweak? Or reverse that: the AI brainstorms, but you write the prose. Which is worse, and why?
Recent studies suggest that AI writing may neutralize the human user’s arguments, and that essays drafted by people contain a range of ideas, whereas AI essays tend to converge around similar points.
On the other hand, chatbots—by disseminating ideas so much more gamely than unmoving written texts—may stir our creativity. In theory, future personalized AI models could allow users to vary the “temperature” (randomness) of LLM responses, increasing their novelty.
But without social norms, dystopian futures come readily to mind, with humans generating AI text for other humans, who employ AI to read it and craft replies, downgrading our species’ intellectual conversation into a feedback loop of artifice, in which we are mindless mouthpieces of literate machines.
That feels far-fetched today. Yet some people are already cramming past emails and blogs and tweets into AI, training agentic systems to capture their “voice,” such that they needn’t cultivate one.
If leading institutions care to keep human writing, they must establish clear and coordinated policies in haste, so people know the professional and social costs of their choices. This depends on detecting what is human and what is machine.
Detection Technology
Disgrace is an enforcer. Soon after the explosion of chatbots in late 2022, nobody could reliably identify whether text had been written or generated. But AI-writing detection software—notably, the Pangram app—has become strong.
You can tell because it keeps precipitating scandals. But institutions have yet to apply detection tools as widely as they should. That would mean standardized detection everywhere we care about human input and honesty: job applications, dating apps, any publication’s vetting process.
For detection apps, the challenge is twofold: Do not mistake AI writing for human (a false negative); and Do not mistake human writing for AI (a false positive). The first error makes the app worthless, but the second error is especially dangerous. A wrongful accusation—whether against an author, a student, or a public figure—could destroy a career.
Last year, the University of Chicago economists Brian Jabarian and Alex Imas published an evaluation of several detection apps, having had them evaluate around 2,000 human-written passages, including fiction and non-fiction from before 2020 (to avoid texts that were stealth AI), along with the same number of AI-generated passages.
“The results are clear-cut,” Jabarian and Imas wrote. “On medium-length to long passages, Pangram achieves essentially zero FPRs [False Positive Rates] and FNRs [False Negative Rates] within our sample.” The error rate did increase slightly with short passages, under 50 words. (For reference, this paragraph is exactly 50 words long.)
The accused still deserve a fair hearing. Institutions should establish systems of appeal, while independent evaluators must keep testing the accuracy of detection tools. This poses an uncomfortable question: what rate of false accusations is acceptable? The reflexive answer is, “None!” But we accept a degree of error in matters that are even more impactful, such as the justice system. Society considers it better to identify many culprits than to preclude any miscarriage.
Benjamin Franklin preferred that 100 guilty go free than 1 innocent suffer, while the legal profession debates “Blackstone’s ratio,” which proposes 10:1. Developers of detection apps likewise make such moral calculations, with Pangram training its models to average no more wrongful accusations than 1 in 10,000.
But AI developers are pursuing ever-better writing tools, which could raise the opposite problem: increasingly mistaking AI writing for human. In which case, we’ll need further antibodies.
The Off-Switch
“More than 300 years after the reading revolution ushered in a new era of human knowledge, books are dying,” warns the leading chronicler of literary decline, talented young fogey James Marriott. He marshals a range of sorrowful datapoints, from the collapse of reading-for-pleasure, to the inability of college kids to handle any textual analysis more cognitively vexing than Instagram.
If people quit reading, it’s moot who (or what) writes.
But a paradox is that our purportedly post-literate era is the most literate in history, if you judge by how much people are producing and consuming words: all the social-media posts, all the texting back and forth, all the newsletters—not to mention the word-streams on a billion podcasts and video clips.
A future without writing—as opposed to today’s decline of deep reading—is possible only if we find a technology more proficient at transmitting ideas and experiences from brain to brain. Brain-computer interfaces are not near any such text-ending magic.
Until then, human reading contains the “off-switch” antibody: that you just don’t care as much once you discover that nobody said it. It’s a psychological reason for why writing belongs in what the economist Alex Imas calls the relational sector: “the human-intensive, provenance-rich, sometimes artisanal part of the economy where the human aspect is part of the value of the good or service itself.”
Indeed, Imas contends that excellent human writing may rise in value in the AI era, partly because so few people will be able to do it. In other words, the post-literate era of tongue-tied semi-literates could make the expressive few stand out. One caveat: if people develop relationships with AI companions, they may care deeply about what those entities write.
THE HUMANS ARE DEAD. LONG LIVE THE HUMANS.
A classic twist in sci-fi horror is when the main character finally twigs: all the other “people” are actually machines. Lately, I’ve felt like that character, discovering when I drop more and more texts into the AI detector that robots are everywhere.
Not only that literary award-winner, but the award citation itself. Also, an essay—on the importance of reading!—by a thinker I admire(d). Then, another piece of writing landed on my desk. Not onscreen. It landed physically on my desk: a handwritten letter, composed in pen, with “Tom” on the front of the envelope.
I have it before me now, saved as no email would be, because of the fellowship of human beings, that we mind about words—but also about another person’s effort and intent. Machines are so eloquent. But until they struggle to find the words, they won’t compose a letter I’d keep.
Human writings are an improved version of human beings. By escaping the confines of your skull, by pausing time, by revising what you’d blurt, you craft what you mean to mean.
“What I write about is other than me. As what I write is smarter than I am. Because I can rewrite it,” Susan Sontag said. “My books know what I once knew, fitfully, intermittently.”
Writing is an intelligence-increasing technology, allowing you to convert the SOS pulses of the perceiving self into objects to organize, line up, accumulate, frame, and build into more than the capacity of your known mind. For the species, it’s an even more IQ-increasing technology, turning thoughts into a public good that never expires.
AI writing may burn books that don’t yet exist. Perhaps that’s fine. Publishers are always moaning that too many books come out. And maybe we’ve reached the end of our utility as authors: we wrote superb datasets called literature; the machines enjoyed that appetizer.
But our words are more than their feed. Writing is our testament, permitting readers to find companions among the dead, and the author to befriend the yet-to-be-born, granting anyone a way to defect from the present, and providing even the godless with a faith that something may persist after life.
We—briefly thinking machines, stuffed with noisy training data, false memories and nostalgia, jostled by fluke and circumstance, each generating words that nobody else would—we are the worst writers, and the only writers I care to read.
My thanks to those whose ideas and comments bettered my writing, including Alex Imas, Benjamin Bratton, Alison Snyder, Séb Krier, Arthur Goemans, Max Spero & Conor Griffin.
Click Our Rules for AI Writing for how AI Policy Perspectives is approaching this issue—and add your thoughts too. You’ll find more about: The Embarrassment Rule, The Intern Rule, The Ventriloquist’s Rule, The No-Dumping Rule, and the Byline Rule…
This piece was so on-point it deserved a real, live, human-generated comment. I've been living with this question for some time--what does disclosure look like for writing, or any creative work, in the AI era? I'm in the process of setting some rules around this, with the intention of preserving uniquely human cognition while managing the realities of a world trained on synthesizing content at the speed of light. This certainly adds to my thinking!
It’s best to bring the idea and logic of the piece, before using any AI LLMs. Then use the AI Model as a brainstorming partner or tool to nail the final product.