When faced with a blank page and no ideas to fill it, fall back on structure.
I am at my day job, editing 800 words of website content transparently written by artificial intelligence and submitted for publication by a credentialed (human) professional. It has all the hallmarks of large language model text generation — from the em dashes to the rhetorical contrasts, with a habit of over-using bold text for emphasis, rhetorical contrasts, and lists of three.
How do I start making this bad AI writing seem less like bad AI writing? I use AI, of course. The prompt I typically give my preferred model is involved, and includes samples of other web pages for reference as well as a stylebook, tips on institutional voice, and directions regarding the content's intended audience and goals. The prompt also makes it explicit: Do not write like an AI; write this like you are a real person. That's usually enough for structure.
There's work to be done that doesn't involve artificial intelligence. I fact-check and I revise and I revise again. I make a call or send an email to confirm this thing or add another thing. And in the end, the story is OK. Better than OK. In this type of work, there are clear expectations for what language should do and what it can accomplish. Once the structure is in place, it becomes a matter of following a formula. The stories tell themselves.
If language is how we make sense of the world and ourselves, trying to impose some scaffolding on our experience of reality, then increasingly sophisticated artificial intelligence is remaking us in its own image. This is a future that will be familiar to any other hacks like me working in the buzzwordy space of copywriting, content marketing, and communications. Imagine the most hollow and trite motivational corporate LinkedIn post you've ever seen, stomping all over your feed, forever. To preserve anything like truth or beauty in the language we are immersed in every day, we need to break with structure for moments of chaotic inspiration. As with everything, that's easier to say than it is to do.
I fell into my career, an overly dignified word for the series of stumbling decisions I've made about how I spend my weekdays in front of a computer, out of a mix of cowardice and stupidity. I always wanted to be a writer. No young man dreams of writing website material for real estate companies, or even of filing 300 words of news copy on school board meetings, but I lacked the self-confidence to see myself making a living as a writer or teacher of creative fiction. I also lacked the wisdom to listen to the many professors who warned me all the way back in 2008 that the news business was bottoming out. Still, journalism seemed like a craft, a trade I could learn and ply to pay rent while I wrote bad poetry and worse novels on nights and weekends.
I wasn't totally wrong. There were rules to reporting and news writing, and I was smart enough to follow them. Along with a willingness to swallow my anxiety about cold-calling people who didn't want to talk to me, that turned out to be enough to impress some editors. I learned by reading, mostly.
I read a lot of bad writing too, which is instructive. You develop, slowly and without realizing it, a sense for the moment a sentence goes limp. When I was covering city council meetings for $12.50 an hour, I kept a list of phrases I was not allowed to use. The list never stopped growing. Today, that list would include words like "delve" and all the transitional phrases favored by AI.
This may seem obvious, but it's the only advice I've ever trusted myself to give to younger writers. Read everything that's out there: news stories, magazine features, novels, nonfiction, advertisements, pamphlets at the doctor's office, billboards, online forums, new books, old books, whatever. The larger the dataset, the better. Read enough, and the rules get to be second nature. It becomes a matter of probability distribution. Write one particular word, and another particular word is likely to follow it, and another word is likely to follow those two words together, and so on until you have a sentence and a paragraph and a full story that fits within a familiar structure. You don't have to think about it. The words sound right because you have encountered them within this structure before, a million times.
This is also basically, in my dumb guy understanding, how large language models learn. Which doesn't quite explain why the vast majority of AI writing is so lousy, in a manner that is distinct from the lousiness of lousy human writing. AI writing, especially the newer models that have ironed out many of the hallucinations that used to populate answers from ChatGPT, is bland and predictable because it's been trained to be that way, a result of capital incentives and corporate structures that extract more value out of an impersonal, superficially smooth voice.
These incentives have always shaped writing, especially the writing that people do for money. Magazine writers like to talk about "voice," specifically developing your own, and it's true that the best of them have a uniquely identifiable way of writing. Your voice has to cooperate within the typical structure of what we recognize as a feature story, and within the bounds of what an editor and the magazine's salespeople respectively consider publishable in an industry in freefall. But it's still your voice. I tried to come up with a voice of my own, in my very provincial, small way.
After I washed out of journalism and started working full-time in marketing and corporate communications, I became more chameleonic, and more cognizant of the practical incentives behind the language we use on the internet: SEO key terms to drive Google traffic, calls to action to like and subscribe and buy, etc. Every brand and executive leader wants their own "voice," with their own ideas about what that should be, and in this line of work you do your best to square all these interests against your own ideas of what constitutes "good" writing within the various structures provided by the field of digital communications.
AI is changing the tenor of all this language. And most people are OK with the internet's newly dominant house style. The hallmarks of AI writing have infiltrated student essays, social media posts, speeches in the British Parliament. My advice that writers should "read everything" is more suspect as the influence of AI language is making our own language — once bad in unpredictable and occasionally enlivening ways — more uniformly uninteresting.
Of course, not all AI writing is created equal. With the right inputs and revisions, a large language model can be coached into providing totally serviceable marketing copy, institutional language, annual report summaries, website descriptions, executive messaging, all the "content" that keeps people like me employed. It can synthesize tons of data, organize pages of research, and find your typos. I even believe that the Fortune journalist using AI to turn over dozens of stories a day is doing solid and dependable work within the fairly regimented structure of daily news coverage.
Beyond my world of hackdom, in blind tests of literary (and not so literary) fiction, people often prefer the creative writing of AI. There's never been much accounting for taste, and I'm still inclined to agree with the literary writers who have given this some thought and concluded that AI is never going to make great art. We can tap into something deeper that AI can't.
In the last section of his book The MANIAC, Benjamin Labatut gets at this something deeper. He recounts the famous 2016 match of Go, an ancient and extremely complex board game that Labatut likens to an art form, played between a top-ranked human player named Lee Sedol and the computer program AlphaGo. Considered an early indicator that artificial intelligence was getting good enough to be scary, AlphaGo won four of five games in fairly dominant fashion. Sedol stole one game by creating chaos from a move that seemingly came out of nowhere, bewildering the AI that had played millions of games against itself to map out strategic patterns human players could only dimly perceive. Sedol's embrace of chaos made no rational sense to the computer, or to anyone else watching the match, which is why it worked. It was called a "divine move" by one observer. Sedol credited "sheer inspiration."
The poet and novelist Ben Lerner, who has apparently prompted Claude AI to do a decent Keats imitation, conceded to an interviewer that AI writing is good enough to pass the Pepsi challenge. "But…there's some kind of miracle of transmission that happens when a human encodes their voice and heartbeat in a sentence. And that's not something AI does."
So this is all well and good for the board gamers and artists, the poets and novelists who can get in touch with the divine, channeling the strange energies of the universe in their valiant efforts to capture the ineffable. What about us hacks? As long as a human is still required to turn bad AI writing into good AI writing, perhaps we should worry less about the computers taking our jobs and more about the computers totally alienating us from language itself. What happens when AI has wormed its way into every sentence we write?
"Go is a work of art made by two people," Sedol said after his losing Go match. "Now it's totally different. After the advent of AI, the concept of Go itself has changed."
What I do for a living is not art. Far from it. But language always has power, in any medium. Language can let a little bit of beauty slip into the world in the most unexpected places. Writing, even writing some lame marketing copy or dropping 140 characters on social media, can still occasionally feel like a "miracle of transmission."
Prompting an AI feels like something else. With AI writing, there is structure, and there can even be voice, but there is no inspiration. Nothing is transmitted. That feels true for the writer, at least. It's less clear whether that feels true for the reader. If I told you that one paragraph in this little essay was generated by AI, would it make a difference?
The concept of writing has changed. We are changing with it, whether or not we notice. If we're going to find a way to preserve the potential for authenticity in everyday written communication, the distant possibility that the words in an email or a brochure or a college admissions essay will actually say something meaningful and original, we can't fall back on structure.
I think many of us have always grappled with this dilemma, the gap between the regard we have for the almost holy power of the written word and the grubby reality of advertising and institutional communications. AI has made the gap more pronounced. We're not artists. This is work. Us hacks are in trouble.