Why Can't AI Write Copy?
An investigation. A rumination. An obituary.
In 2019, JPMorgan Chase signed a five-year, enterprise deal with a New York startup called Persado. The pitch was simple. Persado’s AI could write marketing copy that outperformed humans, and Chase had three years of pilot data to “prove” it. In tests across the bank’s card and mortgage businesses, Persado’s machine-generated ads produced click-through rates up to 450% higher than copy written by Chase’s own marketing department. The case study that made the rounds was a head-to-head on home equity lines of credit. Behold!
The human team wrote:
“Access cash from the equity in your home.”
But Persado’s algorithm wrote such poetry as:
“It’s true — You can unlock cash from the equity in your home.”
JPMorgan’s CMO at the time, Kristin Lemkau, called this “incredibly promising.” She also was either so hoodwinked by Persado, or so scared of her own job security, that she made the mistake of saying the quiet part out loud: “Machine learning is the path to more humanity in marketing.” This is the person deciding what good communication looks like at the largest bank in America? Nope. Not anymore. She’s now the CEO of JP Morgan Wealth Management. She got promoted.
A few variables are worth exploring here. First of all, this study is valuing effective copywriting only through clicks. A click is not persuasion. A click is curiosity at best and muscle memory at worst. It measures whether someone's thumb moved, not whether their mind changed. By this metric, a typo in a subject line that makes someone think they got a bank alert would outperform Hemingway or Shakespeare. Clicks are the lifeblood of performance marketing, and performance marketing is not actually marketing. It’s sales. Digital direct mail. Copywriting for such ads is not really copywriting. It’s math. And in this sense, sure, Persado might have a leg up on us mortals. But when it comes to AI writing anything that makes us think, makes us feel, or convinces us to change our minds about a product, person, or platform, I’ve yet to be impressed.
More importantly, copywriting itself is an art and a science that requires study. There are experts and there are charlatans, and most are the latter. I wonder which humans wrote the copy in this test in the first place? It certainly was not a team of the world’s best, as they would’ve known that putting the “It’s true!” at the beginning of the statement is a classic, proven, tried-and-true copywriting strategy. The technique is called a "truth claim" or a "belief opener." More precisely in the craft, it's an example of leading with validation before the ask. The Gary Halbert types would call it a "readiness frame." You soften the reader's skepticism before you make the proposition by pre-agreeing with them: Yes, this is real. No, you're not being tricked. A tactic as old as the written word itself. The human version from Persado’s case study is flat, transactional, and forgettable. The AI version is marginally better, the way a splinter is better than a paper cut.
But here’s what nobody in the breathless coverage of this deal bothered to assess: Persado didn’t beat an actual pro-level copywriter. It beat a mortgage marketing department. It beat the kind of copy that gets written at 4pm on a Thursday by someone who has 19 other deliverables due, a hangover, and a compliance review in half an hour. The bar it cleared was so low you'd need sonar to find it. Pit Persado against Dan Kennedy or David Ogilvy and I can guarantee you the machine would not win.
And yet, six years later, Persado has won. It’s booked over $2 billion in “incremental revenue” (how do they define this catchall term? unclear) for its financial services clients. It won a 2025 Tearsheet award for “Best Product Designed for Gen Z,” which, what the fuck? Persado is better than TikTok or Depop? It has helped the AI copywriting industry metastasize into a multi-billion-dollar ecosystem of tools, templates, and frameworks, all promising to replace the expensive, subjective, frustratingly human process of writing words that sell things with…more clicks.
Before the Word
For at least 30,000 years before a human wrote anything down, we were making images. Cave paintings at Chauvet in southeastern France date to roughly 36,000 years ago. Lascaux, 17,000. The primeval artists who made them sourced pigments from up to 155 miles away, built fat-burning lamps to work in total darkness, and used the flicker of firelight to create the illusion of animals in motion. Picasso once visited Lascaux and reportedly said “we have invented nothing.” Every child can draw something, at least once, that’s remarkably moving. Yet most people live their whole lives without being able to put a coherent sentence together on paper.
When writing finally appeared, around 3200 BC in the Sumerian city of Uruk, it had nothing to do with stories, poetry, philosophy, or self-expression. The earliest written tablets are inventories. Grain shipments. Livestock counts. Beer rations. The occasional customer complaint. Approximately 85% of all proto-cuneiform texts recovered are administrative bookkeeping. Writing was created to serve ancient middle management. Cities got too big, transactions got too complex, and memory could no longer hold the ledger. It took 600 years for the Sumerians to attach sounds to these symbols, which finally allowed writing to represent spoken language. Then it took another few centuries before anyone thought to use this accounting technology to create art. The first known author in human history was a priestess named Enheduanna, and she didn’t come around until roughly 2300 BC.
Soon thereafter, writing was being used mainly to sell things. Egyptian merchants hawked papyrus handbills to demonstrate their UVPs. In ancient Rome, politicians hired professional sign-painters to write campaign slogans in red and black paint on the facades of wealthy homes. Over 3,000 electoral inscriptions have been recovered from Pompeii alone, including endorsement deals, competitive slander, and what can only be described as attack ads in the form of graffiti. Enemies of a candidate named Vatia plastered the city with copy like “The petty thieves ask you to elect Vatia as aedile” and “All the drunkards ask you to elect Vatia.” Not bad.
Persuasion through language predates literacy itself. People were selling with words before anyone could read them. The spoken pitch is older than the written sentence. And the craft of choosing which words will move another person to act, which is all that copywriting really is, is as old as human social life itself. It has never been a problem of information or automation. It has always been a problem of nerve. Which brings us back to the machines.
Frameworks All the Way Down
Visit the websites of the major AI copywriting platforms and you will encounter a universe of mediocre nonsense. Jasper, which claims to have “read 10% of the internet” (how would this make you a better writer?) offers over 50 templates organized by advertising frameworks like AIDA (Attention, Interest, Desire, Action), PAS (Problem, Agitate, Solution), “Emotion-Logic,” “Picture-Promise-Prove-Push,” “Awareness-Comprehension-Conviction-Action,” and many more. You input your product, your customer persona, and your unique selling proposition. You select a tone of voice (they suggest “Tony Robbins” as an option). You click generate.
So I asked Jasper to write copy for its own product using the AIDA framework. Here’s the beautiful prose it spit back out:
“Attention: Do you hate writing copy for your marketing campaigns? Interest: We know that not everyone is a great writer, but we also know that you can’t afford to have bad content on your website. That’s why Jasper.ai was created — so you don’t have to worry about it anymore! Desire: Now with just one click of a button, Jasper will create the perfect piece of content for any campaign or promotion you’re running. Action: Click here right now and sign up for a free trial of Jasper today!”
This is the product demonstrating itself. The sample on the showroom floor. And it reads like a hostage letter written by someone who learned English from infomercials. Ignoring the AI giveaways, every sentence performs the mechanical steps of persuasion while achieving none of the emotional effects. It is copy that sounds like copy to someone who has never actually read good copy.
Copy.ai, another competitor, pushes their messaging into even more absurd territory. The brand pivoted from its original copywriting premise to calling itself “the first-ever GTM AI platform,” which is the kind of phrase that tells you everything about what a company wants to be and nothing about what it actually does. Our hero Persado describes itself as “reinventing marketing creative by applying mathematical certainty to words, the foundational DNA of Marketing.” They throw around the term “Motivational AI.” Capital-M Marketing. Mathematical certainty. Applied to words. Ok.
The copy that AI copywriting companies use to promote AI copywriting is the most damning evidence against AI copywriting. They can write, but they can’t sell, even when it comes to their own product. To understand why, you have to appreciate what a large language model actually does.
Mongoose
The people who build AI models don’t fully comprehend how they work, if they can actually “think,” or what goes on inside their digital “minds.” Mikhail Belkin, a computer scientist at UC San Diego, said in a 2024 MIT Technology Review piece:
"Our theoretical analysis is so far off what these models can do. Like, why can they learn language? I think this is very mysterious."
That’s the real question. Can they learn language? Or are they performing a very impressive parlor trick? Anthropic’s Dario Amodei put it even more plainly in his excellent 2025 essay “The Urgency of Interpretability”:
“People outside the field are often surprised and alarmed to learn that we do not understand how our own AI creations work. They are right to be concerned: this lack of understanding is essentially unprecedented in the history of technology.”
In theory, LLMs predict the next most likely token in a sequence. Given “Access cash from the equity in your…” the model assigns probability weights to every possible next word and selects from the highest-weighted candidates. “Home” scores high. “Dreams” scores low. “Mongoose” scores essentially zero, I’d imagine. The architecture of LLM is an engine of likelihood. It is not thinking as much as guessing. It generates language by asking, thousands of times per second, what word would most likely come next? It’s advanced probability theory. It’s not, however, doing anything close to what people do when we write.
Consider some of the most iconic and effective lines in the history of advertising, and how they all contain some kind of tension that would be difficult to predict if you were a math-driven model writing from scratch:
“Think Different.” Grammatically incorrect. The adjective “different” where the adverb “differently” should be. Apple’s agency, Chiat Day, chose the wrong word on purpose. The friction here is the message. It tells you that this company does not follow the rules. It reinforces the line itself. A probability engine would likely edit it, because models are trained on grammatically accurate text and reinforced toward correctness by human evaluators.
“Murder Your Thirst.” Insane. Nonsensical. The tagline takes the most benign consumer act imaginable, drinking water, and casts it as homicide. It works precisely because it shouldn’t, and because it’s wholly unique in the category. An LLM asked to write taglines for a water brand would produce variations on purity, refreshment, and hydration. The semantic field the training data points toward. It would likely never arrive at violence as a metaphor for drinking, because the leap requires a willful disregard for category convention that probability cannot replicate.
“Got Milk?” Incomplete. No subject. No verb conjugation. Two syllables. The California Milk Processor Board’s agency, Goodby Silverstein & Partners, built a $100 million campaign on a sentence fragment. The imperfection creates an accusatory intimacy, like someone checking your fridge at 11pm with a plate full of warm cookies. It feels human, and is one of the most effective taglines ever written because of it.
“Where’s the Beef?” Wendy’s, 1984. A Fortune 500 fast food corporation adopting the register of an irritated elderly woman examining a competitor’s hamburger. The tonal shift from corporate to colloquial is so drastic that it became a national catchphrase and even entered a presidential debate. The training data for QSR advertising does not contain the speech patterns of frustrated grandmothers, and so LLMs would likely not be able to come up with something like this. Like Liquid Death, the phrasing is both deeply realistic and exceptionally bizarre.
“Just Do It.” The most famous line ever. A contextless imperative from a company that sells shoes. It does not mention shoes. It does not mention sport. It does not mention Nike. It is three words of pure command that assume you already know what “it” refers to, and more importantly, that you’ve been failing to do it. It was famously inspired by the last words of a murderer about to be executed by firing squad in Utah in 1977 (he said “Let’s do it,” but close enough). The line works because it meets you in your hesitation, your excuse-making, your cowardice. It addresses the Hamlet in all of us. The 20th Century default human condition of inaction, only exacerbated over the last 25 years. It is timeless. It is perfect. Persado could never.
Every one of these lines is a statistical anomaly. They succeed because the writer chose the word that a probability engine would not. These are stylistic choices, but they’re also deeply human. Structural violations of the logic by which language models operate. This kind of approach is not a bug in the copywriting process. It is the copywriting process. The line that works best is usually the line that has no business working in the first place.
The Comfort Mandate
The problem with LLMs’ ability to write with impact is the same as our own: training. A writer with raw skill can be sharpened by good mentorship, constant rejection, and a decade of killing darlings and getting fired from magazines. Contemporary language models are shaped through reinforcement learning from human feedback (RLHF). Human evaluators rate model outputs on criteria including helpfulness, harmlessness, accuracy, and clarity. Outputs that are ambiguous, confusing, or could create discomfort get penalized. The model learns, through millions of these cycles, to reduce friction. To smooth. To clarify. To complete.
But good copy lives in friction.
What’s happened to design over the last decade or so has been widely bemoaned by professionals in branding. Everything is flat. UX all looks the same because it needs to be easy. God forbid anyone needs to click more than once. Typography has, quite literally, lost its legs, even though we are now experiencing a very predictable (and wonderful!) serif revival. But the same thing that performance marketing and digital commerce did to design, it did to language too. Perhaps even more so. Let’s call this trap The Comfort Mandate.
The Comfort Mandate is not a bug that will be fixed in the next model release, nor is it a prompt engineering problem. This is the fundamental orientation of digital technology. Language models are built to be helpful. Good copy is interruptive. Language models are built to be clear. The best copy is irreducibly ambiguous. Language models are built to be comprehensive. Effective copywriting is an act of savage sketchiness.
We’re losing the cultural value of good writing, and this extends to copywriting, because it often makes us uncomfortable. The economics tell the story. Graydon Carter reportedly paid $500,000 for three long-form articles a year at Vanity Fair. That number sounds absurd now. Not because good journalism isn’t worth it, but because the market has collectively decided that it isn’t. We’ve spent two decades optimizing language for speed, scannability, and SEO. We’ve trained an entire generation of readers to skim and an entire generation of writers to let them. To get a debut novel published today, you need 100k+ social media followers for an agent to even consider you. The mean drops a little every year. AI didn’t cause any of this. It is the product of it. Models trained on the internet write like the internet because the internet is what we asked language to become. AI can write puns. It can identify wordplay. It can explain why a joke is funny, which is of course the most reliable way to make it unfunny. What it cannot do yet is feel the rhythm of a sentence the way a drummer feels a beat, knowing exactly where to place the rest, the silence, the space where the listener’s brain fills in what was left unsaid.
Machines might have the math, but they don’t have the nerve. Chase didn’t sign a five-year deal because the copy was that great. They signed it because it was cheaper and the clicks went up. The AI won not on quality but on economics.
This cycle is how most crafts end up dying a slow, meaningless death. Lampposts used to be beautiful. Now they are featureless aluminum poles with a bulb on top. The old ones were intentionally designed. A product of craft. The people who made them cared about something that didn’t need to scale. The new ones are spec’d. Mass produced. Created to be forgotten. Both light the street. Only one makes you look up. Economics wins out again.
The same thing has happened to language. Not because AI killed it, but because we’ve optimized it into a featureless grey pole. Decades of SEO, A/B testing, character limits, readability scores, and conversion rate optimization have produced a world where most commercial writing reads like it was generated by a machine long before machines could generate it. AI didn’t flatten our vocabulary. We did. LLMs just automate all the flatness.
Expensive Boxes
There are alternatives to RLHF, but none of them escape The Comfort Mandate. Anthropic’s Constitutional AI, the most sophisticated training approach currently in use, replaces human evaluators with a set of written principles that the model uses to critique its own outputs. This is basically AI training AI to be better at being AI, and although it might make your eyes roll, it seems to be creating a vastly superior product in Claude. Humans train humans. Perhaps AI should train AI. Nonetheless, the principles themselves still optimize for clarity, helpfulness, and harmlessness, and I’ve yet to see even Claude’s powerful Opus model be able to write effective copy even when custom-briefed with extensive bespoke skill documents. The rubric changed hands. But it didn’t change.
Effective copywriting comes from a deep knowledge of the product and the customer. Everything else follows. The nerve, the subtraction, the willingness to risk the wrong syntax. These are outputs that a LLM could probably be trained for. But the input is still remarkably analog.
Manhattan Mini Storage has been writing some of the best OOH copy in America for 20 years. Their team is two people. Their media budget is zero, because they own the buildings that they emblazon their ads on. Their product is storage units, the most boring category on Earth. Yet their billboards have become a cultural icon in NYC, the world’s noisiest advertising market. They’ve received front-page coverage and cease-and-desist letters, which is a better metric of effectiveness than any click-through rate Persado can report. The copy is often political, rarely makes perfect sense, and has a mean streak to it, just like New Yorkers! “Your closet’s so narrow it makes Dick Cheney look liberal.” “Does this butt make my room look small?” “Why leave a city that has six professional sports teams, and also the Mets?” Lines like these work because Archie Gottesman, the woman who writes them, knows exactly what she’s selling (expensive boxes in an expensive city) and exactly who she’s selling it to (space-deprived, politically opinionated New Yorkers with a distinctly local sense of humor). The lines come last. The subtext comes first.
This is the kind of real-world context that has never been digitized, which means it has never been scraped, which means it is not in the training data. It is, by definition, out of reach for LLMs. It also explains why local advertisers from injury attorneys’ outrageous jingles to used-car salesmen hawking Hondas at 2am on regional TV become icons. Dollar for dollar, I wouldn't be surprised if some of these campaigns outperform Liquid Death or Nike. They’d certainly outperform whatever Persado could come up with. But no one runs those tests.
Chutzpah
LLMs are extraordinary at research. They are increasingly exceptional at analysis, at synthesis, and at pattern recognition. They can identify market positions and competitive white space and audience segmentations. They can create solid strategic frameworks. For the work that precedes the writing, for the thinking that sets up the line, AI is the most powerful tool we’ve ever had. It can write the brief. But the brief is not the ad.
The distance between “this is a more sustainable bottled water company” and “Murder Your Thirst” is not one that better processing or more data or smarter frameworks can close. It is the gap between knowing what to say and knowing how to say it, and this second kind of knowing requires a human body that has felt thirst, and felt existential dread. Not a logical leap. A stylistic one. It requires, in other words, a person.
The AI copywriting industry has built a $2.5-billion-and-counting business on a category error: the belief that copy is a product of information and structure. It is not. Copy is a product of nerve. Yiddish has words for this particular type of productive anxiety. The best writers have chutzpah, a kind of arrogant self-confidence. But they also have shpilkis (literally “pins” or “pins and needles”), an impatient, nervous energy. A jitteriness. A crushing self-doubt that crescendos into wild self-aggrandizement and then crashes back down again over the movement of a comma or the swap of a synonym for the umpteenth time. Copywriters are writers, after all. And the most effective ones are insecure, imperfect, and prolific procrastinators. “Just Do It” was famously written on deadline.
This is a really fantastic read. Thank you. If the media is indeed the message (or massage) and we are indeed past "peak social media," it would seem there is the real potential for a post-AI-slop copywriting renaissance. Would welcome your thoughts / predictions on that front.
Thank you for the reminder that great writing creates tension!