AI Writes Your Copy — Fast and Good. Now What?

A while back, an old friend of mine who works in marketing invited me out for tea.

He'd barely sat down before it all came pouring out: his team now hands 90% of their first drafts to AI. A post that used to take a whole afternoon now takes ten minutes — three versions of it. He even did the math for me: on copywriting alone, the labor they save adds up to several hundred thousand yuan a year.

I said, that's great. Really great.

Then I asked him: has it ever gotten anything wrong?

He paused. Well, yes, there have been mistakes… but you just fix them, right?

So I told him a few stories. All from 2023 — the year ChatGPT exploded, the year every company on earth piled into generative AI, and the year the tuition came due. Bill by bill, it hurts to look at.

Today, over this pot of tea, let me tell you too.

How It "Writes" in the First Place

Start with how it works under the hood. Bear with me through the dry part — it decides every trap that follows.

What is generative AI, really?

Stripped bare: a giant word-association machine. You give it a sentence, and it starts guessing — what's the most likely next word? Guess one, write it down. Guess the next, write it down. Word after word, sentence after sentence, until the answer is complete.

It has read nearly every article, book, and webpage on the internet, which is why the chain never stumbles.

But notice: it has exactly one goal — to keep the chain smooth.

It doesn't know whether it's right. It has no way to know. It simply picks the highest-probability word, every single time.

It isn't after truth. It's after the appearance of truth.

Please commit that line to memory. Every story that follows grew out of it.

Beyond true and false, there's also voice. Its default tone is the average of the entire internet. What does an average taste like? No taste at all. Everything it writes sounds like everyone — which means it sounds nothing like you.

So the job for a brand is taming. Feed it the copy your company has written over the years — the more you feed it, the better it knows your voice. Then set rules: words that must never appear, how headlines stay consistent, what reading level you're writing for. Platforms like Writer even come with a built-in brand style guide — they'll govern whether you use the Oxford comma.

Guardrails help. But guardrails don't stop everything.

Keep reading.

A $100 Billion Tuition Bill

In February 2023, Google ran an ad for its chatbot, Bard.

In the ad, Bard answered an astronomy question: which telescope took the first photograph of an exoplanet? The answer it gave was wrong.

Just that one mistake.

That single day, $100 billion of Google's market value evaporated — nearly 700 billion yuan. All from one sentence.

Unbelievable.

If Google could stumble like that, the media fared far worse. Men's Journal, a men's magazine, ran an AI-written article about low testosterone, and doctors publicly picked its factual errors apart one by one. CNET had it worse: in its AI-written pieces, the interest calculations and mortgage rates were all wrong — and then it got called out for plagiarism. And before you say "well, a machine wrote it" — CNET, at first, only hedged and dodged, never making things clear to its readers.

Why does this happen?

Two reasons. First, it was trained on the entire internet, and the internet is thick with mistakes and lies. Second, as we said: it cannot judge true from false.

So the flaw got a name — "hallucination." Calling it a hallucination is being polite. Plainly put: it fabricates, with a straight face. And one more thing — whatever biases the internet holds, it absorbs wholesale, faster than anyone.

So what do you do? No shortcut: human fact-checking. Dates, numbers, names, places, institutions — check them one by one. Especially the numbers.

And there's a risk in the opposite direction, the one to fear even more. In 2023, employees at Samsung pasted company secrets directly into an AI chat box. Every sentence you feed it can come back out of its mouth when it answers someone else.

Never feed it your own secrets.

Two Quiet Traps

Compared with making things up, two traps are quieter and better hidden.

One: its knowledge has a cutoff date. That generation of ChatGPT — GPT-3 — knew the world only up to 2021. Ask it about the game that just ended or the movie that just premiered, and it will answer without hesitation. Everything it says will be wrong. And it won't say "I don't know." It will serve you the answer that looks most like an answer.

So for trending topics, emerging trends, and the latest search terms, do it the honest way: use a search engine.

The other: it cannot think of what no one has thought of before.

Think about it. Its entire talent is rearranging what humans have already written. An idea no one has ever put on paper simply has no slot in its probability table — so how could it land on that word?

It's an imitation machine. It will never be the first through the door.

If you want to be a thought leader, it can't help you. It might even grind you down into a very eloquent parrot.

Brands Spent Twenty Years Learning to Speak

This is the part many marketers haven't quite digested.

Over the past twenty years, what is the single most important thing brands have done? Turning themselves into a person. B2B companies built content operations to rival traditional media empires; consumer brands ran co-creation, brand collaborations, and put employees forward to speak for the company. Why? Because people only trust people.

In a 2023 Monmouth University poll, only 9% of Americans believed AI would bring more good than harm. Nine in ten were skeptical.

Switch the byline from a human to an AI at a moment like this, and: media won't cite an AI-written piece; industry conferences won't put an AI on stage; the columns, roundtables, and podcasts — trust assets built on humans — are all closed to it.

And here's the stinging part. Don't assume AI copy performs better, either.

Phrasee and Persado, companies that specialize in performance-copy AI, have been at this for years. They train on your own historical campaign data — in theory, nobody knows better which line gets the click. Yet brands' experience with them has run hot and cold: clicks tick up a little, everything falls apart at the lower funnel, and once the licensing fees are paid, little is left.

If even the specialists look like this, general-purpose generative AI mainly buys you time. Saving time and lifting performance are two different things.

So who has the final say on performance? The A/B test.

And it doesn't understand the craft of the channel, either. Anyone who has done email marketing knows: front-load the keywords in your subject line; make the subject line match the body, so the people most likely to click are the ones who open. Details like these — it can't tell how a promotional email differs from a business email.

What it hands you is an unfinished shell. The finishing work belongs to people who know the channel.

So — How Should You Use It?

After all these traps, don't get me wrong. I'm not telling you to stay away.

The opposite. Use it. Just put it in the right position.

Which position? The supporting role.

Its talents are real: it can write from a single prompt out of thin air, grind through long drafts step by step with you, restyle a manuscript you hand it, and feed you lines while you type. But the three things most worth giving it right now are these.

First: beat the blank page. The hardest part of writing is the blank page at the top. Have it produce 20 ideas in one breath: I sell lawn and garden supplies — give me 20 things customers might do to their yards this spring; I run a childcare center — give me 10 reasons parents hesitate to enroll. In half a minute it fills your screen. Fast — genuinely fast. Cross out the junk; two or three will spark something. Once one does, check it against your historical campaigns, do the research, then run a round of A/B testing.

Second: one draft, many servings. When the long piece is done, have it compressed into a four-sentence summary for the email, then rewritten as a social share line. Ask for a few versions, pick one, and edit it into shape.

Third: one message, spoken to several audiences. The same bookkeeping software's copy — have it rewritten for the small-business owner, for the finance team at a large corporation, for the restaurant owner down the street. Or simply lower the reading barrier so more people can get through it.

Notice something: all three sit either at the very beginning of writing or at the very end. The stretch in the middle — where you stake out ideas and make real judgments — it doesn't touch.

That's exactly right.

One Last Reminder: Writing Is Thinking

Many bosses have the same idea in mind: I supply the thinking, AI writes the words. How nice would that be?

But that misunderstands writing — as if it were transcribing what's already been thought through.

Wrong. The act of putting words down is the act of thinking. To make one sentence flow, you are forced to work out its structure, its details, whether it actually holds up.

Outsource that process, and what you save is no longer just time.

What you save is thinking.

You can hand over the pen. Never the brain.

Three Years On, Looking Back

Right — we should talk about where things stand now.

Back in 2023, there was the mania on one side and my cold-water advice on the other. The most rational call at the time: the technology was still in its awkward, reckless adolescence — flawed all over, but worth betting on long-term. Money would pour in, models would get sharper, the market would consolidate, pricing would mature.

Three years on, those calls have largely come true. Today's models are more than a notch above that generation, and hallucinations are down. Note: down, not gone. And the players really have consolidated into the hands of a few.

So the real exam question has long since shifted from "whether to use it" to "how to use it deeply." Over the past fifty years, the office has been remade by technology, round after round — each round eliminated some jobs and grew a new crop of roles. This round will be no exception.

The method is still the one from back then: start small, and don't expect a single big-bang switchover. Take the workflow apart and reimagine it segment by segment. Each time you touch a segment, ask one question: does this, done this way, make the customer's experience better — or worse?

When the tea ran out that day, my friend asked me: so by your account — do we still dare use it?

I said: use it. Use it boldly.

Just before you sign your name to it, ask yourself one thing:

Of these words, which ones must come from me?

Get that question straight, and AI is your leverage. Don't, and it's your landmine.

And one last wish: may you never lose real money over a line an AI made up.