Should Marketers Panic? How Generative AI Will Reshape the Business

A while back, I came across a piece of news.

Coca-Cola used AI to create a new beverage called Y3000. The recipe, the flavor profile, the packaging visuals — part of it was "dreamed up" by generative AI.

My first reaction: even a business built on taste buds and flavor — the most human of senses — is letting AI get its hands in?

Then I stumbled onto a few more. Unilever, Nestlé, Metro — all using AI to generate advertising creative. Meta publicly announced plans to let AI automatically produce ad assets. And there was a forecast saying marketing would be the department most disrupted by generative AI, unlocking roughly $463 billion in productivity every year.

$463 billion.

You read that right.

So I started wondering: how does this thing actually work? Why, all of a sudden, can it do everything? Should marketers embrace it, or should they be nervous?

I shut myself in my study and read for days, trying to make sense of it. Today, I want to walk you through what I figured out.

First, Let's Get Clear: Why Can Generative AI Do Everything?

What exactly is generative AI?

You give it a sentence, a prompt, and it spits out an article, an image, a video — even a piece of code.

It sounds almost mystical, but the underlying logic isn't really that mysterious.

We used to train AI through "supervised learning" — humans fed it standard answers, and it learned to match them. The problem? Standard answers are expensive. They require labeling, manual labor. So the data volume hit a ceiling, and the model's capabilities got stuck there.

Then someone had a different idea: stop feeding it answers. Instead, dig out part of the data and let the model guess what's missing. This is called "self-supervised learning."

Here's an example. You give it a sentence: "This is a ___ article." It has to figure out that the blank is probably "commentary" or "science" — and probably not "umbrella."

What's the advantage? The answers are already hidden in the data itself. No human labeling needed. You can pile on as much data as you want — text, images, video, audio, all of it. The models trained this way are called "foundation models."

So how does it generate new things? Honestly, it's basically playing the next-word game. You give it a starting point, and based on everything it saw during training, it calculates what the next word is most likely to be — and keeps going, word by word.

Here's the key: every time it generates a word, there's randomness built in. The same starting prompt can produce different outputs each time.

That means it's inherently capable of "inventing" new things.

Think about it. This is crucial. Randomness, combined with the sheer volume of data it has seen, means it can connect two completely unrelated concepts. And historically, isn't that how many great inventions happened — from medical breakthroughs to Post-it Notes to shatterproof glass?

But being "new" isn't enough.

What you come up with also has to be "right." It has to be the kind of "right" that consumers can accept and are willing to pay for.

Someone ran a comparison experiment. They had ChatGPT and a group of MBA students each come up with ten product ideas, then scored them. The result? ChatGPT's ideas were, on average, more popular than the MBA students'. Even more interesting — in one group, the researchers fed ChatGPT a batch of "high-quality ideas" as reference. The result? Whether they fed it or not, the scores were about the same.

Why?

Because it had already seen them all. In its training data, it had already absorbed what humans consider "appropriate." Adding more is redundant.

So you see, generative AI fundamentally has two genuine strengths: one is "novelty," the other is "appropriateness." Put them together, and in academia, there's a word for it: creativity.

But It's Not a God — It Has Bugs

Having said all that, I need to pour some cold water on the hype.

In February 2023, Google unveiled Bard — its answer to ChatGPT — with great fanfare. During the launch event, Bard answered a question about the James Webb Space Telescope.

Astronomers took one look and spotted the error.

That single mistake wiped $100 billion off Google's stock value in a day.

Two weeks earlier, Meta had released Galactica, a large model designed "for scientists." It was pulled offline after just three days. The papers it generated looked incredibly convincing — but the facts were entirely wrong.

This kind of "authoritative-sounding nonsense" has a name in the industry: hallucination.

Why does it hallucinate? Computer scientists' explanation: most current large models are trained only on text. They've never experienced the physical world. They don't know that a cup dropped will fall, that water flows downhill — they've only seen words describing these things.

So when it comes to reasoning and causal judgment, they often crack under pressure.

But don't jump to conclusions just yet. Multimodal models are already starting to incorporate images, audio, and video into training. Whether this gap can be closed in the future — nobody can say for certain.

What we can say is, right now, what it's best at is the kind of work that has "no single right answer." Drawing an image, brainstorming product ideas, writing ad copy — these are where it shines. When it comes to tasks requiring rigorous fact-checking and causal reasoning, you need to be the gatekeeper yourself.

The Real Show: It's Going to Rewrite Innovation Itself

Alright, the technology is clear. Let's get to the main event — how exactly does it change marketing?

I'll break it into four moves. These four cover basically the entire pipeline from a company having an idea to keeping users engaged: come up with it, test it, tell people about it, keep them around.

Let's go through them one by one.

One: Come Up With It — AI Helps You Ideate, but Watch Out, It Might "Homogenize" You

When companies want new ideas, they used to rely on a common tactic: crowdsourcing. Bring consumers in, set up an open innovation platform, let everyone pitch ideas. Dell and Procter & Gamble (P&G) both did this.

Now generative AI is here, and with its help, a single consumer can generate a pile of proposals in minutes.

Sounds great, right? But there's a hidden trap.

When people brainstorm together and see each other's ideas, two reactions kick in: one is "cognitive anchoring" — getting pulled along by the other person's idea; the other is "cognitive stimulation" — actually wanting to be more different from the other person.

After generative AI enters the picture, both reactions still happen. But here's the problem: what it spits out, everyone has access to.

If one user asks AI to generate ten ideas, and another user asks the same AI for ten ideas, how likely is it that the two sets collide?

Very likely. Because behind them is the same model, the same training data.

Here's the awkward part. The whole point of crowdsourcing innovation is that "a hundred people have a hundred different ideas." But once AI walks in, a hundred people's ideas might all look the same.

So what do you do? One researcher proposed an interesting approach — a kind of "reverse engineering." Instead of having AI rush to give answers, have it ask consumers questions instead, pulling them out of the answers AI already provided and pushing them to think further.

Whether this approach will actually work, nobody knows yet. But this question is worth every marketer thinking about.

Two: Test It — Can AI Replace Consumers in Answering Surveys?

In the past, when companies tested new products, they had to do market research, send out surveys, run focus groups. Slow, expensive, and often inaccurate.

Now some people are saying large models might be able to replace a portion of consumers in answering surveys.

You might think that sounds like science fiction. But someone has already proven it.

Researchers treated GPT as a "virtual consumer" and asked it questions. They found that the demand curves it produced were highly consistent with real human survey results. Reactions to price sensitivity among higher-income groups, inertia in product choices — all matched up. Some even used it to reconstruct brand perception maps that looked almost identical to those drawn from real surveys.

There's even research showing that when it comes to tagging marketing content, ChatGPT is far more reliable than crowdsourcing platform workers — nearly as accurate as experts.

What's going on? The researchers call it "algorithmic fidelity." You give the model a description — say, "you are a 35-year-old woman living in a tier-two city, earning 10,000 yuan a month" — and the answers it produces are highly correlated with what real people matching that profile would say.

Because in its training data, it has seen an enormous amount of expression from this type of person.

But here's the catch: the more it's seen of a particular group, the more it sounds like them. For niche categories and lesser-known brands, its performance drops noticeably. Because there simply isn't enough relevant content in its training data.

So yes, you can use it for research — but you need to know what it's good at and what it isn't.

There's also a more advanced play: don't just stick to one model. Each model has different training data, different architecture, different strengths and blind spots. Combining results from several models might be far more accurate than relying on any single one.

Three: Tell People About It — Can Ads Written by AI Actually Move People?

Gartner predicted that by 2025, 30% of brands' marketing messages would be generated by AI.

That's a significant proportion.

But here's the question: can things written by AI actually persuade consumers?

The research is deeply divided. One study found that AI-written political persuasion messages were about as effective as human-written ones. Another found that copy written by ChatGPT actually achieved higher user satisfaction and willingness to pay than human-written copy — even experts couldn't beat it.

My own experience after using it: AI tends to be "long-winded" when writing. Turn the temperature parameter up a bit, and it starts rambling. And verbose expressions actually tend to score higher in creative evaluations, because they read as "rich and full-bodied."

But "rich and full-bodied" and "persuasive" are two different things.

So the real research question isn't "can AI write," but "which parameter set to which value produces content that's most effective for which type of audience."

This is not something marketers can hand off to the tech department. This is marketing's own job.

Four: Keep Them Around — AI Helps You Create Images, but Does the Sense of "I Made This" Survive?

After a customer buys something, you need to keep them engaged. This is called customer engagement.

How do you keep them engaged? Have them do things. Run creative contests, let them make works using your materials, have them vote. Once someone has poured effort into something, they feel "I made this" — in psychology, this is called psychological ownership. Ownership goes up, stickiness follows.

Coca-Cola did something along these lines. They partnered with OpenAI to launch the Real Magic platform, where users could use GPT-4, DALL·E, and Coca-Cola's historical brand asset library to generate original images. Winning entries would be displayed on the big screens at Times Square in New York and Piccadilly Circus in London.

The barrier to entry was demolished overnight. People who couldn't draw at all could now participate.

But the problem also arrived.

The reason people feel attached to a piece of work is the effort they put into it. Now the effort is being supplied by AI — will they still feel emotionally connected to the work?

Whether this is a zero-sum or positive-sum situation, nobody can say for certain right now.

From the company's perspective, the benefit is obvious. Before, you could only attract creatively inclined users; now anyone can participate, and the participation base widens instantly. But from the individual user's perspective, stickiness might actually weaken.

That's the tension: breadth versus depth.

How to balance it is something every brand builder needs to think about.

The Deeper Question: Over Time, Will It Change Us in Return?

Those four moves above were all about "how we use it."

But after talking with a group of executives running real businesses, I realized what they're genuinely anxious about is a different question entirely.

As we use it more and more, will we — and our customers — be changed by it?

One executive put it bluntly: "Right now, everyone on the board is asking what value the marketing department can still create for the company. Cost-cutting is something they've always wanted. AI arrives, and that blade just got sharper."

That hit close to home.

The value marketing creates for a company rests on a pile of "market assets": brand, customer relationships, customer insights, channel capabilities. These things are, in theory, scarce, hard for others to imitate — and therefore valuable.

But with generative AI's arrival, two things are happening simultaneously.

One: many capabilities that used to be valuable suddenly aren't. Writing copy, generating ideas, conducting research — these used to require maintaining an entire team. Now a single API call handles it. You can do it, and so can everyone else. The scarcity is gone.

The other: it's not omnipotent either. Its output depends on what you feed it. If you want it to produce something genuinely tailored to the current market, you need the freshest customer data, the most accurate market insights. Those things still require humans to accumulate and maintain.

So at this stage, "human-in-the-loop" probably can't be avoided. AI produces the draft, humans make the call. But what about further down the road?

Nobody can give a definitive answer. The verdict on this one may have to be left to time.

There's Another Angle That's Even More Chilling

I read a study saying that people who rely on GPS navigation long-term show weakened hippocampal function and deteriorated sense of direction.

Use smartphones too much, and your attention span gets shorter.

After the printing press became widespread, humanity's ability to memorize and recite from memory basically atrophied.

So what about generative AI?

It generates creative ideas for you, writes copy for you, does design for you. At first, you feel like it's doing your work for you. But gradually — will your "creative muscles" atrophy?

Consumers too. When AI handles all the creative work for them, will they still value "creative" things? Will their preferences for innovation change?

We are facing a consumer who may have been reshaped by AI. Their tastes, their preferences, the role they play in marketing — all of it could look different from today.

When you think about it carefully, it sends a chill down your spine.

Wrapping Up — A Few Words About the Bigger Picture

Beyond the main thread above, as I was reading through the research, I noticed three more directions worth everyone's vigilance.

One is privacy. For companies to use AI well, they need to fine-tune models with their proprietary data. But feeding that data to a third-party model is essentially revealing half the cards in your hand. Scarcity can vanish just like that. Which companies should feed what, and how much — that's a strategic question, not a technical one.

Another is misinformation. AI can churn out ten thousand fake news stories a day at nearly zero cost. This used to be considered a public policy problem, but organized attacks on brands have already begun — Western vaccine makers, Starbucks, H&M have all been hit. Marketers can no longer pretend this has nothing to do with them.

The last one is a bit counterintuitive. We're all discussing how AI can help marketing. But flip it around — can the accumulated knowledge of marketing science, in turn, help us better understand and harness AI?

The marketing field has decades of research tradition on human cognitive bias. How to design surveys without biasing respondents, why people make irrational decisions. This knowledge, applied to analyzing why the AI "black box" spouts nonsense at certain moments, might be more effective than computer scientists puzzling over it alone.

Because AI, in a sense, thinks in human language, mimicking human patterns. To understand it, you first need to understand humans.

This was the insight that excited me most in my days of reading.

Finally

Back to the question from the beginning: should marketers panic?

My answer is, panic isn't necessary — but not panicking would be a mistake.

Generative AI is not just another tech buzzword blowing past in the wind. It is substantively taking marketing apart at the foundation and reassembling it. Which capabilities will be diluted by it, which ones will become more valuable, what consumers will turn into, how brands should retell their stories — none of these have standard answers.

But people who are asking the questions are already one step ahead of those who aren't.

As for the answers — we'll find them as we go.

I'm still thinking about it too.