Contents
For a few years now, like pretty much everyone, I’ve had an ambivalent relationship with AI.
At first, it was wonder… like many people, I saw these tools as an incredible accelerator.
In a few seconds, you could generate ideas, structure presentations, write content.
I dove in quickly myself, probably too quickly, into this logic of exploration. Automation, agents, content generation: everything seemed possible.
Then, slowly, something started to bother me. Not because the results were bad.
Quite the opposite.
They were often good. Too good, even. Coherent, fluid, well structured. Despite that, over time, a persistent feeling set in: one of déjà vu.
As if, behind the apparent diversity, everything was starting to look the same.
Historically, in data, we talk about garbage in, garbage out.
Bad data produces bad output; a visible error, easy to spot and correct.
But with generative AI, we enter a different dynamic.
It’s no longer garbage out.
I call it average out.
AI doesn’t simply reproduce errors. It absorbs them, blends them and turns them into something plausible, credible, but rarely distinctive.
As if AI, rather than elevating creativity, were helping to normalize it.
From the visible error to the invisible average
What fascinates me (and worries me) is that by moving from error to average, we fall into a new paradigm where content is credible enough to go unnoticed, but rarely distinctive enough to create value.
An error is visible. It gets corrected. It triggers a critical reflex. But average content slips under the radar.
It follows the codes, strikes the right tone, ticks every box.
It often even gives an impression of mastery.
I think back to an example that stayed with me: nutritionists analyzed AI-generated turkey recipes.
On paper, everything was perfect: consistent ingredients, clear steps, an appetizing result.
But digging deeper, they found errors in cooking times and temperatures, serious enough to pose a real health risk.
The recipe was good… until it wasn’t.
That is exactly what average out is.
Creativity that converges
To understand this phenomenon, we need to go back to the very nature of these tools.
AI models don’t create in the human sense.
They predict.
From vast datasets, they generate what is most likely in a given context.
This mechanism, sometimes described as manifold filling, consists of interpolating between existing ideas. The result may seem new, but it remains anchored in what is already known.
A recent example illustrates this limit well.
Researchers developed a model called Talkie, trained exclusively on data from before 1930. The result is fascinating: the model produces coherent, credible language… but deeply rooted in the cultural references, expressions and logic of that era.
In other words, it doesn’t “understand” the modern world. It simply extends the one it was trained on.
This case highlights an often underestimated reality: an AI never truly goes beyond its corpus. It only explores the boundaries of what it already knows.
In practice, this means AI excels at optimizing and recombining… but not at breaking away.
Recent studies also show that, even in very different strategic contexts, AI-generated recommendations tend to converge toward similar answers. A form of average thinking, amplified at scale.
When it’s easy, everyone does it
Marketing is built on differentiation.
Creating an emotion, standing out, occupying a unique space in the consumer’s mind.
So isn’t the unreflective use of AI pushing us in the opposite direction?
I was recently struck by a rather revealing example.
During an April Fools’ contest in Quebec, several organizations, independently of one another, came up with exactly the same concept: linking two places with a zip line.
Nothing in their history suggested this kind of idea. And yet they reached the same conclusion.
Hard to see it as a coincidence.
This kind of convergence illustrates a broader phenomenon: when thousands of professionals rely on the same models to “brainstorm,” the ideas generated tend to look alike, even when the initial intention is to stand out.
By democratizing access to content production, AI is gradually turning certain skills into commodities. What used to be difficult (producing coherent, structured, credible content) becomes trivial.
In my view, this democratization has a side effect: it dilutes scarcity.
In an environment where anyone can produce “decent” content, the value of that content drops mechanically. We are seeing a kind of inflation of perceived quality… where “good” becomes the norm and is therefore no longer enough.
The productivity illusion
People often tell me: “Yes, but it’s still a huge productivity gain.”
It’s true. But only partly.
Yes, we produce faster. But are we producing more value?
What I observe is a shift in work. Less raw production, but more sorting. More validation. More effort to tell what is worth keeping.
Some now talk about workslop: clean, fast content with no real value.
And it creates an interesting paradox: the more we produce, the harder it becomes to make an impact.
The real competitive advantage: knowing when not to use AI?
Looking back, I’ve come to a fairly simple conclusion.
The problem isn’t AI.
It’s the way we use it.
AI is a terrific tool for exploring, testing, structuring.
An amplifier. But it isn’t a creator.
The most important elements in marketing (insight, positioning, tone, a strong idea) rest on dimensions AI handles poorly: context, culture, intuition.
In this context, a reversal is taking place.
The advantage no longer lies in adopting AI; it is now accessible to everyone.
It lies in the ability to use it with discernment.
And sometimes, in the ability not to use it.
Knowing when to slow down. When to step outside the frame. When to favour a human approach, imperfect but singular.
In the age of average out, the discipline is no longer to produce more, but to resist the average.