You don't have an artificial intelligence problem. You have an artificial stupidity problem.

In 1907, a retired schoolteacher named Wilhelm von Osten invited journalists, scientists, and a commission from the Prussian Academy of Sciences to watch his horse do arithmetic.

The horse was called Clever Hans. Von Osten would ask a question — "What is twelve divided by three?" — and Hans would tap his hoof four times, then stop. The audience gasped. The horse was tested repeatedly, under controlled conditions. He got the answers right roughly ninety percent of the time. For a brief, extraordinary moment, people genuinely believed a horse could think.

He couldn't. A psychologist named Oskar Pfungst eventually worked out what was happening. Hans wasn't solving problems. He was reading the room. Each time he tapped, he watched the questioner's posture, their breathing, the micro-tension in their face. When he reached the right number, the person unconsciously relaxed. Hans stopped tapping. The answer was correct. The thinking was absent.

He wasn't intelligent. He was the most sophisticated pattern-matcher in the stable.

I think about Clever Hans every time I see a marketing team celebrate what their AI just produced. The output looks right. It reads well. It hits the brief. And underneath it all, the same thing is happening: pattern-matching so fluent it passes for thought.

The horse in the machine

Here's the uncomfortable question most marketing teams aren't asking: is your AI actually thinking, or is it just very good at tapping its hoof?

A large language model is, at its core, a prediction engine. It calculates the statistically most probable next word, given everything that came before it. It does this extraordinarily well. The result is text that sounds informed, structured, and confident. It has the cadence of thought. The architecture of argument. The surface texture of intelligence.

But prediction is not reasoning. Fluency is not insight. And producing text that looks like it was written by someone who understood the audience is not the same as understanding the audience.

When you ask an AI to "write a nurture email for CFOs in financial services," it doesn't pause to wonder what a CFO in financial services is actually worried about right now. It doesn't consider whether this particular buyer is in growth mode or under board pressure. It doesn't weigh which psychological lever — autonomy, risk aversion, status, fear of irrelevance — would move this person at this moment. It doesn't figure anything out.

It retrieves. It patterns. It taps.

And we call this artificial intelligence.

Artificial stupidity, at scale

What most organisations are actually practising isn't artificial intelligence. It's artificial stupidity: the industrialisation of content that has the appearance of thought but contains none of the thinking.

The output is grammatically impeccable. Structurally sound. Hits all the keywords. Passes every readability score. And it says nothing a thoughtful reader hasn't read a hundred times before — because it was engineered to produce the statistical average of everything that already exists.

This is the specific, structural limitation that matters. An AI trained on the corpus of existing marketing content will produce content that sounds like existing marketing content. It will regress to the mean, by design. The most probable next word is, by definition, the most ordinary one. The most common argument. The most familiar framing. The tone that blends in.

In any other field, we'd recognise this for what it is. A doctor who diagnosed every patient with the most statistically common condition wouldn't be considered intelligent. They'd be considered dangerous. A lawyer who filed the same brief regardless of the case wouldn't last a week. Intelligence, in every domain that matters, is defined by the ability to respond to the specific situation in front of you — especially when that situation is one you haven't seen before.

What intelligence actually is

The developmental psychologist Jean Piaget offered a definition of intelligence that has stuck with researchers for decades: "Intelligence is what you use when you don't know what to do."

Not when you know the answer. Not when you've seen this before. When you haven't. Intelligence is the capacity to face a genuinely novel situation and figure out the right response. It's adaptive. It's contextual. It requires understanding, not just retrieval.

David Wechsler, who created the most widely used intelligence tests in history, framed it differently but arrived at the same place: intelligence is "the capacity to act purposefully, to think rationally, and to deal effectively with the environment." The operative word is deal — not recite, not reproduce, not recall. Engage. Navigate. Figure out.

By either definition, most AI use in marketing is not intelligent. It is not dealing with anything. It is not navigating the specific, unrepeatable reality of a particular buyer in a particular situation with particular pressures. It is producing a statistically average response to a statistically average version of the question. And because the result is fluent, nobody notices that nothing has actually been figured out.

Every buyer is a new situation

This matters in marketing more than in almost any other domain, because every buyer is a genuinely new problem.

Not a new segment. Not a new persona. A new person. With a specific role in a specific company in a specific industry at a specific moment, under specific pressures, with a specific buying group around them — each bringing their own motivations and fears to the decision.

The CFO at a mid-market SaaS company under board pressure to cut costs is not the same buyer as the CFO at an enterprise manufacturer in growth mode. Same title. Different universe. The content that would earn trust from one would alienate the other. One needs reassurance. The other needs vision. One is in prevention mode, protecting what they have. The other is in promotion mode, chasing what they could gain.

No single piece of content, however fluent, addresses both. And no pattern-matching engine, however sophisticated, can figure out which one is sitting on the other side of the screen — because figuring that out requires something the models weren't built to do: understand the person before you write for them.

This is where the definition of intelligence starts to bite. If intelligence is what you use when you don't know what to do, then the real test of an AI system isn't whether it can produce a plausible email. It's whether it can face a buyer it has never encountered, in a context it has never seen, and work out what that specific person needs to hear.

Most can't. Most don't even try. They produce the same email for everyone and rely on the law of large numbers to make the metrics look acceptable.

That's not intelligence. It's the industrialisation of guesswork.

The cost of not figuring it out

The cost isn't just mediocre content. It's the systematic erosion of the one thing B2B marketing is supposed to build: trust.

Trust, as the behavioural psychologist Dr Paul Marsden frames it, runs on two dimensions: competence and care. Competence is whether you can do what you claim. Care is whether you've done the work to understand what the buyer actually needs

Content that was never grounded in understanding — content that was pattern-matched into existence without anyone figuring out who it's for — fails the care test every time. It might be competent. It might be correct. But the reader can feel that nobody stopped to think about them specifically. And in that moment, you don't just lose attention. You train the buyer to stop trusting that you understand their world.

Sixty-five percent of B2B content goes unused (Forrester). Not because it's badly written. Because it wasn't written for anyone in particular.

The distinction that matters

There's a difference between a system that can produce and a system that can figure out.

Production is the easy part. The models are extraordinary at it. Give them clear instructions and they'll generate fluent, structured, polished prose in seconds. That capability is real, and it's valuable.

But production without figuring out is just sophisticated stenography. Someone — or something — still needs to do the thinking: Who is this person? What's driving them? What's their situation? What would change their behaviour? What should we say, and why should we say it this way rather than any other way?

That's the work. That's where intelligence lives. And it's the part that most AI deployments skip entirely, because the speed of the output creates an illusion that the thinking has already been done. You typed a prompt, the machine responded, therefore a process occurred. But the process that occurred was autocomplete, not intelligence. The machine completed your sentence. It didn't figure anything out.

Using intelligence intelligently

The answer isn't to abandon AI. The technology is genuinely extraordinary. The answer is to stop using it stupidly.

That means pointing AI at the hard problem, not the easy one. The easy problem is production: words on a page, an email in a template, a post in a feed. The hard problem is the one that precedes it — understanding the buyer deeply enough that what you produce is worth their time.

What would it mean to use AI intelligently? It would mean deploying it to profile an audience before a single word is written. To identify what drives a specific reader, what they resist, what would change their behaviour. To match content not to a job title but to a psychological reality. To figure out what to say — not just how to say it.

That's the distinction Ada was built around. Not a faster way to produce content, but a structured way to understand the reader first: their industry pressures, their cultural context, their role psychology, their decision-making patterns. The understanding is the product. The content is the proof.

Clever Hans captivated the world because his performance was indistinguishable from thought. But performance and thought are not the same thing. The difference only showed up when someone asked the right question: is the horse actually thinking, or just reading the room?

Your AI is reading the room. The question is whether anyone's asked it to think.