AI Isn’t Magic: It’s a Matter of Asking the Right Questions
In conference rooms and cafés alike, the same rumor keeps growing: artificial intelligence is a near-mystical creature, a spirit living in servers, able to guess, predict, invent. People whisper that it replaces judgment, that it swallows the world the way an engine gulps air. And yet, once you strip away the ornament and the vertigo, what remains is something very human, very old: statistics and probabilities, arranged by algorithms.
Ten years ago, a fourteen-year-old had to beg an adult to install a Python environment, slog through a book on linear regression, and tinker for hours. Today, with ready-made libraries and assistants like ChatGPT, that same teenager can put together a working “AI” app in a weekend, image recognition, a chatbot, automatic summarization. The modules click, the APIs answer, the magic happens. But that’s not where the real stakes lie. The real challenge is no longer to “make AI work.” The challenge is asking the right question and collecting the right data, the kind that truly answers that question.
This idea, which sounds almost banal once spoken aloud, springs from a story engineers know well, but the general public rarely hears. A story of war, airplanes, bullet holes, and a bias that proved costly until someone finally saw it.
The Planes That Come Back… and the Ones That Don’t
Picture an air base during World War II. In waves, aircraft return from missions. Mechanics rush in, take notes, and mark on silhouettes of wings, fuselages, and tails the areas riddled with bullet holes. A pattern quickly emerges: here, the wings are often hit; there, the tail; elsewhere, the mid-fuselage. The obvious conclusion: reinforce where bullets hit most.
So they add armor. The plane gets heavier, fuel burn rises, range drops. They hope losses will fall. They don’t.
Until one day a sharp-eyed statistician asks a disarmingly simple question: “On which planes are we taking these measurements?” Answer: the ones that come back. And the whole edifice collapses. The observed bullet holes don’t show where planes are shot down—they show where a plane can take hits and still make it home. In other words, they mark the least vital areas.
The real conclusion is the reverse of the first intuition: armor the places where there are no holes on the surviving planes. By definition, those are the critical points: hits there prevent the aircraft from returning, so you never observe those impacts.
This story—what statisticians now call a classic case of survivorship bias, isn’t just a military anecdote. It’s a fable about the quality of the question and the relevance of the dataset. It says this: if you only look at what survives, you’ll miss the secret of what dies. And if you only query the data you have, you risk ignoring the data you don’t, data whose absence holds the answer.
Asking the Right Question: A Discipline
AI, first and foremost, is about defining a question that correctly targets the phenomenon you want to model. It’s less “What’s the best architecture?” and more “What are we trying to predict, for whom, when, and with what consequences?” It’s deciding whether you want a score (ranking), an alert (threshold), a recommendation (action), or an explanation (risk factors). Above all, it’s stating what you agree to ignore and what you refuse to forget.
The Beauty of Well-Framed Questions
There’s an elegance—almost literary—to framing a question well. It’s choosing the angle of the light. The same scene under a warm glow tells an intimate story; under a cold neon, a clinical documentary. With AI, framing is everything. “Churn” can be defined as contract termination or as declining usage. The first definition triggers action too late; the second gives you a head start. “Patient risk” can mean probability of a complication or probability of ignoring a care recommendation. The former concerns clinical practice; the latter, behavior, and calls for different strategies.
We marvel, rightly, at the feats of large language models. But if you ask them poorly, they produce beautifully crafted off-topic answers. Give them a clear objective, context, constraints, and positive/negative examples, and everything changes. In the same way, a predictive system is only as good as the intent that guides it.
Pedigree vs. Practice: The Fetish Trap
With so much on offer, it’s tempting to fetishize the tool. We compare models like sports car engines: this model beats that one on a benchmark; that framework promises a ×-fold speedup. Useful, until it isn’t. A middling model fed with good data and guided by a plain engineering question will almost always beat a top-tier model poorly framed. An organization that documents its sensors, versions its transformations, records its assumptions, and listens to its operators moves more surely than one that stacks opaque bricks.
Epilogue: Armor Where There Are No Holes
It’s striking that such an old story still illuminates our brand-new software. What the air base teaches isn’t just a statistical trick; it’s an ethic of curiosity. When the “obvious” lines up a little too neatly, turn toward the silence. Where the data falls quiet is often where the causes hide.
So yes, AI is just stats and probabilities turned into algorithms. And that’s a good thing. It means the power remains in our hands: in the art of asking the right question, in the rigor of choosing or creating the right data, in the patience to return to reality. Tomorrow, a teenager will still build an app in three clicks. What will make the difference won’t be the polish of the UI or the size of the model, but the rightness of the problem and the quality of observation.
Armor where there are no holes. Ask the question that reveals the missing. Make the silence speak. The rest—the accuracy, the performance, the rising curves—will follow.
