AI Gives You Answers. Your Value Is Asking Better Questions
There is a line I learned early in research and have never managed to unlearn: a well-posed question is half the paper. Not half the effort — half the *result*. Once the question is framed correctly, the analysis often falls out of it almost mechanically. Frame it badly, and no amount of clever method will save you. You can run the finest statistics in the world on the wrong question and produce a confident, publishable, useless answer.
I have been thinking about that line a great deal lately, because the tools have changed and the line has not. Ask a large language model almost anything and it returns a fluent answer in seconds. The cost of an answer has collapsed. But the cost of the *question* — the work of deciding what is actually worth asking — has not moved at all. If anything, it has gone up, because now the whole system waits on it.
This is the quiet inversion of the AI moment, and it is why I keep telling worried young professionals the same thing I argue at length in how AI makes critical thinkers more valuable, not less: the machine is not coming for your judgment. It is coming for your answers, which means your judgment is about to be the only thing you are really paid for. And the sharpest edge of judgment, the part that sits furthest upstream of everything else, is problem framing.
What framing actually is
Let me be precise, because “asking better questions” can sound like a motivational poster.
Framing is the act of converting a vague, tangled situation into a specific, answerable question — and choosing *which* answerable question, out of the many available, is the one worth answering. It has three moving parts. First, you decide what the real problem is, as opposed to the symptom that announced itself. Second, you decide what a good answer would even look like, so you would recognise one if it arrived. Third, you decide what you are willing to ignore, because every frame excludes something, and pretending otherwise is how frames go wrong.
An AI does none of this. It takes your frame as given and optimises inside it with unnerving literalism. Ask it to make your report shorter and it will make your report shorter, including the three sentences that were the only reason to write the report at all. The model is a superb answer-engine bolted to whatever question you hand it. The quality of the output is capped by the quality of the frame, and the frame is entirely yours.
EVIDENCE GRADE: STRONG
That last claim is as strong as they come, and it is structural rather than empirical: a system that predicts the next token conditions on your prompt. It cannot want a different question than the one you gave it. Change the frame and the entire answer changes — same model, same data, same afternoon. The leverage lives in the framing, and the framing is a human act.
Where framing comes from: observe, then decompose
If framing is the skill, the fair next question is whether it can be learned or whether some people are simply born asking sharp questions. I think it is mostly learnable, and the method is older than any of the tools.
In how to think like a scientist I laid out a five-step procedure, and the first two steps are the entire engine of framing. *Observe* — look at the situation as it actually is, not as the tidy story you arrived with. *Decompose* — break the mess into parts small enough that each one is answerable on its own. A good research question is almost never one large question. It is a stack of small, brutally specific ones, arranged so that answering them in order dissolves the big vague one you started with.
Most bad questions fail at exactly these two steps. They skip observation and inherit someone else’s framing wholesale — “we need to increase engagement” — without ever asking whether engagement is the thing that is actually broken. Or they refuse to decompose, and ask a question so large that any answer is defensible and none is testable. “Is our strategy working?” is not a question. It is a mood. “Did the customers who saw the new onboarding return in week two at a higher rate than those who did not?” is a question — narrow, answerable, and wrong in an obvious way if it is wrong, which is the whole point.
The reason this matters more now, not less, is that the answer-engine amplifies whatever you feed it. Hand it a mood and it will generate a thousand fluent words of mood. Hand it a sharp, decomposed question and it becomes genuinely useful — a fast instrument pointed at a target you chose. The instrument got cheaper. Aiming it did not.
The badly framed question that cost me weeks
FIELD NOTE — ERLANGEN
Early in my move into neuroimaging I set out to answer what felt like the obvious question: *which brain regions differ between patients and controls?* I built the pipeline, ran the comparison across the whole brain, and spent the better part of three weeks chasing and characterising a difference that looked real. It was not. Once we accounted properly for how much the two groups had moved their heads in the scanner, the effect thinned to nothing — I had partly measured motion, not the brain. The painful lesson was not that my analysis was sloppy; it was competent. The lesson was that my *question* was wrong. “Where do the groups differ?” quietly assumes the difference is in the brain. The question I should have framed first was “what, other than the brain, could make these groups look different?” Weeks separated the two questions. The method could not tell me which one I was answering — only I could, and I had not stopped to ask.
I tell that story often, because it is the cleanest example I own of framing being the real work. The three weeks were not wasted because I did them badly. They were wasted because they answered a question I had never properly interrogated. An AI, dropped into that situation, would have accelerated my mistake beautifully — delivered the wrong answer in an afternoon instead of a fortnight. Speed does not fix a bad frame. It ships it sooner.
Answers are cheap; frames are the constraint
Here is the economic core of it, stated plainly. When answers were expensive to produce, the person who could produce them held the value. Framing mattered, but it was hidden inside the cost of execution — you had to be good at answers to survive long enough for your framing to show. Now that answers are close to free, the hidden variable is exposed. The bottleneck moves upstream to the only step the machine cannot perform: deciding what is worth asking.
EVIDENCE GRADE: MODERATE
I grade the *economic* claim Moderate, and I want to be honest about why. The structural fact — that models optimise inside a given frame and cannot choose it — is Strong. But the leap to “therefore framing is where career value concentrates” rests on the broader pattern of automation, where value has historically migrated to the tasks the tools could not do, plus the early evidence from these specific systems. The mechanism is sound and I would stake real conviction on the direction. The decade of workplace data that would make it Strong is not in yet. Matching confidence to evidence is the whole ethic of this site, so I will not dress a well-founded expectation as a settled law.
| The answer-engine does this cheaply | You still own this | |
|---|---|---|
| Choosing what to ask | — | Framing the real, answerable question |
| Producing an answer | Drafting, code, summary, synthesis | — |
| Judging the answer | — | Is this true, and is it the right question? |
| Deciding what to ignore | — | The trade-off every frame silently makes |
Read the middle column as the part being commoditised and the right column as the part appreciating. The most valuable move you can make this year is not to get faster at the middle column — you will not out-type a model — but to get deliberately, unfashionably good at the right one. Framing compounds. Every sharp question you learn to ask makes the next one easier, and every hour spent racing the machine at answer-production is an hour spent losing.
How to practise framing on purpose
Framing feels innate in the people who are good at it, which fools everyone else into thinking it cannot be trained. It can. It is the habit of inserting a deliberate step between the situation and the work — a refusal to start answering until you have looked hard at the question. Below is the version I use, stripped to something you can run before your next AI-assisted task.
Try this today
Before you type your next prompt, spend five minutes writing the question by hand, badly, and then fixing it. First write the vague version you actually feel — “make this better,” “figure out the pricing,” “is this working.” Then apply two moves. *Observe:* what is the real problem here, as opposed to the first symptom I noticed? *Decompose:* what are the two or three smaller, specifically answerable questions hiding inside it, each one sharp enough that a wrong answer would be obviously wrong? Now write the single assumption your frame is making — the thing it treats as settled — and ask whether that is where the real problem lives. Only then open the tool. You will notice that the answer arrives better because the question did, and that the five minutes upstream saved you the three weeks I once did not save.
Do this often enough and it stops being a five-minute exercise and becomes the reflex that runs before you reach for any tool. That reflex is the thing no model will run for you, because running it requires knowing what you are actually trying to accomplish — a question the machine can only ever inherit from you.
The skill that survives the tool
I have changed fields once already, from engineering to neuroscience, and the one capability that transferred intact was not any technique. Techniques expire. What transferred was the habit of interrogating the question before committing to the answer — the same habit whether the answer was a hearing-aid filter or a brain map. That is why I am unbothered by tools that make answers cheaper. They lower the cost of the part I was always willing to delegate and raise the value of the part I was never able to.
If you take one thing from this: stop measuring your worth by how fast you can produce answers, and start measuring it by how well you can frame the questions worth answering. That is the skill the machine cannot replace, because it is the skill the machine depends on. Problem framing is not one of several mental models for the age of AI — it is the one that decides whether the others get pointed anywhere useful. It sits at the head of the toolkit I would give any young professional, alongside the five mental models every young professional needs first, which are the tools that operate before the question is asked and after the answer arrives — exactly the ground the machine does not touch.
If this way of thinking is useful to you, the natural next step is the free guide this site is built around: [5 Mental Models to Future-Proof Your Career](/newsletter/) — five models chosen and stress-tested by a brain researcher, each with an honest grade of the evidence behind it. You’ll also get *Signal*, my monthly email: one idea from neuroscience you can use at work. No productivity spam, no AI panic.
*Mageshwar Selvakumar is a doctoral researcher in neuroscience in Erlangen, Germany, studying how chronic pain reshapes the brain using multi-parametric MRI.*