Most advice about AI-proofing a career is a list of things that sound safely human. It is comforting and close to useless, because it never explains the mechanism, and without a mechanism you cannot tell whether the list will still be right next year.
The mechanism is simple. When something becomes cheap to produce, the value moves to whatever is still scarce next to it. So the question is not which skills feel human. It is which skills become more necessary precisely because output got cheap.
Why cheap output makes judgement expensive
When writing a competent report took two days, the report was itself a filter. Only people who understood the subject well enough could produce one, so the artefact was weak evidence of understanding.
That filter is gone. Anyone can produce the artefact now, which means the artefact proves nothing and somebody still has to determine whether its contents are true. That determination did not get easier. It got more frequent, because there is far more plausible material arriving that needs checking.
This is the whole moat in one sentence: the value moved from making the thing to being able to tell whether the thing is any good.
The skills that got more valuable
These are not soft skills in the dismissive sense. Each one is testable, each one takes years to build, and each one is now load-bearing in a way it was not when production was the bottleneck.
- Domain judgement. Knowing the field deeply enough that a wrong answer feels wrong before you can prove it. This is the single hardest thing to fake and the hardest to automate, because it is built from consequences rather than from text.
- Deciding what matters. Prioritisation is a claim about the world, accountable to outcomes rather than to how well it was argued. Generated evidence does not make the claim for you.
- Persuasion and holding a position. Most bad decisions are not analysis failures. They are what happens when nobody will tell a senior person their idea is not worth a quarter.
- Taste. The ability to look at three acceptable options and know which one is actually good. Models produce the average of what they have seen, which is by definition not distinctive.
- Negotiation and dealing with conflicting interests. Nothing about this is an information problem, which is the only kind of problem a model helps with.
- Knowing what to ask. A model rewards a precise question and punishes a vague one, so the ability to frame the real question has become a direct multiplier on output quality.
The skills that lost value, said plainly
It is not kind to leave this part out. Fluent writing on its own, producing a competent first draft, formatting and structuring documents, summarising material somebody else wrote, and routine research have all fallen sharply in price. Not to zero, but far enough that they no longer distinguish anybody.
If your reputation rests mostly on producing polished artefacts quickly, that reputation is resting on the part that got automated. That is worth knowing early rather than late.
How to build the moat without waiting for permission
None of these are courses. They are built from repetitions with feedback, which means the practical move is to arrange for repetitions with feedback.
- Make explicit predictions before you decide something, then check them. Judgement without a scorecard is just confidence.
- Volunteer for the decisions rather than the deliverables, even when the deliverable is more comfortable.
- Review generated work critically as a habit. Being the person who finds the flaw is how you become the person who is trusted.
- Go deep in one domain rather than shallow across five. Depth is what the wrongness detector is made of.