Why "learn to prompt" is weak career advice

"Learn to prompt" is advice to get good at operating an interface that is being deliberately engineered, by very well-funded people, to need less skill every quarter. It is not wrong, exactly. Prompting is a real competence, it is genuinely harder than it looks, and someone who cannot get a useful answer out of a model in 2026 is at a disadvantage. But as career advice it fails three tests at once. It is not scarce, because roughly a quarter of employed people already used AI for work in a given week by late 2024. It is not durable, because every model release absorbs more of the skill into the product. And it is not the constraint, because the thing that actually decides whether AI helps or harms your work is not how you phrase the request but whether you knew what to ask for and can tell whether the answer is right. Those two capabilities are getting more valuable as prompting gets easier, which is the opposite of what the advice implies.

The strongest case for the advice#

It deserves a fair hearing, because the people giving it are not fools and the evidence is not entirely on my side. Zamfirescu-Pereira and colleagues, in a 2023 CHI paper with the excellent title "Why Johnny Can't Prompt", gave non-experts a purpose-built tool for designing and evaluating prompts and found they struggled badly: they approached prompting opportunistically rather than systematically, over-generalised from single successes and failures, and had persistent difficulty forming an accurate model of what the system would do. So prompting is not trivially easy, and the gap between someone who does it well and someone who does it badly is real and measurable today.

There is a second, better argument. Fluency is a precondition for everything else. You cannot develop a feel for where a model is reliable and where it is confidently wrong without using it a great deal, and you cannot use it a great deal without basic competence. On that reading, "learn to prompt" is not the destination but the entry fee, and the people saying it mean something closer to "engage seriously with the tool". I have no quarrel with that at all. My quarrel is with treating the entry fee as the strategy.

Why it fails as career advice#

It is not scarce. Bick, Blandin and Deming found that by late 2024, nearly forty percent of the US population aged 18 to 64 used generative AI, twenty-three percent of employed respondents had used it for work in the previous week, and nine percent used it every working day, with work adoption as fast as the personal computer and overall adoption faster than the internet. A capability that a quarter of the workforce already exercises weekly, two years into the technology, is not a moat. It is a baseline.

It is not durable. The commercial incentive of every model provider is to make careful prompting unnecessary, because the market for a tool that requires skill is smaller than the market for one that does not. Reasoning models, system prompts, memory and agentic scaffolding all absorb work the user used to do by hand. Any capability whose supplier is actively investing in its obsolescence is a poor place to anchor a career, and the Johnny study describes a difficulty that is being engineered away rather than a permanent human edge.

It is not the constraint. This is the substantive objection. Dell'Acqua and colleagues, working with BCG and researchers at Harvard, MIT and Wharton, gave 758 consultants GPT-4 access and found that on a task just outside the model's competence, AI-assisted consultants performed worse than consultants with no AI at all. They did not fail because they prompted badly. They failed because they could not tell that a fluent, confident answer was wrong. Vaccaro, Almaatouq and Malone's 2024 meta-analysis of 106 studies makes the same point at scale: human-AI combinations underperformed the better of human or AI alone, with losses concentrated in decision-making, and what determined the outcome was the allocation of the task, not the quality of the interaction.

And the returns run the wrong way. Brynjolfsson, Li and Raymond found AI assistance raised productivity by thirty percent for the newest workers and almost nothing for the most skilled. Read as career advice, that is uncomfortable: the tool levels the floor. A capability that raises novices to near-expert output is, by definition, a capability that stops distinguishing you.

The strongest objections to this#

The Johnny study is from 2023 and used the models of that moment, so it may now understate how much easier prompting has become, which strengthens my argument, or it may understate a persistent difficulty, which weakens it. Nobody has run the equivalent study on current models. The adoption figures are self-reported and count any use, which is a low bar and not evidence of skilled use. And it remains possible that a durable, senior version of prompt design survives inside engineering and product roles, in which case "learn to prompt" is sound advice for a narrow population and poor advice for the general one. That is a real caveat and I would not pretend otherwise.

There is also a fair objection to my position, which is that this is a distinction without a difference: if fluency is the entry fee and judgement is the differentiator, telling people to learn to prompt is a harmless first step. My answer is that offering it as the whole answer does real harm, because it directs finite effort towards the part that is commoditising and away from the part that is appreciating.

What the durable version of the advice would be#

Autor and Thompson give the cleanest way to see why the advice misfires. Across four decades and 303 occupations, they found that automation removing the less expert tasks from a job raised wages, while automation removing the expert tasks lowered them. Value follows the expertise content of what remains. Prompting is not the expert content of anybody's job. It is the interface to the tool that is removing content, and getting better at the interface does nothing to change which direction your role is moving.

So the durable version of the advice is different in kind, not in degree. Learn to know what to ask for, which is problem definition and is the hardest part of most professional work. Learn where the frontier runs in your own domain, meaning the specific categories of problem where the model sounds most confident and is most often wrong, because that knowledge is local, non-transferable and does not commoditise. And learn to be accountable for the output, which is the one thing that cannot be delegated to the system at all.

Put crudely: prompting is asking well. The scarce skills are knowing what is worth asking, and knowing when the answer is wrong. AI makes the first of those cheaper every year and the other two more valuable every year, and the advice everyone is repeating points at the wrong one.

What to do instead#

Spend an afternoon on prompting, not a career. Get competent, then stop optimising it. The marginal return on your two-hundredth hour of prompt craft is close to zero and falling.

Build your frontier map. Keep a running note of where the model has been confidently wrong in your specific domain. After six months this is a rare asset and nobody can copy it from you.

Practise problem definition deliberately. Write the brief before you write the prompt: what is actually being decided, what would count as a good answer, what constraints are non-negotiable. This is the work that survives.

Keep the reps that build the judgement to verify. You cannot check an answer in a domain where you never built competence. See how humans learn with AI and using AI without dependency.

Make your judgement visible. When anyone can produce competent-looking output, competent-looking output stops being evidence. See staying valuable in the age of AI.

Development of the idea#

I have argued the underlying position since well before the current wave. In the Observer in July 2026 I made the case that the next AI power class will not build the models, and in Entrepreneur UK the same month that AI does not create bad decisions, it exposes them faster. Knowledge Is No Longer Power develops the argument about where the human edge moves once information and fluency are free. The framework is in SuperSkills (Kogan Page, 2026).

Key research and primary sources

• Zamfirescu-Pereira, J. D., Wong, R. Y., Hartmann, B. and Yang, Q. (2023). Why Johnny Can't Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts. CHI 2023.
• Bick, A., Blandin, A. and Deming, D. J. (2024). The Rapid Adoption of Generative AI. NBER Working Paper 32966.
• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier. Harvard Business School and BCG working paper.
• Vaccaro, M., Almaatouq, A. and Malone, T. (2024). When combinations of humans and AI are useful: a systematic review and meta-analysis. Nature Human Behaviour, 8, 2293-2303.
• Autor, D. and Thompson, N. (2025). Expertise. NBER Working Paper 33941; published in the Journal of the European Economic Association, 23(4).
• Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889-942. Earlier version NBER Working Paper 31161.

Related SuperSkills research#

On the career question, staying valuable in the age of AI and will AI replace my job. On verification as the scarce work, the verifier's discount and human and AI decision making. On what to build instead, human skills in the age of AI.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. This page is part of a deliberate disagreement layer: it argues against a widely repeated position, states the strongest case for that position before answering it, and marks where the evidence could go the other way. Findings are attributed to the studies that produced them and kept separate from the interpretation, which is the author's.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.