The AI skills to learn are mostly not technical. The research points to a mix: enough working knowledge of artificial intelligence to use it with care, deep expertise in your own field, social skills that machines handle poorly, and the discipline to keep thinking for yourself while a machine offers ready answers.
That answer runs against four assumptions that dominate most advice on the subject. The first is that the valuable skill is the technical one, so everyone should learn to code or to write clever instructions for a chatbot. The second is that artificial intelligence levels the playing field, so years of expertise in a field matter less once a machine can produce expert sounding answers. The third is that people skills are pleasant extras, a soft topic for a training day rather than something employers pay for. The fourth, and the most dangerous, is that using artificial intelligence while you learn is the same as learning.
Each assumption contains a grain of truth, which is why it survives. Technical skills with artificial intelligence do carry a pay premium. Less experienced workers sometimes do gain the most from these tools. Yet when you strip each claim back to what the studies actually measured, a different picture appears. The pay premium for technical skills depends heavily on what those skills are combined with. Expertise turns out to be the thing that catches the machine's mistakes. Social skills have been rising in value for four decades and show no sign of stopping. And several careful experiments now show that people who lean on artificial intelligence while practicing perform better in the moment and worse once the tool is taken away.
Why Do Most Lists of AI Skills to Learn Miss the Point?
Most popular lists of skills for the age of artificial intelligence share a flaw. They treat skills as items on a shelf, each with a fixed value, and simply rank the items. Prompt writing goes near the top. Coding goes near the top. Data literacy goes near the top. Then, almost as an afterthought, comes a line about communication or creativity.
The trouble is that a skill does not have a fixed value. Its worth depends on what else the person can do, on the job they are doing, and on whether the skill is something a machine can now do cheaply. A skill that a large language model (software trained on vast amounts of text to predict and produce language) can perform in seconds is unlikely to command a premium for long, however fashionable it is today. A skill that makes the machine's output more useful, or that catches its errors, is a different matter.
There is a second gap. Lists of skills to learn rarely say anything about how to learn them. That omission matters more now than it ever has, because the same tools that people are told to master can quietly undermine the learning process itself. A worker who asks a chatbot to draft every analysis may produce better looking work this month and be less capable next year. The skill list says nothing about this, yet the evidence on it is among the strongest in the whole field.
A third gap is the assumption that everyone gains equally. Some of the best designed field experiments show the opposite: the same tool, offered to everyone, widened the gap between strong and weak performers. Who gains depends on judgment, on knowing which advice to act upon, and on an accurate sense of one's own limits. Those are skills too, and they rarely appear on any list.
So the useful question is not "which skills are hot?" It is "which skills become more valuable when combined with artificial intelligence, and how do I build them without letting the tool do my thinking for me?" The rest of this article answers that question, study by study.
What Does the Labor Market Actually Pay For?
Start with the money, because pay is the most honest signal of what employers value. An early large study of online job advertisements in the United States tracked demand for artificial intelligence skills from 2010 to 2019. That labor market study found that demand rose sharply across almost every industry and occupation, led by information technology, then engineering, science and management roles. Job postings that asked for artificial intelligence skills offered pay about 11% higher than other postings at the same firm, and about 5% higher than other postings with the same job title. The largest premium was in management roles, not in purely technical ones. Firms that wanted these skills also tended to pay more for their other jobs, which hints that part of the premium reflects the kind of firm doing the hiring rather than the skill alone.
That finding about management deserves attention. It suggests that the market was rewarding people who could combine an understanding of artificial intelligence with responsibility for decisions, budgets and people, rather than rewarding technical skill in isolation.
A study published in 2024 made this idea of combination its central question. Working across 962 distinct skills, research on skill complementarity argued that the value of any skill depends mostly on how many other valuable skills it can be combined with. A skill's worth also depends on the worker's existing background, and for most skills the value is highest when paired with skills of a different type. Applied to artificial intelligence, the study estimated that these skills raise wages by about 21% on average, and it attributed that premium to how well they combine with many other skills, not to scarcity alone. Put plainly: a skill in artificial intelligence is worth most to someone who already brings something else to the table.
More recent evidence from the United Kingdom points the same way and adds a twist about formal qualifications. Drawing on about 11 million online job vacancies between 2018 and mid 2024, a 2025 vacancy analysis found that artificial intelligence roles grew by 21% as a share of all postings between 2018 and 2023. Over the same period, mentions of university degree requirements for these roles fell by 15%. Skills in artificial intelligence carried a wage premium of about 23%, which was larger than the premium attached to any degree below doctorate level (the doctorate premium was about 33%). Employers, in other words, were paying for demonstrated skill more than for the credential.
These three studies share a limitation worth stating. Job advertisements show what employers say they want and what they offer, not what they eventually pay or how the hired person performs. Advertised wage premiums can also shrink as more workers acquire the skill. Still, the direction is consistent across countries, data sets and years: artificial intelligence skills command a premium, and the premium is largest when those skills are combined with other expertise.
Which Skills Rise in Value Alongside Artificial Intelligence?
If technical skills pay most when combined with something else, the next question is what that something else is. A working paper that has not yet completed peer review offers the most direct answer so far. Using nearly 30 million job postings from the United States, the United Kingdom and Australia between 2018 and 2024, one large working paper asked whether adopting artificial intelligence increases demand for human abilities that complement it.
Its findings fall into three parts. First, jobs that involve substantial work with artificial intelligence are more likely to ask for analytical thinking, resilience (the capacity to keep working well under pressure and setbacks), digital literacy and ethical judgment. Second, these complementary abilities are linked to meaningful pay gains, especially in management, sales and finance roles that involve working with artificial intelligence. Third, the effect spreads beyond jobs that are about artificial intelligence. As firms, industries and regions adopt more of it, demand for these complementary abilities rises in ordinary roles too, while demand for skills that machines can more easily substitute, such as summarizing, translation and routine customer service, declines. The patterns held in all three countries.
Two cautions apply. This is a working paper, so its methods have not yet been fully tested by independent reviewers. And like the studies above, it relies on what employers write in advertisements, which can reflect fashion as much as need. Even so, it fits neatly with the pay evidence. The skills that hold or gain value are the ones that help a person decide what to do with a machine's output: reasoning about whether it is right, persisting when it is wrong, and judging whether it should be used at all.
Notice what drops in value: summarizing and translation. These are exactly the tasks that a large language model performs well. A worker whose main contribution is turning long text into short text is competing directly with the tool. A worker whose contribution is deciding which summary matters, and why, is working alongside it.
Why Does Domain Expertise Matter More, Not Less?
A common claim holds that artificial intelligence democratizes expertise. If a machine can produce a competent legal memo or a decent market analysis, why spend years learning the field? The evidence suggests the opposite. Expertise is what allows a person to see where the machine has gone wrong, and machines go wrong in systematic ways that only an expert notices.
An early and careful demonstration came from patent examination, the work of deciding whether an invention is new enough to deserve a patent. A patent examination experiment found that a machine learning tool (software that learns patterns from past data) used to search for earlier, similar inventions was biased toward finding text that looked similar to the new application. The most relevant earlier inventions are not always the ones that use the same words. Finding them required domain expertise, which corrected the tool's bias. The study concluded that domain knowledge and machine learning are complements: the tool is worth more in expert hands.
Medicine offers a sharper test. In an experiment with radiologists, a working paper from the National Bureau of Economic Research that has not yet completed peer review, professional radiologists read medical images with and without predictions from an artificial intelligence system. On average, the predictions did not improve their accuracy. Additional clinical information about the patient did help. The radiologists failed to gain from the machine because they made predictable errors in combining its advice with their own judgment: they gave the machine's prediction too little weight, and they treated their own reading and the machine's reading as if they were unrelated pieces of information. The authors concluded that, unless those errors could be corrected, it was often better to send a case to the radiologist alone or to the machine alone than to a radiologist assisted by the machine.
Read together, the patent and radiology studies make a strong case. The value a person adds to a machine depends on knowing the domain well enough to be better than the machine at something, and on knowing when to trust it and when not to. Expertise is not made obsolete by artificial intelligence. It is the very thing that decides whether working with artificial intelligence helps or hurts.
Why Do Social Skills Keep Gaining Ground?
Social skills are often filed under "soft," as though they were less real or less valuable than technical ones. The labor market data have said otherwise for a long time, well before the arrival of large language models.
The economist David Deming showed this in a landmark 2017 study. Between 1980 and 2012, the share of the United States workforce in jobs requiring high levels of social interaction rose by nearly 12 percentage points. Over the same years, jobs that were heavy on mathematics but light on social interaction, including many science and engineering roles, shrank by 3.3 percentage points. Jobs that demanded both strong mathematical and strong social skills saw the fastest growth in employment and pay. The pay reward for social skills was much larger in the 2000s than it had been in the 1980s and 1990s. The explanation the study offered is simple: social skills lower the cost of coordinating with others, which lets people specialize and trade tasks more easily. Computers had already taken over much routine work, but they remained poor at reading people, negotiating and adapting to a colleague's needs.
A later experiment showed what social skill does in practice. One study of teams randomly assigned people to several different teams and compared each team's performance with what the members' individual skills predicted. Some individuals consistently pushed their teams above that prediction. These team players scored higher on a well established test of social intelligence, the ability to read other people's thoughts and feelings. They did not differ from others in general intelligence, personality, education or gender. When the researchers combined social intelligence and the team player effect into a single measure of social skill, it improved team performance about as much as general intelligence did. There was also suggestive evidence that team players raised the effort of the people around them.
Why should this matter more as artificial intelligence spreads? Because coordination does not go away when machines do more of the individual work. It grows. Someone still has to decide who does what, persuade colleagues to change a process, resolve disagreement about a machine's recommendation, and keep a team motivated. The labor market data above, which show complementary human abilities gaining value as firms adopt artificial intelligence, are consistent with that view. A fair caution is that the social skills research predates the newest language models, so we do not yet have long run evidence on how those models change the reward for social skill. The direction of four decades of data, however, is hard to ignore.
Who Gains Most From Artificial Intelligence Advice, and Why?
A hopeful story holds that artificial intelligence helps weaker performers most and so closes skill gaps. Newer evidence complicates that story in useful ways.
The clearest complication comes from a field experiment with several hundred small business owners in Kenya, published in 2026. The entrepreneur field experiment randomly gave some entrepreneurs access to a business advice assistant built on a large language model, delivered through a messaging application. On average, the assistant made no detectable difference to revenue or profit. That average hid a split. Entrepreneurs who were already performing well at the start may have gained more than 15% from the advice. Those who were performing poorly did nearly 10% worse. The difference did not come from the questions they asked or the quality of the advice they received. It came from which pieces of advice they chose to act upon. The strong performers had the judgment to pick the suggestions that fit their business. The weak performers did not, and the machine's confident advice led some of them astray.
A controlled experiment adds a further piece. Research on calibration found that artificial intelligence helped people with lower baseline ability more, which fits the earlier hopeful story. But holding ability constant, the assistance was worth more to people who were well calibrated, meaning they had an accurate sense of how good they actually were. People who knew they had low ability gained the most of all. The researchers estimated that if everyone were well calibrated, artificial intelligence would reduce inequality in performance nearly twice as much as it already did.
These two studies do not contradict each other as much as they first appear to. The second one shows that lower ability people can gain the most, but only if they know where their weaknesses lie. The first shows what happens when people lack the judgment to sort good advice from bad: the gains go to those who already know their business. In both, the deciding skill is not technical. It is judgment about the advice and honesty about oneself.
Can Using Artificial Intelligence Stop You From Learning?
This is the question that most skill lists skip, and it may be the most important one. The concern is not that artificial intelligence makes people lazy in some general moral sense. It is narrower and better supported: learning requires effortful thinking, and a tool that removes the effort can remove the learning with it.
The psychology behind this is old and well tested. A 2007 meta analysis of the generation effect combined 86 studies and found that people remember information better when they produce it themselves than when they simply read it, with a moderate and reliable advantage. Researchers have also studied cognitive offloading, which a 2016 review defined as using physical action or an outside aid to reduce the mental demands of a task, such as writing a reminder rather than remembering. Offloading is often sensible. The question is what happens to the skill when the offloading becomes total.
A mathematics field experiment, published in 2025, gave a direct answer. Nearly 1,000 high school students practiced mathematics with one of two tutors built on a large language model, or with no tutor at all. One tutor behaved like a standard chatbot. The other was designed with guardrails: it gave hints and guided students toward answers rather than handing them over. During practice, both groups did far better than students without access. Grades rose 48% for the standard chatbot group and 127% for the guarded tutor group. Then the tools were taken away for an exam. Students who had practiced with the standard chatbot scored 17% lower than students who had never had access at all. The guardrails largely prevented this harm. The students with the standard chatbot had used it as a crutch, leaning on it for answers rather than working through problems, and their apparent progress during practice did not survive the removal of the tool.
The same pattern appears in adult learners. A randomized writing experiment assigned 117 university students to complete a writing task with help from a chatbot, a human expert, writing analytics tools, or no extra support. The chatbot group improved their essay scores the most. Yet they gained no more knowledge than the other groups and were no better at applying what they had learned to new tasks. The researchers found that the chatbot group went about planning, monitoring and reviewing their work in different patterns from the other groups, and they warned that the tool may breed dependence and what they called metacognitive laziness, meaning a tendency to hand over the thinking about one's own thinking. The sample was small and the task was a single session, so the finding needs replication, but it lines up with the larger mathematics experiment.
The lesson for anyone building skills is stark. Better output during practice is not evidence of learning. If the machine produced the answer, you may have learned how to ask the machine, which is a much narrower and more easily replaced skill than understanding the problem.
What Happens to Critical Thinking When the Machine Does the Thinking?
If learning can suffer, what about the thinking skills that experienced professionals already have? Here the evidence is younger and mostly based on surveys, so it supports careful conclusions rather than alarming headlines.
A knowledge worker survey, presented at a leading peer reviewed computing conference in 2025, asked 319 professionals to describe 936 real examples of using generative artificial intelligence at work. The more confidence a worker had in the tool's ability to do a task, the less critical thinking they reported applying to it. The more confidence they had in their own ability, both to do the task and to evaluate the machine's output, the more critical thinking they reported. Workers described their effort shifting away from gathering information and producing work and toward checking, combining and supervising what the machine produced. The study is based on self reports and cannot prove cause, but it identifies a useful mechanism: trust in the tool and trust in oneself pull critical thinking in opposite directions.
A 2025 British survey of 666 people found that more frequent use of artificial intelligence tools went together with lower scores on a measure of critical thinking, and that this link ran largely through cognitive offloading. Younger participants relied more on the tools and scored lower on critical thinking. Higher education was linked to better critical thinking regardless of tool use. This was a single correlational survey with self reported measures, so it shows that the two things move together, not that one causes the other. People who think less critically may simply be more inclined to lean on the tools. It should be read as a warning sign, not a verdict.
These findings connect directly to the earlier result on calibration. Gains from artificial intelligence depend on knowing your own strengths and weaknesses, and leaning on the machine without checking yourself is no way to learn where those lie. The skill that protects against this is deliberate self checking: forming your own view before asking the machine, comparing the two, and keeping track of when you were right.
Which AI Skills to Learn First, and How Should You Learn Them?
If you are deciding where to spend your learning time, the evidence suggests an order of priority that looks quite different from the usual lists.
Begin with depth in your own field. The patent and radiology studies point the same way: you add value to a machine only where you are better than it at something, and you can only judge its output in a field you understand well. If you are early in your career, this means resisting the temptation to let the tool do the work that builds expertise. Do the analysis yourself first, then compare it with the machine's version.
Next, learn enough about artificial intelligence to use it with judgment. You do not need to become an engineer. You need to understand what these tools do well (summarizing, drafting, generating options) and where they fail (confident errors, bland and similar output, weakness on problems that need context they do not have). The wage studies show a real premium for these skills, and the premium is largest when you combine them with something else you already do well.
Then invest in social skills. Coordinating people, persuading, negotiating and reading a room have gained value for four decades. A team that uses artificial intelligence still needs someone to decide who does what and to settle disagreements about what the machine recommended.
Finally, protect your own thinking. Practice without the tool when you are learning something new. Write your own first draft before asking for the machine's. Predict an answer, check it, and keep an honest record of when the machine was right and when you were. These habits build exactly the calibration that the experiments show makes the difference between gaining from artificial intelligence and being misled by it.
Key Takeaways
- Employers do pay more for skills in artificial intelligence, but the reward is largest when those skills sit on top of other expertise, particularly management and judgment, rather than standing alone.
- Knowing your field deeply is what lets you catch a machine's systematic errors; where the human is weaker than the machine, adding a human to the loop can make results worse.
- Social skills have earned a growing premium since 1980, and people with high social intelligence lift the performance of whole teams about as much as raw intelligence does.
- The same advice tool can help strong performers and harm weak ones; the deciding factors are which advice a person chooses to act upon and how accurately they judge their own ability.
- Practicing with a chatbot that hands over answers can raise scores during practice and lower them once the tool is gone, while guided tools that offer hints instead largely avoid this harm.
- Greater trust in the machine goes with less critical thinking, while greater trust in your own ability to check its work goes with more.
Implications for Practice
For human resources leaders designing development programs, the first implication is to stop treating training in artificial intelligence as a standalone technical topic. The pay evidence shows that the premium comes from combination. Build programs that teach artificial intelligence inside real work in each function, so that a finance analyst learns it through finance problems and a recruiter through recruitment problems. This keeps domain expertise at the center, which the patent examination and radiology findings show is where value is created.
Second, protect the conditions for learning. The mathematics and writing experiments show that unrestricted access to a chatbot during practice can produce impressive output and weak learning. For new hires, graduate trainees and anyone acquiring a core skill, design learning tasks so that people attempt the work before consulting the tool, and where possible use tools configured to give hints rather than full answers. Assess learning with the tool removed, because performance with the tool is not a reliable measure of capability.
Third, invest in calibration. The entrepreneur and calibration experiments show that gains depend on judging which advice to act upon and knowing one's own limits. Build routines in which people record their own judgment before seeing the machine's output, then review where each was right. Managers can make this part of regular feedback conversations, which also counters the overconfidence that the metacognition experiments uncovered.
Fourth, give social skills a place in selection and development that reflects their market value. Coordination, persuasion and the ability to raise the performance of a team are rising in worth, and they are skills machines do poorly. Treat them as core capabilities, not as optional extras on a training calendar.
