The New Skill Premium

If a job advertisement asks for analytical thinking alongside AI skills, that tells us something about what an employer wants. It does not tell us that AI caused the requirement, that the applicant will use AI every day, or that taking an analytical-thinking course will produce a pay rise.

Those distinctions matter when turning labour-market research into career advice. A worker needs a way to decide what to learn. A striking headline can be a useful lead, but it is a poor employment contract.

What the postings actually say

Complement or substitute? How AI increases the demand for human skills, by Elina Mäkelä and colleagues, examines skill requirements in online job advertisements. I use the June 8, 2026 revision here. The analysis concerns postings from 2018 to 2024, not a live snapshot of the 2026 labour market.

Its within-role analysis uses 9,998,342 US postings. The broader company, industry and regional analyses extend to about 30 million postings; the methods section describes coverage including the UK, Australia and New Zealand. This is a large sample of advertisements, not 30 million observed workers or completed hires.

The researchers identify AI-related roles through specified AI and machine-learning skill mentions. They organise other skills into literature-derived clusters labelled complementary or substitutable. Those labels define what they study. They are not experimental proof that every task in one cluster must remain human and every task in the other can be automated.

In the main adjusted within-role analysis, five of seven complementary skill clusters were more likely to appear in AI-related postings. Ethics and digital literacy did not have statistically significant effects in that comparison. The paper also finds broader associations between AI-related hiring and skill requirements in non-AI roles.

That is a more qualified result than “all human skills are rising because of AI.” Advertisements reveal employer language and intended requirements. The study does not randomly introduce AI into otherwise identical firms and observe their response. Its significance statement appropriately describes the evidence as correlative.

The wage claim needs its own check

The paper compares posted salaries using regression controls. In its pooled analysis of AI roles, the resilience skill cluster was associated with an approximately 5% salary premium. It says premiums for the other complementary clusters could not be confidently detected in that analysis. Descriptive salary differences for selected occupations are not interchangeable with that adjusted result.

A posted salary is not necessarily the salary someone eventually receives. Jobs that mention resilience may also differ in responsibility, employer, demands or other ways that the analysis cannot fully capture. The association is not a claim that learning resilience causes a 5% increase in pay.

Test the inference before using a wage headline

Supported: in this dataset and regression, postings for AI roles that mention skills in the resilience cluster have higher advertised salaries on average, after the included controls.

Not established: the same employee would be paid more after completing a course, the relationship applies equally in every country or occupation, or AI adoption caused the premium.

Worth investigating locally: what responsibilities the better-paid roles carry, how applicants demonstrate them, and whether current postings in your target market show the same pattern.

The distinction is not academic fussiness. A training budget, a job move and a hiring rule all have consequences. They deserve more than a plausible story pasted over an observational result.

Turn a broad skill into work someone can demonstrate

“Develop judgment” is difficult advice to act on. A smaller target is more useful: explain why a proposed answer is unsupported, identify the missing information, and decide what can responsibly happen next.

Consider a hypothetical analyst working with an AI-generated sales forecast. A useful development exercise would ask the analyst to check whether the inputs match the forecast period, identify a change in the customer mix, and explain why a past relationship might not hold. The analyst can use software while doing this. The skill is in testing the claim, not proving they can work without tools.

That exercise requires domain knowledge. Removing technical learning from a development plan because “AI does the execution” would make it harder to know what to question. Communication helps the analyst explain the uncertainty; it does not replace understanding the data.

For a manager, the equivalent improvement is to replace vague screening language with a relevant work sample. Ask candidates to inspect a flawed analysis and explain their decisions under clear conditions. Allow reasonable accommodations and consistent tool access. Do not use confidence, extroversion, or familiarity with fashionable terminology as a proxy for sound reasoning.

This is a proposed way to assess a particular job requirement. The posting study does not validate the exercise or establish that it predicts performance. That needs evaluation too.

Complementary does not mean immune

It is tempting to make a list of abilities AI will never replicate and advise everyone to move there. The evidence here cannot support that list. Models can already contribute to analysis, writing and idea generation. How much they contribute depends on the task, tools and standards, not on whether the skill's name sounds human.

Accountability is a separate matter. An organisation still has to assign responsibility for decisions and consequences. That requirement can justify a human role without proving that a person always outperforms a model at the underlying task.

The reverse claim is also too strong. Falling mentions of a skill in job postings do not show that existing workers no longer use it, or that every position containing it is disappearing. The skill may be assumed, renamed, bundled with something else, or supplied through another arrangement. A job advertisement is one measurement surface, not the job itself.

A development plan with room to change

I would start with the actual work a person wants to do, then inspect several sources of evidence: current job requirements, conversations with people doing the role, and samples of the work itself. Use the research to ask better questions, rather than treating its categories as a ranking of people.

Choose a capability that can be practised and observed. Keep the technical foundation needed to exercise it. Agree on time for that practice and review the result after a meaningful piece of work, not merely after a course-completion badge.

Employers have obligations in that transition too. If a role changes, name the new expectations, provide support and discuss workload and progression honestly. Telling people to become adaptable while giving them no time to learn is not a workforce strategy.

The new skill premium is a question to investigate in a particular market. The useful response is to build demonstrable capability, not to promise anyone an AI-proof career.

Source and revision note

Mäkelä and colleagues, Complement or substitute? How AI increases the demand for human skills, arXiv 2412.19754v4, June 2026. See Methods, Analysis of Internal Effects, Valuation of Skills and the supplementary classification details. This revision distinguishes advertised skill demand and salaries from causal effects on workers. It removes unsupported claims about guaranteed job transformation, permanently human-only skills and private organisational results.