AI Is Changing Professional Work: Which Skills Are Becoming More Valuable?

01 / The story

The developer paradox: faster and slower at once

The most cited experiment in this field is also the most misunderstood. In a July 2025 randomized controlled trial, METR gave 16 experienced open-source developers 246 real tasks in repositories they knew well. Allowing early-2025 AI tools increased completion time by 19%—yet the developers believed AI had sped them up. [1]

METR's February 2026 follow-up complicates the picture further: raw estimates were consistent with possible speedups, but the researchers judged them unreliable because developers increasingly declined to work without AI and concurrent agents complicated measurement. METR now explicitly warns that its 2025 result is historical, not a current verdict. [2]

On the other side, a February 2025 working paper covering randomized experiments with 4,867 developers at Microsoft, Accenture and another large company found completed tasks rose about 26% with AI assistance, with less-experienced developers gaining most. And a 2023 controlled study found a bounded coding exercise completed 55.8% faster with GitHub Copilot. Company involvement in both should be disclosed, and completed tasks are not the same as hours saved or long-term software quality. [3, 4]

The honest reading: AI assistance helps in some settings and hurts in others, and the outcome depends on the task, the codebase, the worker's experience and whether review time is counted.

02 / The story

Law, medicine and engineering: assistance, not autonomy

A preregistered randomized trial published as an NBER working paper in September 2026 gave 133 patent lawyers at 11 U.S. firms access to a custom drafting assistant. Blinded expert scoring found higher drafting quality at 10 days and 90 days. But on later unaided work, the average benefit was concentrated among senior lawyers; juniors showed no average improvement. Better assisted output is not the same as durable expertise. [5]

In medicine, a February 2025 Nature Medicine trial randomized 92 physicians on simulated management-reasoning cases with GPT-4 assistance—bounded reasoning help, not real-world autonomous treatment. A November 2025 NEJM AI trial of 238 physicians found one ambient AI scribe reduced note-writing time by 9.5% while another's 1.7% reduction was not statistically significant. Products and outcomes differ, and note time is not total clinical productivity. [6, 7]

Professional bodies are drawing the accountability line clearly. The American Bar Association's July 2024 Formal Opinion 512 holds that competence, confidentiality, client communication and reasonable fees remain the lawyer's obligations when using AI. The National Society of Professional Engineers' position statement, revised February 2026, advocates rigorous verification and professional accountability for safety-critical AI. These are professional obligations, not suggestions the tools can absorb. [8, 9]

03 / The story

Exposure is not automation, and automation is not job loss

Three different quantities get collapsed into one scary number. The U.S. Bureau of Labor Statistics, which published explicit AI exposure categories on August 27, 2026, states plainly that exposure does not imply job loss, productivity gains, automation probability or wage effects. [10]

The International Labour Organization's May 2025 global index estimates one in four workers is in an occupation with some generative-AI exposure, and 3.3% of global employment falls in the highest exposure category. These are occupational exposure estimates—not percentages of jobs already automated or expected to disappear. [11]

The most concrete employment evidence comes from the Stanford Digital Economy Lab. Using ADP payroll data through June 2026, its August 2026 update found employment among 22-to-25-year-olds in highly exposed occupations about 19% below a counterfactual keeping pace with less-exposed peers—with reduced hiring, not increased firings, as the main adjustment, and no widespread economy-wide displacement. This is observational evidence, not proof that AI explains every hiring change. [12]

04 / The story

Which skills are becoming more valuable

Synthesizing the experiments and the professional guidance, six skills stand out—an editorial judgment, not a measured wage premium.

Verification and evaluation: writing tests, checking citations, assessing whether an output is plausible. The heterogeneous experimental outcomes make this the most clearly supported need. [1, 2, 5]

Domain expertise and problem framing: the patent trial shows assisted performance and underlying expertise are different things, and seniors benefited most. [5]

Workflow design and measurement: matching tools to bounded tasks and measuring total time including review—knowing when not to use AI. [1, 2]

Communication and accountable decisions: explaining uncertainty, obtaining consent, resolving competing goals and taking responsibility. The ABA and NSPE positions make this a duty, not a preference. [8, 9]

AI and data literacy: the World Economic Forum's 2025 employer survey identifies AI, technological literacy and cybersecurity among growing skills—directional employer expectations, not guaranteed demand. [13]

Learning without dependence: maintaining unaided practice and mentorship, especially for juniors. The patent trial's junior results and the entry-level hiring evidence both point here. [5, 12]

05 / The story

What to watch

The open questions are empirical. Do the entry-level hiring gaps close as tools and training improve, or widen? Do productivity gains in company trials survive independent replication with review time counted? Does junior expertise erode when assisted work replaces unaided practice?

One cautionary tale for everyone covering this field: a widely circulated claim that AI made materials-science researchers dramatically more productive was disavowed by MIT in May 2025, which said it had no confidence in the data and requested the paper's withdrawal. In a field this noisy, the skill that matters most may be the one this article tries to model: check the source before believing the number. [14]