Auto-generated, not human-reviewed

Skill Acquisition

Updated Jun 14, 2026

Papers 195 (127 full-text)

Claims 316

Evidence strength: Mixed: reliable short-run performance gains; learning and independent-skill development effects are inconsistent or negative; most long-run evidence is observational.

Bottom Line

AI assistance reliably boosts immediate task performance. Several randomized and controlled studies also find reduced persistence and weaker later unassisted performance after brief exposure; a programming meta-analysis finds no consistent learning gains Liu et al. (2026); Gardella et al. (2026); Maier et al. (2026).

Observational and natural-experiment studies (using policy shocks) link AI exposure to higher demand for complementary human skills (analytical thinking, teamwork, resilience) and advanced digital skills, and to lower demand for routine cognitive skills Stephany et al. (2026); Zhang & Zhang.

What This Means in Practice

Separate AI-for-output from AI-for-learning. Require some AI-free practice, add short reflective debriefs, and test unassisted performance after AI use. Even minutes of exposure can reduce persistence and later unassisted performance while lifting immediate results Liu et al. (2026); Gardella et al. (2026); Maier et al. (2026).
If you use AI in training, add human coaching and gate access. Instructor-controlled large language model (LLM) support and targeted conversational coaching improve writing and communication, with effects that depend on user traits Wang et al. (2026); Kumar et al. (2026); Duddu et al. (2026).
When revising roles or curricula, emphasize complementary human skills (analysis, collaboration, resilience) and advanced digital literacy. Demand for these rises in AI-intensive and adjacent roles Stephany et al. (2026); Wang et al. (2026).
If you fund AI adoption, also fund broadband, devices, and training. Better connectivity supports AI uptake and ICT (information and communications technology) role growth, but gender, regional, and organizational gaps in advanced digital skills persist Bilgin & Ottaviano; Whelan et al..
If entry-level work is thinning, rebuild early-career formation. Use structured apprenticeships and supervised AI use to preserve mentoring, feedback, and judgment development De Crescentis & Baker (2026); Rosenthal & Iqbal (2026).

What the Research Finds

1) Immediate AI assistance vs. learning and persistence

Randomized and controlled studies find AI assistance raises short-run task performance but reduces persistence and can impair later unassisted performance after minutes of exposure Liu et al. (2026); Gardella et al. (2026).
In education, randomized ChatGPT access raises short-term knowledge scores, but a programming meta-analysis finds no consistent learning gains (pooled effect ~0.14, not statistically significant). Heavy on-demand use is associated with weaker learning Conde et al. (2026); Maier et al. (2026).
Providing high-information AI boosts immediate performance without average post-AI harm, but effects vary widely across users Wu et al. (2026).
Generative AI (GenAI) can dampen motivation and distort self-judgment. An experiment finds lower intrinsic motivation and self-perceived creativity with GenAI; reviews suggest LLM use can reduce the accuracy of self-assessment and narrow confidence gaps between novices and experts Endres et al. (2026); Koch (2026).
Brief, personalized AI coaching can build specific interpersonal skills. A preregistered randomized trial shows higher empathic communication after short LLM coaching, but the same messages labeled as AI receive lower validation ratings; effects differ by personality Kumar et al. (2026); Duddu et al. (2026).

2) AI is shifting skill demand toward complements and advanced digital capabilities

In the US, UK, and Australia, AI-intensive roles show higher demand for analytical thinking, resilience, teamwork, and digital literacy; nearby non-AI roles move in the same direction Stephany et al. (2026).
Better connectivity and digitalization are associated with more AI adoption, growth in ICT roles, and shifts toward higher-skill workers; firm digital transformation is linked to higher labor demand and improved employee digital literacy Bilgin & Ottaviano; Fan (2026); Wang et al. (2026).
In 67 million postings in China, AI augmentation exposure is associated with higher shares of nonroutine analytical skills, while displacement exposure is associated with lower shares of routine cognitive skills Zhang & Zhang.
On a logo-design platform, embedded GenAI shifted requested skill diversity, partly via stronger freelancer competition Sun & Liu (2026).
Germany shows a move toward frontier skills, and wind energy postings highlight sector-specific digital priorities: scientific programming and numerical modeling are core; machine learning, IoT, and cybersecurity vary by role Genz et al. (2026); Stoltman et al. (2026).

3) Early-career formation, on-ramps, and how AI changes pathways

Entry-level functions that once supported apprenticeships are compressing. Interviews indicate AI is disrupting informal mentoring and feedback channels that support career growth De Crescentis & Baker (2026); Rosenthal & Iqbal (2026).
After GenAI rollouts, an AI-like practice pattern spread. It predicts smaller rating gains in unsupervised contests, but higher non-AI scores among applicants screened through AI-prohibited gates Yao (2026).
Adopting an AI coding assistant coincides with persistent increases in developer activity and expansion into new languages. Copilot-adopting firms hire more entry-level software engineers, and new hires list more non-programming skills without coding-skill declines Quispe (2026); Baird et al..
Large US retraining programs seldom move participants into less-automatable roles; wage recovery often reflects returns to similar work. Among options analyzed, apprenticeships perform best Jacobs & Canedy (2026).
Freelancers use GenAI to structure short-term learning but treat it as supplementary and inconsistent; they face invisible competencies that are hard to signal and shift from learning as growth to survival Imteyaz et al. (2026).

4) Equity, access, and who can acquire AI-era skills

Across Europe, women are about 15 percentage points less likely than men to perform advanced digital tasks; the gap is largest at the top tail, 26 points in Ireland, and only about 30 percent is explained by observables Whelan et al..
Low-income regions lag in education and digital access; countries such as Nigeria and Uzbekistan report acute shortages in advanced AI competencies and mismatches between education and employer needs Horobets; Ovili et al. (2026); Rajabovna.
Only 24.4 percent of at-risk workers in an Egyptian jobs graph have viable skill-overlap pathways to safer roles Dawoud et al..
On-premise retrieval-augmented generation (RAG) can match cloud options on task quality but requires in-house machine learning operations (MLOps) and engineering. Among small US firms in underserved communities, higher AI adoption is associated with better financial literacy and profit growth but is constrained by training disparities You & Hong; Akuffo.

5) Collaboration design can prevent or cause deskilling

Models and simulations show that even rational AI adoption can erode human skills over time; mandatory practice or occasional induced AI failures can preserve capability in simulations Caosun & Aral (2026); Park et al. (2026).
Treating AI retrieval tools as external memory encourages unbounded note-taking without abstraction; adding consolidation mechanisms may better support long-term learning Xu et al. (2026).
People adjust to AI confidence when confidence tracks accuracy, but many fail when it does not Li & Steyvers (2026).
User fluency shapes outcomes: fluent users attempt harder tasks, expose visible failures, and solve more hard problems; novices risk invisible failures that look successful Potts & Sudhof (2026).

What We Still Don't Know

We lack causal, long-run field evidence on how AI assistance affects workplace skill trajectories. Most evaluated trainings report short-term learner gains but no organization-level results or task reallocation (Kirkpatrick Level 4) Woods et al. (2026); Maier et al. (2026).
Which collaboration guardrails (for example, mandatory practice quotas, delayed reveals, scaffolds) preserve human capability across job families remains untested outside simulations and lab tasks Park et al. (2026); Liu et al. (2026).
The net career impact of redesigned apprenticeships and structured early-career pathways in AI-rich firms has not been causally evaluated at scale De Crescentis & Baker (2026).
Beyond Europe and selected case studies, we lack comparable, longitudinal measures of AI-era upskilling by gender, age, region, and firm size tied to hiring, wages, and task exposure Whelan et al.; Horobets.
The returns to complementary human skills in non-tech sectors under real AI adoption, on pay, promotion, and mobility, are still mostly inferred from postings rather than linked to realized worker outcomes Stephany et al. (2026).