AI coding assistants are reshaping software engineering rather than replacing it: firms are substituting automation for routine junior work—cutting entry-level hiring—while value and risk move toward architecture, verification and governance; unreviewed AI-generated code raises measurable security and technical-debt concerns.

Main Finding

AI coding tools are reshaping software engineering rather than eliminating it: routine implementation is increasingly automated, value shifts toward architecture/verification/accountability, entry-level hiring is contracting, and unreviewed AI-generated code raises measurable security and technical-debt risks. The net effect is a reallocation of labor and skills within the software sector with short-term contraction for junior roles and medium-term growth for judgment- and AI-oversight–intensive roles.

Key Points

• Productivity effects are heterogeneous:
  • RCTs and trials report results from a ~19% slowdown (early-2025 tools) to an ~18–56% speed-up (newer models or bounded tasks), depending on model generation, task scope, and developer experience.
  • AI helps most on routine tasks (reported ~30–60% time savings in some surveys) and least on complex, unfamiliar systems.
  • Perceived speed-ups often exceed measured gains.
• Employment and skill formation:
  • Payroll data (2021–2025) show ~20% decline in employment for developers aged 22–25 (late-2022 to mid-2025), while experienced developer employment rose (~6–12%).
  • Job postings remain below pre-pandemic levels, with sharper falls for junior roles.
  • New roles (AI-governance, AI-validation, hybrid oversight) command wage premiums; AI fluency pays more in entry-level roles that require evaluation/directing of AI.
  • Risk of skill atrophy: automating code generation can weaken apprenticeship pathways because evaluating AI output often requires hands-on generative experience.
• Code quality, security, and technical debt:
  • Large-sample security audits find only ~55% of AI-generated samples free of known security issues; earlier Copilot evaluations found ~40% of generated code contained vulnerabilities.
  • Merged AI contributions without adequate review are associated with higher code churn, mismatched dependencies, architectural debt, and more maintenance costs later.
  • AI shifts effort from writing to evaluating code, and evaluation is often more error-prone and context-dependent.
• Opportunities:
  • Lowers barriers for domain experts to prototype (“vibe coding”).
  • Frees experienced engineers to focus on architecture and product judgment if organizations deliberately reallocate freed time.
  • Creates premium roles around AI oversight, governance, and review pipelines.
• Policy and organizational recommendations (summarized):
  • Mandate human review of AI-generated code for critical systems; redesign junior roles to preserve skill formation; keep foundational CS education AI-independent while adding AI-collaboration training; support transparency, reskilling, and independent research.

Data & Methods

• Approach: structured narrative synthesis combining randomized trials, payroll-based labor analyses, industry surveys, and code-security audits (preprints, peer-reviewed studies, and institutional reports).
• Key empirical sources and findings referenced in the paper:
  • METR RCTs with experienced developers: early-2025 model slowed performance (~19%); follow-up with newer models showed ~18% speed-up.
  • Peng et al. RCT on GitHub Copilot on bounded tasks: ~56% faster completion for assisted group.
  • Industry surveys (McKinsey, Google DORA) reporting sizeable routine-task savings but mixed effects on delivery stability.
  • Payroll/administrative data showing ~20% fall in employment among 22–25-year-old developers (late-2022 to mid-2025) and growth for experienced engineers.
  • Security audits across >100 LLMs: only ~55% of samples free of known security issues; earlier Copilot studies found ~40% of generated code with vulnerabilities.
  • Production-repository analyses linking AI-assisted merges to higher churn and architectural debt.
• Limitations acknowledged:
  • Evidence base is young and rapidly changing; many labour-market findings are correlational (not strictly causal).
  • Heterogeneity by model generation, task type, developer experience, and organizational practices reduces external comparability.
  • Formal meta-analysis was not feasible due to differing designs; recommendations are based on synthesis rather than pooled causal estimates.

Implications for AI Economics

• Labor reallocation, not net elimination (short-to-medium run):
  • Demand shifts from routine implementation to higher-skilled roles (architecture, verification, AI oversight). This produces heterogeneous wage and employment effects: contraction for junior, growth and wage premiums for AI-adjacent senior roles.
• Human capital formation risk:
  • If junior hiring and hands-on coding opportunities shrink, long-term supply of senior engineers with deep evaluative judgment may fall, raising future labor scarcity and wage pressures for senior talent.
• Hidden/shifted costs and productivity accounting:
  • Measured time savings on coding can mask increased downstream maintenance, review costs, and delivery instability. Economic evaluation of AI should pair productivity metrics with quality, churn, and defect costs.
• Market structure and task bundling:
  • Firms may substitute AI for junior hires where budget-constrained, concentrating judgment tasks among fewer, more expensive senior staff or new specialist roles—altering team composition and career ladders.
• Policy and redistribution considerations:
  • Public investment in reskilling, apprenticeship replacements, and independent security research can mitigate short-term dislocation and longer-term skill-supply risks.
  • Regulation or standards that require transparency and mandatory review for AI-generated code in safety-critical domains can internalize risks and change firm incentives.
• Research priorities for economists:
  • Causal identification of AI’s effect on employment and wages, longitudinal studies of skill development under AI use, and firm-level evaluations of organizational practices that best capture AI benefits while containing risks.
• Strategic implication for firms and policymakers:
  • To realize productivity gains without increasing systemic risk, organizations must invest in verification pipelines, training, and deliberate hiring practices; policymakers should support metrics, transparency, and programs that preserve human capital formation.

If you want, I can (a) extract a one-page bulletable policy brief for employers or (b) produce a short table mapping research gaps to specific empirical designs economists could use. Which would you prefer?