AI and the Labour Market, Signal, Noise and Structural Change

Narrative

Practitioner evidence shows rapid adoption of AI-assisted development but not a clean measure of labour-market displacement. DORA's 2024 survey of technology workers found that more than 75% used AI for at least one daily responsibility and associated higher adoption with reported improvements in documentation, code quality and review speed, while also associating it with lower delivery throughput and stability. Its 2025 follow-up used survey responses from nearly 5,000 technology professionals and more than 100 hours of qualitative material, concluding that AI amplifies pre-existing organisational strengths and weaknesses rather than producing an automatic performance gain.

The immediate work redesign is clearer than any headcount effect. Thoughtworks, Martin Fowler contributors and InfoQ practitioners describe a movement from writing implementation line by line towards specifying intent, supplying organisational context, validating output and governing agents. These sources repeatedly identify testing, architecture, security review, observability and shared context as new or enlarged constraints. They also warn that measures such as generated lines, commits and pull-request volume can rise without proving better delivery, quality or customer outcomes.

Distributional signals are mixed and mostly perceptual. Stack Overflow's 2024 survey found lower trust among professional developers than among people learning to code, while its 2025 survey reported that 52% of respondents viewed AI tools or agents as positive for productivity and that agent use remained far from universal. Practitioner accounts argue that junior staff may gain faster access to unfamiliar codebases, but that mentorship, verification and deliberate unassisted learning remain necessary; this is a work-design hypothesis, not evidence that junior vacancies have fallen. GitHub and CNCF materials add evidence of tool uptake and case studies, but they are vendor surveys or organisational examples, not representative employment, wage or redundancy statistics.

For the wider labour-market question, this lane supplies leading indicators rather than causal estimates: adoption in daily engineering work, changes in the bottleneck from coding to specification and verification, review burden, skill requirements for context engineering, and platform investment. It provides no credible basis to attribute aggregate tech employment, vacancies, wages or productivity growth to AI rather than interest rates, post-pandemic correction, firm restructuring or the business cycle. The strongest operational implication is to assess AI with end-to-end delivery, reliability, quality, incident and workforce-development measures, with explicit pre-adoption baselines, rather than use activity metrics as proxies for labour savings.