Developer Productivity in the Age of Generative AI: A Psychological Perspective
Citation. Edwards, R., & Appiah, R. (2026). Developer productivity in the age of generative AI: A psychological perspective. PsyArXiv. https://doi.org/10.31234/osf.io/7gwx2_v1
Summary. This PsyArXiv preprint reports a reflexive thematic analysis of interviews with 12 senior software and platform engineers who actively used AI tools. Through an SDT-informed psychological lens, it describes a transition from implementation toward systems thinking, a redefinition of productivity around impact rather than code volume, and a tension between autonomy-enhancing leverage and concerns about competence, de-skilling, and raised performance expectations.
Key claims
- Eleven of the 12 interviewed senior engineers who crossed a basic trust threshold for AI described moving attention from line-by-line implementation toward systems-level architecture, planning, and strategic judgment; one specialist’s repeated experience of inaccurate domain-specific outputs limited use to trivial tasks.
- Participants described productivity less as code volume or speed than as the quality, durability, usefulness, and total cost of the product delivered; some also reported that AI-assisted speed raised their perceived baseline for an adequately productive day.
- Participants framed AI both as a force multiplier and as a non-judgmental mentor, collaborator, rehearsal partner, or “rubber duck,” creating a private space to ask basic questions, clarify requirements, test arguments, and prototype ideas before social interaction.
- The same cognitive offloading that supported autonomy and higher-level work also produced competence concerns: participants reported forgetting basic syntax, worrying about loss of foundational practice, and needing architectural mastery and metacognitive oversight to direct and verify AI-generated work.
Evidence & limitations
- This is a February 2026 PsyArXiv preprint, not peer-reviewed. Its public OSF download exactly matches the ingested Drive binary by SHA-256.
- The study uses 45-minute semi-structured interviews and reflexive thematic analysis with a purposive sample of 12 active AI users who were senior engineers with at least five years of experience; it supplies rich mechanism and identity accounts, not prevalence estimates, causal effects, or evidence about junior developers or non-software work.
- Recruitment through LinkedIn and the DORA community, a predominantly male sample (11 of 12), Western technology-hub context, and the lead author’s Google/DORA affiliation constrain transferability; the paper says the work was conducted independently of Google and that the second author provided external analytical distance.