Research, Pillar 01
AI in Software Engineering
AI is changing software development. Whether it is changing it in the ways the conversation assumes is an empirical question.
Most commentary on AI in software engineering moves faster than the evidence. This research programme slows that down: measuring what is actually happening in engineering teams, tracking adoption over time, and applying the same analytical standards used across the rest of this work. The findings are more nuanced than the headlines suggest, and more useful for that reason. The work now runs at three scales: engineers and teams adopting AI tools, a national programme closing the adoption gap, and the study of ecosystems where autonomous agents are contributors.
Current projects
Human-Centred AI for Software Engineering. A Grand Solutions programme funded by Innovation Fund Denmark, studying how Danish software organisations adopt AI for software engineering at national scale, with Daniel Russo as co-principal investigator.
Adoption Dynamics
The most common assumption in AI tool rollouts is that adoption follows perceived usefulness: if engineers believe a tool helps, they will use it. The empirical record qualifies this significantly.
Workflow compatibility is the stronger predictor. Engineers adopt AI tools that integrate into how they already work. Tools that require changes to established practices face sustained resistance, regardless of how capable they appear on paper. An organisation introducing AI coding assistants should map integration points to existing workflows before measuring uptake, and should not interpret early resistance as evidence of permanent rejection.
A second finding: social proof within the team matters as much as the tool’s inherent qualities. Engineers who initially resist AI tools become high adopters when a trusted colleague demonstrates integration into a shared working context. Adoption is partly a social process, not only a capability evaluation.
Creativity and Generative AI
When routine coding tasks are automated, developers gain time. What they do with that time is not automatic.
Early findings from this work suggest that the quality of non-routine creative work after AI adoption depends heavily on team structure and explicit expectations, not on the AI tool’s capabilities alone. Teams that restructure around the time freed by AI show different outcomes from teams that simply add the tool to an unchanged workflow.
This complicates simple productivity narratives. The question is not whether AI makes engineers faster. It is whether AI makes engineering teams better, and under what conditions that happens.
Human-Centred Adoption at National Scale
The adoption question does not stop at the team boundary. Roughly half of Danish software organisations do not yet consider AI for software development, and closing that gap rigorously is a research problem in its own right.
AI4SE1DK, a Grand Solutions programme funded by Innovation Fund Denmark, addresses it empirically. A national survey and longitudinal action research map how AI-assisted practices spread across fourteen partner organisations, and adoption is linked explicitly to workforce wellbeing rather than raw automation, measured with validated instruments through 2029.
AI-Native Software Ecosystems
As coding agents move from assistants to contributors, the unit of analysis shifts from the individual tool to the ecosystem it acts in. Agents that each pass their own tests still leave repositories that accumulate problems no single contribution accounts for.
The AI-Native Software Ecosystems line studies these ecosystems as complex adaptive systems. First evidence from more than 930,000 agent-authored pull requests indicates that integration risk is a property of the repository rather than the individual agent, which moves the governance question from certifying agents to monitoring ecosystems.
Defining publications
- Navigating the Complexity of Generative AI Adoption in Software Engineering — ACM TOSEM 33(5), 2024
- Creativity, Generative AI, and Software Development: A Research Agenda — ACM TOSEM 34(5), 2025
- Exploring Individual Factors in the Adoption of LLMs for Specific Software Engineering Purposes — Information and Software Technology 200, 2026
- More Is Different: Toward a Theory of Emergence in AI-Native Software Ecosystems — preprint, 2026
All 62 papers, filterable by year and type
Frameworks
Human-AI Collaboration and Adaptation Framework (HACAF)
HACAF describes how software engineers absorb AI assistance into workflows they already have. Its central claim, which the adoption studies on this page test and support, is that workflow compatibility predicts adoption more strongly than perceived usefulness: engineers take up tools that fit how they already work, and quietly abandon those that ask them to work differently. The framework is the reference point for the adoption research in this pillar.
Russo, D. (2024). Navigating the Complexity of Generative AI Adoption in Software Engineering. ACM Transactions on Software Engineering and Methodology, 33(5). 10.1145/3652154
The Copenhagen Manifesto
The Manifesto sets out what it means for generative AI in software engineering to be human-centred. It argues for reversing the common order of adoption: begin with what engineers and teams actually need, then decide whether and how AI helps, and judge the result on well-being, learning, and autonomy rather than on speed alone. It emerged from the Copenhagen Symposium and provides the ethical frame for the adoption work in this pillar.
Russo, D., et al. (2024). Generative AI in Software Engineering Must Be Human-Centered: The Copenhagen Manifesto. Journal of Systems and Software, 112115. Read the Manifesto