Research Software in an Age of AI-Assisted Development: Reflections from Edinburgh

Important: This document began as a draft vision statement prepared in advance of the “Research Software Engineering in the Age of Generative AI” workshop, March 2026. It was intended to lead to discussion in the document before and discussion in-person at the workshop, including disagreement, and refinement. The current version of the document is a snapshot of that draft vision statement and some of the discussion. It captures a sense of the identified principles, benefits, challenges, and open questions, rather than having a single point of view that represents all of the workshop attendees and contributors. Not all authors necessarily agree with all the points in the document.

A note on terminology: We use AI-assisted software development throughout this document to refer to tools and practices often described as Generative AI, GenAI, LLM-enabled development, or AI-assisted coding. We chose this term because it focuses on the activity we care about here, namely the production, verification, documentation, and maintenance of research software, rather than on a particular technology label or marketing category.

Research software, the source code, scripts, computational workflows, and infrastructure that power modern research, have never been more central to human discovery. And the scholarly community (funders, publishers, societies, etc.) increasingly recognizes this, though there is still a long way to go before such recognition is widespread or leads to systemic change.

At the same time, in the last few years, the rise of AI-assisted tools, i.e., those that can generate code and documentation by using large language and multimodal trained models, and act in semi-autonomous “agentic modes,” has caused significant disruption in software creation and society at large. These tools raise legitimate questions about training data provenance, licensing, quality, security, attribution, environmental cost, and accountability that are hard to ignore. Yet the conversation around AI-assisted tools too often seems to oscillate between hype and dystopian predictions. Neither of these extremes serve research well. This document also starts from the assumption that these tools are and will continue to be used, regardless of the concerns raised above, and aims to guide further usage within this context.

This document outlines some ideas about how research software can and should be produced in the era of AI-assisted research tools by Research Software Engineers, computational researchers, and domain experts writing their own code. It focuses on the software production function: building and maintaining software that meaningfully advances research, but also includes the people involved. Workshop participants discussed both, including the role of RSEs and how it will change, such as leveraging the ADSA/USE-RSE-led Position Statement on Generative AI in the RSE Workplace and related work in this space.

We believe that developing a shared vision of the future of research software is urgent: AI tools act as an amplifier of existing practices at every level: individual coding habits, institutional incentives, and the commercial priorities of the companies building these tools – accelerating whatever patterns they encounter, both good and bad. Applied without thought, AI-assisted tools can proliferate bugs, embed and reinforce biases, and generate plausible-looking results faster than our existing systems of peer review and verification can check and absorb them. Applied with care and discipline, it has the potential to expand access to computational methods from those with software development skills to those without them, at least for those who can afford it, and allow all researchers to focus on the work that most requires human skill and judgment.