Intention Drifting

The project investigates the phenomenon of intention drifting, as the gradual divergence between the developer's expressed intentions, the AI system's interpretations, and the actual behavior of the code in AI-mediated programming. The programming domain can be translated to a broader societal phenomenon: how decision-making power is subtly transferred from humans to AI systems through shifts that often go unnoticed. The project is funded by the Wallenberg foundations (WASP and WASP-HS) and runs over two years in a collaboration between Södertörn University and Lund University.

Intention Drifting: Ethics, Authorship and Critical Judgment in AI-Mediated Programming

AI-assisted programming is fundamentally changing how software is created. The developer's work is shifting from writing code to articulating intentions and interpreting what the AI system produces. Within this shift, a phenomenon emerges that the project names intention drifting: the gradual displacement between developers' expressed intentions, AI systems' interpretations, and the actual behavior of the resulting code.

Related drift phenomena have been observed across generative AI research, from autonomous agents that gradually deviate from their initially specified objectives to novice programmers who drift towards the suggestions offered to them. What these accounts share is a treatment of drift as a technical problem to be mitigated or a cognitive risk to be managed, resting on an implicit assumption of intention as a stable specification. The project departs from this assumption. Drawing on science and technology studies and post-phenomenological accounts of intention as situated and distributed, intention drifting is approached not as departure from a fixed origin but as the continuous reshaping of intentionality in human, AI and code entanglements. Fundamental questions of authorship, accountability and critical judgement follow, questions that cannot be answered by technical means alone.

The project pursues three integrated aims. The first is to develop intention drifting as a conceptual framework grounded in critical-theoretical perspectives on distributed agency, authorship and responsibility. The second is to investigate empirically how intention drifting arises, is recognised and is negotiated in programming practice, with attention to the ethical and critical moments these episodes contain. The third is to develop, through research-through-design, probes and provocations that make the phenomenon visible and support critical reflection and ethical judgement within the work of programming itself.

Methodologically, empirical inquiry, conceptual development and exploratory design work are combined in an iterative process. The empirical studies take the form of workshops with professional developers working with contemporary AI coding tools, followed by think-aloud sessions and retrospective interviews. The design work yields a small set of focused design probes that are tested in further workshops and that function as research artefacts in their own right.

Programming here serves as a window onto a considerably broader societal phenomenon. As AI takes on a larger role in decision-making across public administration, healthcare, education and infrastructure, questions of control, accountability and human agency become pressing, and are often framed in binary terms. Intention drifting complicates that framing by showing that control is rarely lost in single visible moments of delegation, but through small and often unnoticed shifts that accumulate over time. The project contributes a vocabulary that makes such shifts possible to identify and discuss.

The project runs for 24 months from December 2026 and is a collaboration between Södertörn University and Lund University. Martin Jonsson is main applicant and WASP-HS PI, while Emma Söderberg is WASP PI and hosts a postdoctoral researcher that will work at the Department of Computer Science in Lund. Funding is provided by the Wallenberg foundations through the common call between WASP and WASP-HS.