Responsible AI (RAI) in journalism is most often discussed in terms of the actions that systems should take: transparency, accountability, and fairness. Codes of ethics, regulatory frameworks and design principles all share this normative orientation. While these are important, they have a structural blind spot. They assume that responsibility can be defined in advance and stabilised through technical or governance measures. However, this overlooks the more interesting question of how responsibility is actually produced in practice when journalists interact with AI systems under deadline pressure, using imperfect tools within organisations that have their own cultures and sometimes contradictions.
This is the topic I have been working on for the past ten years, at the intersection of journalism studies, science and technology studies, philosophy and information systems research. The first learning that feed the responsible AI framework on which I am currently working at a more theoretical level are that there are different levels of use of technology and that rather than moving too fast towards its strongest intensity, in terms of adoption or appropriation, framing the uses of AI as a smooth integration appears to me what’s the most accurate to describe what’s currently going on in journalism.
One personal hypothesis is that Rogers’ diffusion of innovation theory is still widely used as a reference, even though it treats adoption as an event rather than a process (which I believe is misleading). Rogers’ model positions the adopter as a rational decision-maker choosing to accept or reject an innovation. Nevertheless, this reasoning struggles with situations where « adoption » is pressured, partial, reluctant, or contradictory. This theory also implicitly posits that once adopted, the technology disappears analytically. There is no mechanism for studying what happens after the threshold, how practices transform, how trust is recalibrated, and how breakdowns occur. This is why Jouët’s appropriation was already a critique of this tradition in the 1990s; she was pushing back against the diffusionist model by insisting on the creative, transformative work users do with technologies. De Certeau’s bricolage (DIY) does similar work.
The second learning, derived from my PhD thesis as well as from my research within the NORDIS hub on responsible fact-checking technology, is that the formation of use is a very complex process that stems from both external and internal factors: cultural and professional representation interplay with organisational, ethical, and practical variables that make the analysis complex to conduct (see my recent paper with Arjen van Dalen, presenting an adaptation of the Technology Acceptance Model for journalism, the TAM-J framework).
The third learning is that it is particularly complicated to embed legal and ethical constraints into technological systems without considering the broader conditions that shape responsible use. In this process, works from Science and Technology Studies, including the French sociological school of use, helped me make sense of what I observed. Layers from ethics and epistemology, as well as from human-machine interaction, provided fruitful angles to get a broader picture of what I was analysing.
These observations form the basis of the responsible AI framework for journalism. Drawing on Giddens’ theory of structuration, it consists of a four-dimensional analytical framework for examining how responsibility is enacted in AI-mediated journalism. These dimensions focus on journalistic workflows and routines, user experience and trust formation, ethics and epistemology, and information integrity and accountability.
The core argument is that responsibility in AI-mediated journalism is not a property of systems, nor a product of compliance. It is a recursive sociotechnical accomplishment that is continuously produced, stabilised, and destabilised through everyday practice.
One component of the quadrant is particularly critical in the age of generative AI: the question of information integrity. Unlike previous AI tools, generative AI does not retrieve or process existing information. It produces outputs that simulate epistemic authority while remaining disconnected from verifiable sources. The traditional markers of information integrity (traceability, source transparency, verifiability) are structurally undermined.
