There is a danger, as there was in the 1990s when digital technologies were adopted by the commercial world, that humanities researchers will assume corporate capture of artificial intelligence (AI) is inevitable because of the technical nature of the technology. If this occurs, DH risks having our traditional focus on open data, open code, and open methods undermined. Because of this there is an urgent need for DH researchers to engage with AI at a technical level, experimenting with methods that support technical, cultural, and epistemological transparency. This needs to be augmented with methods that align such work to existing standards such as FAIR (including FAIR software) and CARE, emerging standards in Indigenous data sovereignty, and principles of responsible AI. This paper describes an LLM Retrieval Augmented Generation (RAG) tool developed by the AI as Infrastructure (AIINFRA) project. ATLAS (Analysis and Testing of Language Models for Archival Systems) enables reproducible and transparent experiments with multiple LLMs, vector stores (databases), and word embeddings, and is built using open source code. This has allowed us to design the tool with scholarly and Indigenous values at its core, providing detailed technical information about the calibration of each experiment (model version, word embedding, vector store characteristics, system prompt), as well as the source documents used by the foundation model to generate its response. Although limited by lack of scale and funding, ATLAS demonstrates how scholars can test the ability of AI to conceptualise the fragmentation, clustering, dispersion, and interconnection of data in DH.

The opportunities and challenges presented by generative AI for humanities research are staggering. The disruption wrought upon the IT industry from 1995 was extraordinary, but we had 3 years to adapt. Generative AI snuck up (machine learning folks were always promising, but always 10 years away from delivering anything useful) and ambushed us in the last year or so. Instead of years, we get months to adapt. The significance of generative AI for humanities research methods seems poorly understood and greatly underestimated. The most common responses are anecdotes about where it makes mistakes, hallucinates and how students are using it to cheat. Not untrue, but verging on trivialisation.

The capabilities of generative AI are quite astonishing. When generative AI can expertly summarise academic papers, can generate near-perfect Sanskrit and then translate and analyse grammar instantly, conventional research practice has changed radically. When generative AI acts as an instant master’s level assistant able to research, collate, analyse and present instantly, then the practice of a digital humanist has changed radically. The methodological challenge is what questions to ask, how to ask them, and how to validate the responses.

So, where do we stand? As researchers, we are ground truth. We need to engage in workflows that generate, validate and rectify the results produced by generative AI. The most productive way to engage with generative AI is as active collaborators rather than passive consumers—building workflows that reject or emend outputs, embedding corrections back into model training, and aligning results with scholarly standards. Through collaboration, scholars can own domain-specific models—grounded in peer-reviewed scholarship and tailored to the demands of research.

This presentation will explore how generative AI is being integrated into three research workbenches: Glycerine, TLCMap, and Omeka S.

• Glycerine Workbench has integrated an open-source Image AI and IIIF annotation pipeline, supporting iterative, scalable workflows for training models in image segmentation, captioning, and semantic tagging. This architecture is being developed collaboratively with the IIIF community, ensuring that training and deployment pipelines remain open, extensible, and fit for scholarly use.

• TLCMap has implemented an open-source mapping pipeline to extract and geolocate place names from large texts, with a focus on Australian contexts. Researchers can review, emend, and validate results within the workbench. The resulting data layers can be visualised on maps and analysed spatially—supporting use cases from literary geography to discursive museum catalogues.

• Omeka S, now being established as national infrastructure, is distinguished by its capacity for relationship graphing—incremental, collaborative, and semantically rich linking of heterogeneous content. Unlike paradigms based on repeatable experiments, these workflows thrive on cumulative annotation and knowledge construction. Through an ARDC CDL initiative based at the University of Sydney, we are designing Omeka S modules that embed writing, translation, annotation, and visualisation tasks—powered by generative AI—directly into researcher workflows.