In May, researchers, heritage professionals, data specialists, and other stakeholders gathered in Lisbon for our workshop dedicated to a question that is becoming increasingly urgent across archaeology and cultural heritage: how can we develop applications and use archaeological comparative collections to support AI development while ensuring high-quality training datasets, transparency, and the continued value of human expertise?
The workshop was designed to bring together stakeholders from different sectors to discuss future challenges related to the use of archaeological comparative collections for AI applications and to strengthen competencies in building training datasets for AI systems. What emerged throughout the day was a complex conversation about people, skills and responsibility.
The day opened with an introduction to MAIA and updates from its Working Groups, followed by a series of short presentations from stakeholders representing different backgrounds and perspectives. Although participants approached the topic from different angles, a common theme quickly became clear: AI is only as good as the data that supports it.
Again and again, speakers returned to the importance of high-quality, structured, and well-annotated datasets. Several highlighted the risks posed by fragmented infrastructures, inconsistent documentation practices, and the lack of shared standards across institutions and countries. Others stressed the importance of open code, accountability, and sustainable digital infrastructures. Without common frameworks and clear responsibilities, even the most advanced AI tools risk producing unreliable results. One participant summed up a challenge that many recognised: archaeological data does not exist in isolation. Geological, geophysical, environmental, and spatial datasets can all contribute valuable archaeological insights. The real task is learning how to connect and structure these different sources of information in meaningful ways.
If data quality was one recurring theme, human expertise was another: the phrase “keeping humans in the loop” appeared repeatedly throughout the discussions. Various stakeholders expressed concern that the growing use of AI could lead to a gradual loss of specialist skills if human oversight is not maintained. Participants pointed to a growing skills gap affecting the sector from different directions. Senior professionals often have decades of practical knowledge but limited time to develop new digital competencies. Younger professionals may be more comfortable with emerging technologies but have had fewer opportunities to acquire field experience and specialist expertise. This raises a fundamental question for archaeology and many other disciplines: how do we ensure that technology supports expert judgement rather than replacing it?
Following, a roundtable discussion allowed participants to explore some of the most pressing issues facing archaeological AI today. The first was the widening skills gap. Developing intelligent agents and AI-supported workflows requires technical competencies that are still unevenly distributed across the heritage sector. Participants discussed the need for programming skills, shared standards, and stronger collaborations with software developers to ensure that archaeologists retain ownership and control over their data. Legal uncertainty also emerged as a major concern. Questions around licensing, data reuse, intellectual property, and ownership of AI-generated outputs remain difficult to navigate for many institutions and researchers.
What is especially fascinating is that we are once again discussing what is exactly archaeological data. This debate is hardly new. Archaeologists have been discussing the nature of archaeological evidence and documentation for decades. However, the rapid growth of digital infrastructures, large datasets, and AI technologies has brought the issue back into focus. In a world increasingly shaped by datafication, defining what counts as archaeological data is no longer a purely theoretical exercise. It directly influences how datasets are created, shared, curated, and used to train AI systems.
In the afternoon, participants heard from MAIA Working Group 2, which is developing guidelines to support the responsible use of archaeological data in AI applications. The group’s ultimate goal is to produce practical recommendations addressing key areas such as:
- Data quality
- Data use guidance
- Persistent and unique identifiers
- Long-term sustainability
This work is being developed through seven selected case studies that explore real-world challenges and opportunities. The subsequent discussion reinforced a message that had already emerged throughout the day: trustworthy AI depends on trustworthy data. Questions of data quality, standardisation, ownership, availability, curation, and trustworthiness are not technical details to solve later. They are the foundations upon which future AI applications will be built.
The workshop concluded with a series of dilemma games that invited participants to step into the role of archaeologists facing difficult decisions involving archaeological data and AI. These exercises moved the conversation from abstract principles to practical scenarios. Participants were asked to weigh competing priorities, balance risks and benefits, and reflect on the ethical implications of their choices.
–
The Lisbon stakeholder workshop demonstrated that the future of AI in archaeology is not simply a technological question: data quality, shared standards, legal frameworks, professional skills, and ethical responsibility need to be built together. As AI tools become increasingly integrated into archaeological research and heritage practice, the community faces a dual challenge: building the datasets and infrastructures that make these technologies possible while preserving the expertise, critical thinking, and contextual understanding that give archaeological knowledge its value.