GLAMorous AI TL;DR — April 2026

Welcome back to GLAMorous AI, a short roundup of how artificial intelligence is being used in museums, archives, libraries, and archaeology.

This month, we were in Vienna. Here is what stood out.

🏛 CAA International 2026 — A Quick Summary

The 53rd CAA International Conference brought together archaeologists, heritage professionals, and computational researchers in Vienna from 31 March to 3 April 2026. Across six sessions and roughly 90 papers, the programme covered an exceptionally wide range of themes: from Linked Open Data, CIDOC CRM, and FAIR data practices through to AI, remote sensing, sensory archaeology, ethics, infrastructure, and community-led approaches.

What stood out was how each session felt distinct in focus yet still complemented the others. Whether moving between highly technical discussions, theoretical reflections, or applied case studies, there was a strong sense of coherence across the conference as a whole. One of the clearest wins was minimising session overlap, which meant it was possible to attend most of the AI-focused programme without clashing.

It was also notable how many sessions extended beyond the immediate scope of AI; a sign that these topics are growing in relevance across the field as a whole, not just within specialist communities.

🤖 AI Across the Pipeline

Several sessions connected directly with the MAIA COST Action (Managing Artificial Intelligence in Archaeology), covering AI applications in cultural heritage protection, digital collections as training data, generative AI and text mining, vision foundation models for remote sensing, and questions about AI in the post-digital era of archaeology.

A major new systematic review of machine learning in archaeology examined 135 articles published between 1997 and 2022, finding a significant rise in publications from 2019 onwards. Automatic structure detection and artefact classification were the most represented tasks, with a consistent pattern of mixed or unsuccessful results in those same high-volume areas.

More models. More papers. More patchy results. The methodology problem has not gone away.

Read the review →

📄 LLMs and the Legacy Data Problem

One of the most practically relevant threads at Vienna was applying large language models to archaeological documentation — excavation reports, grey literature, archival records.

Recent benchmarking found that multimodal LLMs can transcribe historical manuscripts with character error rates of around 5–7% on 18th and 19th century English documents, a notable improvement over Transkribus, and reduceable further to around 1.8% with LLM-based post-correction (see Levchenko, 2025). Impressive on paper. But archaeological documents are not well represented in LLM training data, and domain-specific fine-tuning remains under-resourced across most of the sector.

Studies comparing LLMs against Transkribus on multilingual historical datasets found no consistent overall winner. Context dependency is high. Ground truth datasets for archaeological documents are scarce. Without them, these workflows cannot be safely evaluated — or deployed.

Read more → 

⚖️ Training Data, Ethics, and the FAIR Question

The session on digital archaeological collections as AI training data may have been the most important at the whole conference. It asked a question the sector has been avoiding: what data are we actually training models on, and do we have the right to use it?

MAIA has identified one of AI's most challenging aspects as not developing appropriate algorithms but creating the datasets essential to train them — and noted that archaeology presents an extreme example of this problem. Despite the field being highly digitalised, archaeological information is often stored in digital formats but rarely structured in a way that makes it machine-actionable.

Community data, indigenous heritage records, and contextually sensitive finds all sit in collections digitised without AI use in mind. Using them as training corpora without consent frameworks is a governance failure, not just an oversight.

Read more →

🔭 MAIA in Hainburg — Just Before Vienna

The MAIA COST Action held its Second General Meeting in Hainburg on 30–31 March, co-located with the conference. Working Group 1 shared results from a literature review of 1,297 bibliographic records and an initial global AI mapping survey in archaeology. Working Group 2 focused on guidelines for training datasets across use cases including numismatics and image recognition.
The afternoon workshop featured talks covering multimodal language models for experimental archaeology, explainable AI in archaeological interpretation, AI-assisted 3D digitisation, and the use of AI as a tool for expanding the hermeneutic circle.

The explainable AI thread is one to watch. If archaeologists cannot interrogate why a model reached a conclusion, the output cannot enter the scholarly record.

📖 Watch This Space

CAA proceedings are published on a rolling basis through PCI Archaeology — peer-reviewed papers from Vienna will begin appearing over the coming months. Worth bookmarking: proceedings.caa-international.org

Also recently published in JCAA: a paper on quantifying inherited uncertainty in archaeological legacy data using fuzzy logic — directly relevant to questions about data quality and AI training. Read more → 

❓ Big Question

If AI can classify artefacts, detect sites, and extract knowledge from legacy reports — but the data it trains on is undocumented, unconsented, and not FAIR — who is responsible when the outputs are wrong?

💬 About

I'm Alfie, a heritage scientist and archaeologist exploring where heritage, ethics, and AI meet. I also chair the CAA Special Interest Group on Machine Learning and AI, and was in Vienna presenting on OCR and NER pipelines for archaeological and archival data.

This digest keeps things short, critical, and useful — no jargon, no hype.

👉 Read or subscribe at glamorousai.beehiiv.com 👉 Send papers, reports, or ideas for May