GLAMorous AI TL;DR — August 2026

Who Gets to Be AI-Ready? The Equity Crisis Nobody Is Discussing

Author

Alphaeus Lien-Talks
August 16, 2026

Multimodal AI can work across diverse cultural contexts. But only for institutions that can afford to prepare their data, train their staff, and audit their systems.

Adapting from Abu Talib et al. (2026)

Welcome to August's edition. This month we're stepping back from the models and guidelines to ask a harder question: who gets resources to adopt AI responsibly, and who gets left behind?

The technical conversation is thriving. The equity conversation is nearly silent.

Abu Talib, Ibrahim & Abusirdaneh – Reusability and Benchmarking Potential of Architectural Cultural Heritage Datasets for Generative AI: An Analytical Study

A careful study examining which cultural heritage datasets are actually usable for AI training. Finding: architectural datasets from Western European institutions are well-documented, standardised, and ready for multimodal learning. Datasets from Middle Eastern, Asian, and African institutions? Fragmented, inconsistently formatted, and often not machine-readable.

The result: generative AI trained on reusable datasets will reflect which regions, cultures, and institutions have the resources to standardise their metadata.

This is not a problem with the model. It is a governance and funding problem.

Appear2Meaning Benchmark Contributors – Appear2Meaning: A Cross-Cultural Benchmark for Structured Cultural Metadata Inference from Images

An important benchmark testing whether multimodal AI can infer accurate cultural metadata across different contexts. The work is valuable precisely because it acknowledges: generative AI systems perform systematically worse on non-Western cultural objects, even when trained on diverse datasets.

Why? Because "diverse" training data is often sourced from well-resourced Western institutions that happen to hold global collections, not from local institutions that hold culturally contextualised knowledge.

This creates a two-tier system: global institutions set standards for how cultural objects are described; smaller, local institutions adapt to those standards rather than defining their own.

🏛️ What People Are Saying

Halaychik – The Shrinking Vendor Landscape: A Wake-Up Call for Libraries

Libraries have handed over control to vendors—companies providing books, databases, discovery systems, and now AI tools. Smaller institutions are particularly vulnerable. They lack the technical expertise to build systems independently, so they outsource to vendors. But outsourcing means losing internal expertise, becoming locked into proprietary contracts, and being unable to interrogate how systems actually work.

The result: the institutions least able to afford dependence become the most dependent.

The alternative Halaychik proposes—pooling resources to create independent networks—requires coordination and funding that smaller institutions struggle to afford. So the status quo persists: outsourcing as a necessity, not a choice.

Breeding – Library Systems Report 2026: Innovation Under Constraint: How Libraries and Vendors Navigate Austerity and AI Disruption

Marshall Breeding's annual systems report documents the real landscape: smaller libraries are adopting fully hosted, vendor-managed solutions (OPALS, MEDAD, Atriuum ILS) that require no local technical expertise. No servers to maintain. No staff to hire. But also no control, no transparency, and vulnerability to vendor decisions about pricing, feature updates, and data practices.

For smaller institutions, this is rational. For the sector, it creates an archipelago of isolated, vendor-dependent libraries. When vendors consolidate (as they do), options disappear. When vendors fail or pivot to AI products, smaller institutions have no recourse.

Sui, Yang & Sui – A Systematic Review of AI in Cultural Heritage Preservation: Technological Frameworks, Applications, and Future Directions

A comprehensive map of where AI is actually working in heritage: Restoration (damage diagnosis, digital reconstruction), Understanding (visual analysis, semantic interpretation), and Management (object monitoring, systemic oversight). The paper identifies six future directions: multimodal AI, explainable systems, safe generative restoration, cultural consistency methods, heritage large models, and digital twins.

The implication is clear: these capabilities are technically achievable. But they require sustained investment, cross-institutional collaboration, and institutional capacity that is unevenly distributed. A well-resourced institution can build toward digital twins and heritage large models. A small archive cannot.

🚨 What People Are NOT Saying

The Regional Disparities We Are Ignoring

There is a clear geographic gradient in AI readiness across GLAM institutions. Well-resourced Western institutions (US, UK, Germany, Scandinavia, etc.) are adopting AI with support, guidance, and peer networks. Institutions in Eastern Europe, the Global South, and smaller nations are watching from the margins.

This is not accidental. It reflects decades of funding disparities, digitisation inequities, and who gets to publish in English-language journals.

Now AI adoption is following the same patterns. The rich get richer. The under-resourced fall further behind.

The Cost of Non-Adoption We Are Not Measuring

What happens to an institution that chooses not to adopt AI? In the short term, nothing. In the medium term: competitors adopt AI and become more efficient. Funders expect AI integration in grant applications. Job candidates ask about AI capabilities. Researchers prefer datasets that are AI-indexed and searchable.

Over time, non-adoption becomes a penalty. Your collections become less discoverable, less useful, less funded.

But this cost is invisible. It is not published. It is felt as slow institutional decline.

The Liability Vacuum We Are Creating

As institutions rush to adopt AI—often without full preparation—who bears responsibility when systems fail?

An archive deploys AI transcription without sufficient training data in their language. The model makes systematic errors on proper nouns, place names, and Indigenous terminology. Researchers build work on corrupted metadata. Years later, the errors are discovered.

Who is liable? The institution? The vendor? The AI company? The internal team that implemented the system without sufficient oversight?

Every institution is navigating these questions alone, without published guidance, without peer precedent, without insurance.

The Vendor Lock-In We Are Normalising

Smaller institutions are adopting fully hosted, vendor-managed solutions. No internal servers. No technical staff needed. It is rational for institutions with limited resources. It is catastrophic for sector resilience.

When a vendor consolidates (which they do), options disappear. When a vendor decides to pivot from library management to AI services, smaller institutions have no alternative infrastructure to move to. Their data is held hostage in proprietary systems.

This is not a technology problem. It is a market structure problem. And no one in the GLAM sector is discussing how to break it.

The Institutional Colonialism Built Into AI Tools

Smaller institutions increasingly rely on external vendors—or external experts—to deploy AI systems. They cannot afford to hire data scientists or build oversight committees. So they outsource.

Outsourcing means losing internal expertise, losing the ability to interrogate how systems work, and becoming dependent on external providers' decisions about what counts as "responsible AI."

The most vulnerable institutions—those with the smallest budgets and fewest technical staff—end up most dependent on external vendors. This creates a new form of institutional colonialism: technical systems are designed and controlled elsewhere, imposed on local collections, with no local agency to change them.

❓ Big Question

If institutions that cannot afford responsible AI adoption are either falling behind or becoming dependent on external vendors—who is responsible for closing this gap?

Is it individual institutions, competing for limited funding? National governments, fragmented across jurisdictions? International bodies like UNESCO, which move slowly? Or is this a market failure that requires intervention we have not yet named?

And more fundamentally: should institutions be expected to adopt AI without resources to do so safely?

💬 About

I'm Alfie, a researcher and archaeologist exploring where heritage, ethics, and AI meet. 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 September. And if you work in a smaller or under-resourced institution, I want to hear what adoption actually looks like on the ground.