In her opening keynote for the 2026 Internet Librarian Connect virtual conference, Shani Evenstein Sigalov, Marie Curie Postdoctoral Fellow; lead researcher at the University of London with AI-BRIDGES; and convener of MOSAIC, the Knowledge Commons AI Consortium, discussed how to bridge the gap between AI system and the Knowledge Commons.
She began by inviting attendees to meet their new colleague, the one who has "read every book in the library". It can answer billions of questions but sometimes, too often, provides the wrong answer. She identified a gap between what she calls a "black box" and a "glass box". The black box is characterized by fluent answers, no sources, no context and no accountability. The glass box has every claim sourced, every change visible, context preserved and communities accountable. "AI has read everything and cataloged nothing," she commented.
There is good news about the gap. You can ask questions in plain language; collections are described and structured at scale; open access spans languages; records are matched to shared identifiers; and AI can surface what's missing. In the knowledge loop, people create, the commons curates, AI learns and people use AI, which restarts the loop. Beware, however, of the possible breakdowns in the loop, particularly around provenance opacity, disintermediation and recursive contamination.
Not Like Human Library Users
From her work with Wikimedia, Sigalov has realized that bots, unlike human library users, are accessing the resources in a way that strips the shelves rather than reading books. While LLMs are trained on massive datasets, they often lack the structured, verifiable, and multilingual richness found in Linked Open Data platforms like Wikidata. She's discovered that LLMs are learning from Wikipedia but not from Wikidata. Somewhat frighteningly, AI agents are editing and creating Wikipedia pages. The licenses under which Wikipedia makes its information freely available were made with humans in mind, not AI. And AI apparently lacks good manners and has no understanding of reciprocity.
Librarians are left confronting what Sigalov termed the "holy trinity"—misinformation, disinformation and missing information—which puts entire institutions as well as society in general at risk. She wants an open and trustworthy that runs on our terms in lieu of depending on large technology companies.
Looking at "The Six Pack of Care", formulated by Audrey Tang and Caroline Green at the Institute for Ethics in AI, University of Oxford. The six elements are attentiveness, responsibility, competitiveness, responsiveness., solidarity and symbiosis.
Developing a Trustworthy Model
At AI-BRIDGES, with its mission to connect institutional data, linked open data and AI systems, her team is working on phasing in a model institutions can trust.
- Phase 1, teaching AI to check the catalog (the corpus is on Hugging Face; the harness, code and benchmark on GitHub; and the report on ArXiv)
- Phase 2: train small, open Wikidata search models on the corpus
- Phase 3: humanities and GLAM scholars evaluate them on real research questions
- Polyphonic AI; many voices, languages and collections, not one flattened story
Right now, librarians can build by linking library catalogs to Wikidata and opening up metadata so that AI learns from the glass box. Librarians can advocate by asking AI vendors where they got their sources. To educate users, bring AI into teaching information literacy and stress the importance of asking "Where did you read that?" as part of the education endeavor.
We should not fear our new colleague, she advised, but we should train them. Citing sources is important but even more critical is the fight for truth.