Somebody in last week’s AI4LIB Open Hour asked a question that stopped the room for a second. Not “should we use ChatGPT in the library,” which used to be the question every single week for the better part of a year. This time it was, “who in my library is actually responsible if the AI gets something wrong?” Nobody on the call had a ready answer. That silence, more than any survey or dataset, told me the profession has quietly crossed over from one debate into another.
For most of the past two years, the AI conversation in libraries was about adoption. Should we try it. Which tool. Is it allowed. That question is more or less settled now. AI already sits inside cataloguing, reference, discovery, and research support, whether a library formally approved it or not. The question worth asking today is not whether AI belongs in the library. It is how a library governs something that is already inside its walls.
Start with the job market, because numbers rarely lie the way opinions do. Going through more than 800 library and information science job advertisements across India, the pattern is hard to miss. MLIS, UGC-NET, SET, a Ph.D. where the post demands it, these still open the door. But almost every listing now expects working familiarity with Koha or DSpace, comfort with MARC21 and Dublin Core, exposure to research databases and plagiarism tools, and increasingly, some demonstrated ability to use AI responsibly. I won’t be wrong if I say the degree gets you shortlisted, the skillset gets you hired. For a student writing to me on WhatsApp asking what to learn next, that is the honest answer, not the diplomatic one.
Governance is where things get interesting, and where most libraries are furthest behind. Walk into any library today and you will find staff already using ChatGPT, Gemini, Copilot, NotebookLM, on their own initiative, often on their own devices, usually without telling anyone. Very few institutions have written down what is and is not acceptable. The University of Arizona Library’s AI strategy is worth a close read here, precisely because it does not obsess over which tool to bless. It builds around accountability, who decides, staff training, privacy, transparency, and risk. That is the correct instinct. We went through the same lesson with the internet, then with institutional repositories, then with open access. The tool is never the hard part. The framework around the tool is.
This is where I keep returning to a phrase I have used before on this page: the Fluency Trap. AI can produce an answer that reads beautifully, cites confidently, and is completely wrong. Fluency and accuracy are not the same thing, and a profession built on evaluating information for a living cannot afford to forget that distinction just because the output sounds polished. If anything, AI-generated fluency raises the stakes on the oldest skill in librarianship: teaching people to ask, how do you know this is true.
Naturally the question that keeps coming back, in every workshop and every Open Hour, is whether AI will replace librarians. My answer has not changed and I don’t expect it to. AI has already absorbed the routine end of the job, quick factual lookups, first-pass metadata, basic retrieval. It has not touched, and I doubt it will soon touch, the part of librarianship that was never really about retrieval in the first place. Helping a confused research scholar turn a vague worry into a searchable question. Knowing which gap in the literature actually matters. Reading a community and knowing what it needs before it asks. That is judgement, and judgement does not come out of a training corpus.
There is a quieter opportunity hiding inside all this anxiety, and it belongs to public libraries as much as academic ones. Communities need someone to teach them how AI works, how to spot AI-generated misinformation, how to make an informed choice instead of an easy one. Public libraries are already stepping into that role in a few places, and it fits the mission better than almost anything else on offer right now. Academic and research libraries have their own version of the same opening, in metadata work, in supporting open science, in shaping how AI gets used across a campus rather than reacting to it department by department.
None of this sits in settled law, and it will not for some time. Copyright questions around AI training data, authorship, and licensing are moving through courts across the world as I write this, with well over 150 cases already filed. For anyone using AI in research or teaching, that uncertainty is not a footnote, it is a reason to be deliberate about disclosing what the AI did and what the human did.
If this week’s discussions distilled into one line, it would be the same three instructions I keep repeating on this Open Hour, and I make no apology for repeating them again: verify the source, disclose the tool, validate the output. Simple enough to fit on a sticky note. Hard enough that most institutions still haven’t managed it. Libraries did not ask to become the profession that teaches the world how to think about AI responsibly. But look around, nobody else is doing it half as well, and somebody has to.