Somewhere between the Delhi High Court’s order last week and a poll I ran in the AI4LIB group, I found myself thinking about a librarian I knew years ago who could locate any paper in the archive without once touching the catalogue. She simply knew where things lived. Nobody called that artificial intelligence. Nobody called it anything at all. It was just her job.
I bring her up because this was the week our profession’s fingerprints on the AI industry became impossible to ignore, and also the week it became clear how little the industry has given us back for it.
Start with the court order, since it set the week’s tone. The Delhi High Court, hearing ANI Media’s copyright suit against OpenAI, declined to grant an interim injunction. At this early stage, the Court observed that training an LLM on publicly available literary works could fall within fair dealing under Section 52(1)(a) of the Copyright Act, and that ANI had not shown a prima facie case that ChatGPT’s outputs reproduced its content closely enough to justify blocking the service. Nothing here is final. The main suit continues, and a fuller judgment may say something quite different once the evidence is examined properly. But for a case that lawyers, publishers and technology companies have been watching closely, even a provisional word carries weight, and it will shape how Indian courts think about AI training and fair dealing for a while yet.
Copyright was the backdrop. The AI4LIB group spent the rest of the week asking a more uncomfortable question: who actually built the ideas AI companies are now selling back to us.
I ran a small poll last week, one question. Do you have access to a paid LLM, and if so, who pays for it. Twenty four people answered. Fourteen had no access at all. Seven were paying out of their own pockets, something like four hundred rupees a month, quietly, without reimbursement. Three had both personal and institutional access. And exactly zero, not a low number, zero, worked at a library that had budgeted a single paid AI seat for its staff. I read that last figure three times before it settled in. We have spent two years writing papers and running webinars on how AI will transform library services, drafting policy on ethics and patron opt outs, while the people meant to carry this transformation are funding it from their own salaries. You cannot teach AI literacy to a patron using a tool you were never given yourself.
The fix, I’d argue, is not more vendor spend. It is sequencing. Fund staff first, even one paid reasoning model per person matters more than a fancy discovery layer nobody asked for. Then maximise what’s already free, Semantic Scholar for summarised search, NotebookLM for turning a stack of PDFs into something a new user can actually ask questions of, ResearchRabbit for visual citation mapping that still surprises people the first time they see it. Audit what your library already owns before buying anything new, EBSCO’s natural language search and Libby’s newer features go unused more often than anyone likes to admit. Only after that does it make sense to reach for the paid specialist tools, Elicit for systematic reviews, Scite for citation-backed fact checking. And before renewing anything next cycle, ask one honest question. Did a patron request this tool by name in the last six months. If not, you already know what to do with that line item.
Then came the piece that gave the week its real backbone, the one I have been calling, only half in jest, the room is already ours. Ask anyone in AI what makes a language model intelligent and you’ll hear about transformers, attention, training at scale. All true, none of it the full story. Underneath every system that retrieves and ranks knowledge sits an index, and indexing theory was not built by computer scientists. It was built by Ranganathan, by Mooers, by Lancaster, by generations of information scientists who spent careers working out how knowledge should be organised so it could be found again. Faceted classification, controlled vocabularies, relevance ranking, the inverted index itself, these came out of library science long before anyone thought to rebrand them.Which is really just the setup for the list that followed a day later, seven capabilities the AI industry is celebrating as breakthroughs that libraries were doing decades earlier under other names. Prompt engineering is the reference interview with a bigger salary attached. Retrieval augmented generation is a controlled database search, MEDLARS was doing exactly this in 1964. Knowledge graphs are subject indexing and faceted classification. Named entity recognition is authority control, which the Library of Congress has practised since 1902. Semantic search is what MeSH was built for. Content moderation is a collection development policy that every library has had for over a century. And hallucination detection, that one made me smile, we called it reference verification, and we’ve been doing it since the reference desk was a physical piece of furniture. AI scaled these ideas and automated them, genuinely and usefully. But the foundations were laid by information professionals who are now, a little absurdly, being invited to AI literacy workshops to learn concepts their own profession invented.
I could have let that sting and stopped there, but the week offered a better ending than resentment. A librarian at Berkeley’s East Asian Library is cataloguing nearly two thousand Chinese film posters, and AI drafts a translated description in seconds. Useful, fast, and still entirely subject to her judgement, because as the technical services lead there put it, the AI does not understand the collection, the users, or the cataloguing system the way she does. The same team has used AI to surface agricultural publications from the 1960s that had been effectively lost, and to flag sensitive material like Social Security numbers buried in large digital collections before anything goes public. None of it replaces a person. All of it frees a person from days of tedious searching so they can spend their time on the decisions that actually need a human in the room.
That, I think, is the shape of the answer to everything else the week raised, the copyright question, the provenance debate a paper on scholarly AI disclosure stirred up mid week, the budget gap, even the slightly bruised pride of watching our own ideas get patented back to us. AI does the searching. We keep doing the deciding. We built the index behind the intelligence once, quietly, over a century, without asking anyone’s permission. There is no reason we cannot do it again, this time out loud.