If they have, you might not have thought so from the ISKO-UK annual conference (7 October 2025). The theme for this online conference was “knowledge organization in the age of AI”. For those unfamiliar with the terrain, “KOS” refers to “knowledge organization system”, a catch-all term that comprises controlled vocabularies, taxonomies, thesauri, and the like. For over a hundred years (back at least to the Dewey Decimal Classification, Copilot tells me, and most likely further) these methods have been used to aid discovery, particularly in the context of libraries. Today, knowledge organization is an established academic discipline, with several scholarly journals, with such titles as, unsurprisingly, Knowledge Organization, Advances in Knowledge Organization, and Cataloging and Classification Quarterly.
Knowledge Organization 101
Back to basics: KO is about discovery. Library classifications enabled us to find books, at a time when you only had the choice of looking through the catalogue, or browsing through the books on the shelf, which (you hoped) were ordered by subject. Today, with the advent of full-text indexing, and Google, discovery typically avoids the library catalogue, and we typically search by word or question. Little attention is paid to the accuracy of searching except for some defined business areas. Commercial search, such as looking for things on Amazon, is a highly sophisticated industry, although it’s not so much about finding the right thing as placing things in front of the user that the supplier wants to sell. The euphemism “search engine optimisation” is the commercial application of tools to promote individual results, with those results are optimal for the vendor. That’s how a search for a film on Netflix returns unconnected results, because they don’t have the film you are looking for.
Enterprise search continues as a specialised activity, managed by a few experts. Academic search is largely left to users to sort out for themselves, although systematic reviews for evidence-based medicine are widely used, as you would imagine. The use of taxonomies and classification appears to be something of a sideline for much of content discovery these days.
This conference
So, what’s new in knowledge organization? From the presentations I heard, knowledge organization seems to be in defensive mode. For the most part, speakers emphasised the limitations of AI, and the strengths of human information scientists, in a tone that reminded me somewhat of the early days of internet search engines, such as Yahoo, Lycos, and then Google. The response from the library world at the time, around the year 2000, was horror that the human at the reference or information desk could be replaced – but few libraries have such a human-based resource these days.
The presentation that for me engaged with, and looked for, evidence from using genAI, was by Tony Russell-Rose. He is an academic who combines this with running his own search company, 2d search, which uses an innovative visual interface to carry out discovery. His examples were focused on those areas of search that still use highly managed search strategies, such as recruitment (matching candidates to jobs). The interest of the talk was how the 2dsearch tool has started using genAI tools to provide search results. He provided a simple history of discovery mechanisms, starting with “term selection”, which he described as “a common source of inefficiency and error” (although it is still used for systematic reviews), for example:
(“business analyst” or “systems analyst” or “system analyst” or “data analyst” or “requirements analyst” or “functional analyst”) and crystal and report* and analy*and data near analy* and not inventory and not retail and not (ecommerce or “e-commerce” or b2b or b2c)
This was followed by what he described as first-generation language models such as word2vec, which produced better results on standard tests, but these were then superseded by the use of LLMs, specifically, OpenAI. On looking at the latest 2dsearch interface, the results still appear to be very much around the term selection method, but the significance appears to be that the search engine has taken your query and done all the term selection for you – with better results than any manually-curated resource, such as linked open data. You still have the option of tweaking the automatic selection of terms, if you so desire.
Tony Russell-Rose was followed by Julain Clunis, who reverted to a traditional defence of KOS practice to improve (or replace) AI tools, but it all seemed somewhat less relevant after the 2dsearch presentation.
Does KO influence how you present information?
Stressing the limitations of genAI was not the only theme common to most of the presentations. Sadly, many presentations were barely disguised academic papers. In fact, this conference established (for me) new records in the number of slides for one presentation (30 slides in 30 minutes), and a new record for the number of words per slide (an average of around 125 words per slide, in a presentation of 22 slides). The “concluding remarks” for one presentation comprised three full slides:
Slides such as the one below, presented at a rate of one slide every 60 seconds, made it well-nigh impossible to understand what is being stated:
I found this graphic incomprehensible. “FT” perhaps stands for fine-tuning, but what does “STD” and “bCoT” stand for? How do these results show that “training on a multidisciplinary dataset produces higher results”? Your only chance of understanding is to turn to the original article from which the presentation is derived.
I would have thought that the concept “knowledge organization” might include a component of how to present knowledge in an intelligible way – sadly not the case here.
Complaints about AI
Many of the complaints about AI were of the familiar kind. There were references to “algorithmic bias”, and claims that algorithms are “far from neutral” – not surprising, since they are based on corpora of human-created content, with all the prejudice and bias inherent with humans. I would argue that there is no such thing as algorithmic bias – the problems come not from the algorithm, but the data on which the algorithm is applied, for example, using Victoria’s Secret models as the standard type for the female body. Or using an algorithm that identifies as “sexual” images with more pixels of flesh – with the consequence that images of larger women are more frequently labelled as sexual.
Conclusion
Do we want more complaints about genAI and its drawbacks? Following Tony Russell-Rose’s presentation, the first question from the audience was “so what should human knowledge organization experts be working on?”, to which question Tony responded, diplomatically, that he would think about it. It seems to me the best thing for KO experts to be doing would be to carry out a formal evaluation of interfaces such as 2dsearch, to determine how best to make use of genAI as part of a search strategy.
Leave a Reply