Day two of the SEASON 2025 conference in Hamburg starts with a keynote by Anne Beaulieu, reflecting on the purpose and future of search. Search is now transforming considerably, especially given the intrusion of generative AI into the contiguous mix of uses for search, the impact of disinformation, the receding of the public sphere, and the impossibility of opting out from these systems; there is an urgent need to respond to this by rethinking search.
The way that students now engage with search engines and generative AI tools, for instance, reveals some of these changes and points to many of the concerns; if we conceptualise search interfaces as an organised space of interaction between a system and an agent, for instance, then this needs to be understood as layered and complex, and as mediating information, communication, and experience. It provides a platform for interpretive work in knowledge production.
In Science and Technology Studies, search is mostly understood as retrieval, emphasising interfaces dominated by a search box that guides the interface and interaction design for exploring abundant information. This is a very thin interface, though, and is applied to all sorts of engagement with information, well beyond generic search engines. It has replaced previous understandings of information as navigable via curated hyperlinks, for instance.
Such impoverished interfaces via a simple query box provide one dominant mode of interaction that is treated as a solution to a single problem, and offers only one kind of possible interaction. One question we might ask here is who this interface thinks users are, what capabilities they have, and how they might want to use it. With search, this is likely that the user knows what they are looking for, have the right vocabulary to find this information, wants the best match for their query, and does not want to engage with the complexity of the search space.
Can we move beyond such impoverished interfaces, then? What alternative strategies might there be: for instance, chaining (moving from one reference to the next); browsing (enabling serendipitous discovery); and other modes that foreground the richness of the data collection, and present metadata at various scales. These could, for instance, render the scope of the search space more readily visible (e.g. by showing the volume of content items available per year or per other criteria, or offering a thumbnail overview of possible matching items). Alternatively, they might present the search strategies of different users, enabling them to learn from and interact with each other; or they could foreground the narratives (also in multimodal form) attached to data sources and data curators rather than merely the data points themselves.
Such interventions can help to avoid the reproduction of an architecture of universalism, counter the fiction of a frictionless circulation of knowledge, increase reflexivity, challenge the idea of a ‘best match’, and reduce the isolation of the user as a lone evaluator. The question then becomes, of course, which novel interfaces can best foreground such new forms of engagement – and whether these simply represent incremental improvements to conventional search, or might even depart more substantially from established search paradigms. If our needs have changed from finding relevant documents to solving specific problems or supporting complex decision-making, this ought to be reflected in a paradigm shift in our tools for engaging with information, too.
Generative AI might provide one such paradigm shift, yet this depends on its deployment; LLMs are presently being used to augment search engines, and search engines to augment LLMs, and there are problematic practices like post-rationalisation search (where search is used after the fact to find sources that back up the AI-generated answer). Perhaps we need to think differently about the possibilities here.
This requires us to approach search as a knowledge infrastructure. If we are moving from information retrieval to solving specific problems or supporting complex decision-making, then this requires increased infrastructural meaning-making. Knowledge infrastructures are robust networks of people, artefacts, and institutions that have developed into complex phenomena over the long evolution of disciplines and institutions; the knowledge systems they support represent certain approaches to systematicity (knowledge organisation), reflexivity (reflection on what is known), and distributivity (knowledge availability in society).
Viewed through this lens, then, search that draws on generative AI meets different needs: shifting from matching to reasoning, solving specific problems, drawing on context awareness and a proactive recognition of information needs, and utilises the predictive logic of LLMs and their focus on language; it engages in summary and synthesis rather than mere retrieval: positioning search as a conversation, enabling different and multimodal interfacing strategies, constructing an answer through conversation rather than merely providing sources, but also generating considerably greater environmental costs and potentially relying on shallow sources in intransparent ways; generating potentially bad content: this may be of obscure provenance, unreliable due to mere semantic relevance rather than meaningful coherence, seeking to please the user rather than providing clear and potentially critical information, building on proprietary materials used without permission and under unjust labour practices, and including hallucinations caused by insufficient training data; and result in a danger of cannibalisation: where AIs train on AI-generated content and thereby enter into a feedback loop, and where content production incentives are removed because such content is no longer producing sufficient revenue.
Users of generative AI systems are not necessarily aware of or reflecting on these issues: there is an ethos of instrumentality and sufficing, and a belief in the law of large numbers which they hope means that the AI will be generally correct most of the time. Systematicity (through flattening, a short memory, and the dominance of certain inputs), reflexivity (decreased transparency, potential feedback loops), distributivity (through the dominance of big AI companies and the ecological implications of energy-hungry AI processes) are all negatively affected by generative AI technologies.
But equally, what can we learn from comparing search as retrieval to search as generative AI engagement? Which novel interfaces might we imagine beyond these options? How might search as knowledge infrastructure contribute to better knowledge?
