2026-06-03
This presentation highlights the added value of creating environments that foster serendipity as opposed to strict findability in the context of documentary search driven by “AI-research-assistant”.
AI: “automation of cognition” Abbass (2021)
Documentary Research: human led process of research and selection of documents.
Information Retrieval: technical matching between a query and one or more documents.
Google Scholar basics
Google Scholar Labs provides an explanation for the article’s presence in the results list
In order to understand the implications of the integration of “AI-based systems” within information retrieval infrastructure, let’s look at the possible places of its intervention.
Typology based on Tay (2025).
Deep research with Undermind.ai
Evaluation of RAG and AI Research Tools : Archambault and Rincón (2024); Patterson et al. (2025)
Evaluation of recommander systems : Raza et al. (2025).
Broad design principles for RAG for historical corpus (Puren et al. 2026) :
Concrete proposals for Archival AI assistants : Finn et al. (2026) : 7 design principles :
For Research Assistants :
Innovation and research depend heavily on our ability to make new semantic connections.
Two faces:
Thesis:
Current AI research assistants focus mainly on findability to meet basic search engine standards; however, because they inherently incorporate interpretative steps, making serendipity a core design goal could be a valuable direction to pursue.
Serendipity is a process of filtering and connecting with new information : as a design goal, it means putting at the center human judgment.
Serendipity is both the process and the result of a ‘chance encounter’.
Modeling of serendipity (Makri and Blandford 2012)
Importance of the reflexive dimension to distinguish serendipity from chance.
Exploratory and serendipitous approaches in documentary research:
“adjacent possible” (Kauffman 1996; Björneborn 2022): environmental constraints give rise to a negotiation between exploitation and exploration (Monechi et al. 2017)
exploit the profusion to generate unexpected connections (Bates 1989; Sandra Erdelez 1999)
favor disciplinary bridging (Dumas Primbault 2023)
Reproducibility and memory: how can we keep track of an exploratory path? how can we help a person recall the many steps leading to a resolution (Sanda Erdelez 2004)?
Evaluation: how can we evaluate the impact of a finding? or the interest of a new search feature (Pouyllau 2023, 2025)? how can we cater to all in a highly subjective process without falling into personalization bias (e.g. filter bubbles (Pariser 2012))?
How can we design explainable systems that are at the center of environments that favor serendipity and critical exploration? And allow for reflexive interactions with the algorithm?
Complete reports tracing the algorithmic mediation of the user’s request to the response or list of articles returned (reranking criteria, query reformulation prompt, etc.).
Expose the tool’s limitations and capabilities Example: Barista, the query builder assistant on Impresso (Finn et al. 2026)
Finn et al. (2026)
Similar to the “narrow over generalisation”.
Specialized classification assists human judgment.
Example: Evidence-RAG from the Journal of Digital History: assists the editor in presenting the evaluation of an article: situates the relevance of a comment relative to the evaluated article (Guérard et al. 2026).
Expliciting uncertainties and tool limitations.
A paradigm shift that Digital Humanities approach can embrace:
« Can we conceive of models of interface that are genuine instruments for research? That are not merely queries within pre-set data that search and sort according to an immutable agenda? How can we imagine an interface that allows content modeling, intellectual argument, rhetorical engagement? » — (Drucker 2013)
Counterfactualization: identifying which explanatory variable should change in the algorithm to obtain a different result: make alternative possibilities explicit. (Chevillon 2026)
Example: the Provotype (Boer and Donovan 2012)
(evolution theory): “opportunistic selective adaptation, favoring characteristics that are useful for a new function, for which they were not initially selected.”
In other words : allowing the tool to become the result of a misuse in the process of appropriation (Tchounikine 2025).
Example: “Misappropriation” of the search function of Gallica’s famously flawed search engine (Dumas Primbault 2025)
Generative adversarial networks (GANs) (Goodfellow et al. 2014) and antagonistic AI (Cai, Arawjo, and Glassman 2024).
Spirit of contradiction: allowing correction and evaluation by the user. Schneider (2026) tomorrow at INKE !
Example: IEML-RS, a recommander system, (Schneider, Levy, and Vitali-Rosati 2026) allows for a basic semantic decomposition of a concept in 9 fixed facets (i.e. IEML), however the “translations” done by a LLM (prompted along with a few example), are mediocre encouraging the user to correct .
IEML-RS, selection of a keyword not translated into IEML - automatic translation with gemini
Example: françaiS au pluriel (Suchet 2026)
françaiS au pluriel: tree structure
françaiS au pluriel: superposition
françaiS au pluriel: history
Highlight underexplored semantic facets to offer other paths.
Example: typology of citation type -> citation network beyond quantification
Visualization of citation disagreement on a corpus of articles from the journal Itinéraires (in development: classifications are randomized)
Visualization by citation type for an article (in development: classifications and scores randomized)
Frictionless interfaces tend to blur the lines between research and analysis whereas environments fostering a reflective process of selection help put value into the findings themselves.
This work benefited from a grant thanks to funding from the Consulat General de France au Quebec and the Fonds de recherche du Québec, which enabled a research stay at the Huma-Num Lab via the Sophie Germain mobility grant. https://doi.org/10.69777/381645
Research funded by the CRSH through the Revue3.0 partnership project as well as a grant from the Circé Network for mutualization and research for scientific journals.
An ecosystem from article writing to the queries that allow it to be found: interactions between humans and digital tools
Alexia Schneider - 2026
