Report on CHIIR 2026 Workshop on Generative AI and Academic Search (GAI&AS)

The CHIIR 2026 Workshop on Generative AI and Academic Search (GAI&AS) was held on 26 March 2026; its report, by Yifan Liu, Jaime Arguello, Orland Hoeber and colleagues, appears in ACM SIGIR Forum. It gathered researchers in human information interaction and retrieval to examine how Generative AI is reshaping academic search and scholarly practice.

The organizers' premise was that systems built for document retrieval now also summarize, recommend, synthesize, and converse, unsettling the assumption that the system locates sources while interpretation stays with the user. Because scholarship depends on traceability and credibility, academic search became a lens for revisiting learning, cognitive effort, and agency.

Synthesis: This report maps an early research agenda, not measured outcomes. Its three clusters — foundations and principles, applications and opportunities, and search-as-learning — converge on one worry: GenAI search optimizes efficiency while absorbing the interpretive work through which people learn. The classroom-relevant thread is search-as-learning and its notion of "friction," asking which (meta)cognitive processes should not be offloaded — a question tied to AI Literacy, Critical Thinking, Cognitive Offloading, Scaffolding, and Self-Regulated Learning.

Key Findings

1. Participants' sticky notes, sorted by organizers, produced three thematic clusters: (1) Foundations and Principles, (2) Applications & Opportunities, and (3) Search-as-Learning.
2. Systems must be transparent about process, purpose, and performance: how an output was generated, the user's higher-level goal, and the system's confidence.
3. The search-as-learning cluster's most popular topic was "friction": keep users engaged with critical, possibly uncomfortable (meta)cognitive processes rather than offloading them.
4. One talk found highly cited researchers reconstructed at approximately twice the rate of lower-cited peers, testing DeepSeek R1, Llama 4 Scout, and Mixtral 8×7B against OpenAlex and Google Scholar ground truth for 1,596 seed authors across 10 disciplines and 8 global regions.
5. A librarian reported a "trust gap": students often over-trusted GenAI while faculty distrusted it, and single-session workshops were judged insufficient for durable literacy.
6. AI overviews synthesized from the five top-ranked results of a social science academic search engine may reduce mental demand and frustration for some users, though overall effects were mixed.
7. Participants wanted authentic, participant-engaged methods — longitudinal and observational studies, value-sensitive design, control groups, pre- and post-task assessments — plus cross-sector partnerships.

How the workshop was organized

The call for submissions framed four interrelated themes: supporting research and education; design, evaluation, and human-centered interaction; search-as-learning in the era of GenAI; and search integrity and ethics. These covered the roles such systems play in ideation, literature review, evidence synthesis, and teaching, plus risks tied to fairness, accountability, transparency, ethics (FATE) and Cognitive Offloading. Foundations & Principles wanted trustworthy design, including transparency and recognition of biases such as confirmation bias; Applications & Opportunities named personalized summarization, on-device agents, and agentic AI for long-term needs like writing a literature review; Search-as-Learning centered on friction, scaffolding tools, AI literacy, cognitive load, and the differing needs of domain novices and experts.

Evidence the lightning talks presented

Most talks proposed designs; a few reported results. One modeled how users formulate information needs into prompts through three layers and identified three types of metacognitive laziness: plan-deficit, monitoring-deficit, and regulation-deficit. A librarian described a 15-week digital literacy curriculum built on a survey of 2,076 students and 101 librarians. A study of self-directed learning analyzed prompts issued to a GenAI system about diffusion and osmosis: participants asked for definitions, examples, and explanations, but also for work that may have been important to do themselves — differentiating concepts, deciding what to learn, simplifying earlier answers. Another talk mapped intellectual virtues such as curiosity and intellectual humility onto the ACRL Framework and warned that equity gaps for transfer and first-generation students may widen.

Principles, methods, and partnerships

The second roundtable asked how to advance human-centered GenAI academic search. On principle, participants named equal access to these tools, democratization of knowledge, technological humanism over techno-determinism, and augmenting rather than replacing human abilities; promising theories included scaffolding and nudging, friction, self-determination theory, cognitive load, and sensemaking, with a call to revisit models such as the Technology Acceptance Model. On method, most responses urged authentic, participant-engaged research that identifies impacts beyond efficiency and effectiveness, plus studies of cultural, linguistic, and neurological diversity. On partnerships, they wanted collaboration with the information, publishing, and technology companies that build these tools, and with university libraries as front-line sites for studying adoption and search literacy. A second workshop is planned, with an Academic Search and Learning with GenAI Track at IP&MC2026.

What this means for practice

• Instructors. One-shot AI training does not appear to hold: a survey of 2,076 students and 101 librarians fed a 15-week curriculum, and presenters argued single-session workshops leave a misunderstood foundation.
• Librarians. University libraries are named front-line partners for studying adoption and search literacy, and reference librarianship is proposed as a design resource.
• Learners. Proposed interfaces should preserve decision points, prompt sub-question formulation, and make confidence and sources visible, keeping verification in higher education users' hands.
• Tool builders and evaluators. The proposed agenda — pre- and post-task assessments, control groups, and evaluation of long-term effects on affect, behavior, and cognition — sets expectations for AI in Education claims a workshop report cannot supply.

Limitations

• A workshop report can support claims about what a community discussed and prioritized; it cannot support claims about what works. GAI&AS was a single event held on 26 March 2026, reflecting a self-selected pool drawn largely from human information interaction and retrieval.
• Nearly everything above is a stated opinion, design principle, or research question: the report presents no evaluation of any system, no controlled comparison, and no learning outcome data, so nothing here shows that GenAI academic search improves or harms learning.
• The figures it reports belong to individual lightning talks, not the workshop: 1,596 seed authors, 10 disciplines, and 8 global regions bound one bias study's ground truth; 2,076 students and 101 librarians bound one curriculum's survey; approximately 1,700 first-year students describe one librarian's instruction reach; five top-ranked results bound one overview evaluation.
• No attendance figures or levels of participant agreement are recorded, and clusters were assembled from sticky notes, so a theme's prominence is not a measured consensus.

Citation

Liu, Y., Arguello, J., Hoeber, O., et al. (2026). Report on CHIIR 2026 Workshop on Generative AI and Academic Search (GAI&AS).