Taking Bearings: Artificial Intelligence in Knowledge Platforms and Open, Social Scholarship/AI and Scholarship

Research Methods and Practices Arnett, Catherine, Eliot Jones, Ivan P. Yamshchikov, and Pierre-Carl Langlais. 2024. “Toxicity of the Commons: Curating Open-Source Pre-Training Data.” Preprint, arXiv, October 29. https://doi.org/10.48550/arXiv.2410.22587. Open-source pretraining data is increasingly treated as a shared resource for building “open” language models, yet it can embed and reproduce harmful speech patterns at scale. Arnett, Jones, Yamshchikov, and Langlais argue that reducing toxic model behaviour requires intervening upstream in data, and they focus specifically on the distinctive constraints of public-domain corpora (including historical documents and OCR-derived text) where standard web-text toxicity filters can be impractical or ill-suited. They propose a fully open-source curation pipeline designed for these conditions and present three concrete contributions: (1) ToxicCommons, a custom-labelled dataset organized across five toxicity dimensions (racial/origin-based, gender/sex-based, religious, ability-based discrimination, and violence); (2) Celadon, a classifier trained on that dataset to detect toxic content more efficiently at scale in open data; and (3) a “balanced” filtration strategy that explicitly trades off safety filtering against retaining sufficient training material. The paper’s core claim is that open-data model development needs domain-attuned, auditable curation methods paired with task-specific classifiers, so that openness in data does not automatically translate into avoidable harms in downstream model outputs. Bail, Christopher A. “Can Generative Artificial Intelligence Improve Social Science?” https://www.semanticscholar.org/paper/Can-Generative-Artificial-Intelligence-Improve-Bail/6206cc77bb3a3c0b6b9fce1ad68a8b1786a56941. Bail (2023) provides a critical evaluation of how Generative Artificial Intelligence (AI), particularly Large Language Models (LLMs), might fundamentally transform computational social science. Moving beyond superficial administrative use cases, Bail examines the potential of LLMs to simulate complex human behaviours. He suggests that Generative AI could revolutionize agent-based modelling by replacing simplistic, rule-bound agents with dynamic, memory-equipped "silicon samples" capable of emergent group behaviours. While acknowledging the utility of AI for automated content analysis and expanding programming accessibility, Bail rigorously details the methodological and ethical perils of these largely opaque systems. He warns that the proprietary fine-tuning of commercial LLMs introduces severe demographic biases—often skewing toward highly educated, liberal perspectives—which threatens the external validity of AI-assisted research. Furthermore, Bail highlights the "Stack Overflow Problem," cautioning that the unchecked proliferation of AI-generated "junk science" could contaminate future training data and degrade the broader scientific ecosystem. Ultimately, Bail argues that social scientists must not remain passive "end-users" of corporate AI. Instead, they must actively collaborate with computer scientists to reverse-engineer the "social sense" into AI models, ensuring the development of open-source, reproducible infrastructures that genuinely advance the study of human behaviour. Bakharia, Aneesha, Antonette Shibani, Lisa-Angelique Lim, Trish McCluskey, and Simon Buckingham Shum. 2025. “From Transcripts to Themes: A Trustworthy Workflow for Qualitative Analysis Using Large Language Models.” In Joint Proceedings of LAK 2025 Workshops, co-located with the 15th International Conference on Learning Analytics and Knowledge (LAK 2025). Dublin, Ireland: CEUR-WS.org. https://ceur-ws.org/Vol-3995/LLMQUAL_paper1.pdf. Bakharia et al. explore the use of AI in qualitative, rather than quantitative, research in education. The authors define “minimum requirements that an LLM-supported qualitative analysis workflow must satisfy” and present a “LLM-based workflow that derives an initial set of themes” from text data that is “transferable to other LA [learning analytics] contexts performing qualitative analysis of open-ended text.” The authors define two requirements for “LLM-generated inductive coding”: first, it must verify “coded textual extracts… against source data” to make sure the quotes are accurate (i.e., verbatim and not hallucinated) and verify those extracts “are meaningfully classified under the assigned code.” Second, in the interest of transparency, it must “explain that rationale for each code” and “trace every code, whatever level of abstraction, back to its source data.” The article outlines and describes the proposed workflow in detail, including its being built in Python with a Jupyter notebook and Flask application. While they acknowledge challenges and problems in their research (e.g., LLM bias, the lack of comparable human analysis, the risk of alienating the human quality of qualitative research), Bakharia et al. ultimately advocate their approach as one that “improves transparency, verifiability, and interpretability, while addressing limitations of previous methods and enhancing researcher processes in qualitative thematic analysis" (9). By adapting their minimum requirements and workflow, the authors argue, researchers can engage more effectively in LLM-driven qualitative research. Dennstädt, Fabio, Cedric Sivert Möller, Tim Fellerhoff, Felix Busch, Oke Gerke, André Karch, and Oliver Krause. 2024. “Title and Abstract Screening for Literature Reviews Using Large Language Models: An Exploratory Study in the Biomedical Domain.” Systematic Reviews 13 (1): 158. https://doi.org/10.1186/s13643-024-02575-4. Dennstädt et al. perform a study in which they attempt to determine the viability of using LLMs to automate scholarly literature surveys. The researchers developed a method for using LLMs to perform the title and abstract screenings of a systematic literature review method and applied that method with four LLMs across eleven datasets in biomedical literature. A key advantage of implementing LLMs in this process is that it escapes the need to train the machine via pre-selection. Researchers have developed automated and semi-automated processes for systematic literature reviews in the past, but this requires a human agent to provide training data or a corpus of example texts that tell the system what to look for. LLMs do not require this step. Instead, the researchers used a Python script that prompts the LLM “to evaluate the relevance of a scientific publication for inclusion into an SLR,” providing the LLM with both abstract and title, request for a numbered score indicating the relevance of the publication, and a numerical threshold that defines whether a score makes the given publication relevant. The authors characterize the results of the study as “promising” but also “far from perfect,” identifying that such applications could be helpful to researchers and are widely applicable without special training for data or user, even if “fully automated systems… still fail to differentiate… near the level of human evaluation.” The authors conclude that more research into the use of LLMs to automate systematic literature reviews is needed, but it seems very likely that scholars will employ LLMs into this integral research task with greater frequency in the future. How well LLMs will continue to perform in this task is unclear, especially in new research, and Dennstädt et al. clearly state, “we cannot answer the question of to what extent LLMs should be used for conducting literature reviews and for doing research.” Montague-Hellen, Beth. 2024. “Empowering Knowledge through AI: Open Scholarship Proactively Supporting Well Trained Generative AI.” Insights 37 (1). https://doi.org/10.1629/uksg.649. Generative AI is becoming a dominant interface for finding, summarizing, and producing academic information, with consequences for what counts as authoritative knowledge online. Montague-Hellen argues that, rather than focusing primarily on misuse, job displacement, or defensive copyright disputes, scholarly communications actors should actively encourage the inclusion of scholarly literature in generative AI training data to improve reliability and ensure research is represented in emerging discovery platforms. The article develops two main pathways for doing this: making scholarly outputs more machine-actionable and “crawlable” (especially through better linkage, metadata, and structured HTML rather than PDF-only dissemination) and making permissions unambiguous by explicitly addressing AI training in licences and related signals. It foregrounds “garbage in, garbage out” to claim that while libraries and publishers cannot remove low-quality web content from training corpora, they can tilt the balance by lowering friction for high-quality, curated research to be ingested legally. Montague-Hellen also highlights unresolved tensions around attribution and consent under common Creative Commons licences, suggesting that clearer, more specific permission frameworks (or new licensing patterns) may be needed to distinguish human reading from machine training and to communicate “enthusiastic consent” where desired. Mugaanyi, Joseph, Christoph Lehner, and Lia M. Bally. 2024. “Evaluation of Large Language Model Performance and Reliability for Citation Generation Across Scientific Disciplines.” Journal of Medical Internet Research 26: e52935. https://doi.org/10.2196/52935. Mugaanyi, Lehner, and Bally perform a study aimed at gauging the accuracy and reliability of citations generated by ChatGPT (GPT 3.5) in both the natural sciences and humanities. The authors emphasize that “Citations and references serve as the backbone of scholarly communication, providing the necessary context, evidence, and credit to prior works… ensuring the integrity of the research process.” With this in mind, the article frames the related study as aiding researchers in determining whether LLMs are viable as writing assistants in scholarly writing. In short, if an LLM cannot be relied upon to accurately cite and reference sources, it poses a risk to the integrity of research when incorporated into scholarly writing. The study found that ChatGPT hallucinated over a quarter of its references, with a slightly higher margin in the natural sciences. Furthermore, not all of those citations that did exist were accurate, and the LLM hallucinated DOIs for nearly 90% of the references it cited, real or otherwise, in the humanities. The issue, as the authors describe it, is that “in the current iteration of LLMs, since the training is geared toward generalization and the models are probabilistic, they tend to interpolate and fill in the missing information with synthetic text.” The authors conclude that, while domain-specific models may improve performance and reliability, in its current, generic form, researchers need to contemplate whether the strengths of ChatGPT’s reference generation justify its demonstrable limitations and “the importance of robust validation processes to ensure the accuracy and reliability of generated content.” Schroeder, Hope, Marianne Aubin Le Quéré, Casey Randazzo, David Mimno, and Sarita Schoenebeck. 2024. “Large Language Models in Qualitative Research: Uses, Tensions, and Intentions.” Preprint, arXiv, October 9. https://arxiv.org/abs/2410.07362. Schroder et al. perform interviews with 20 qualitative researchers in human computer interaction (HCI) and qualitative research to gauge how researchers are actually using LLMs and concerns over the incorporation of LLMs at different stages of research; the authors also provide a survey of the rise of LLMs and flag potential concerns, using both interviews and survey as a launchpad to outline suggestions and recommendations for researchers in the field contemplating the use of LLMs in their research. The authors are concerned that “the speed of LLM development has outpaced guidance on their ethical use” and “the HCI community is contending with the need for developed policies that tackle how to use AI ethically in research.” The most prominent concerns that the article raises relate to “ethics, unequal adoption of new technologies, model bias, and performance” as well as “concerns regarding participant privacy,” but there is also a larger issue within the field of HCI that “LLMs may lend the impression that qualitative inquiry can be automated, and their integration into Quantitative Data Analysis (QDA) software may increasingly impose positivist approaches that conflict with interpretivist traditions.” Key findings from the interviews are that most researchers are open to responsible use of LLMs but have concerns about established guidelines and norms for its implementation. Additionally, many are already using LLMs “to generate recruitment materials,” “speed up qualitative coding,” and “for ideation and feedback,” even as they acknowledge tensions between the qualitative nature of their work and the prospect of automation. The most significant recommendations that Schroder et al. make include updating consent forms, using dedicated tools rather than defaulting to Chat-GPT, and implementing design decisions specifically for LLM-incorporating methodologies that reconsider “participant privacy,” LLM’s “intentional use,” “transparency and validation,” “researcher context,” “deep engagement with data” despite this automation, and ways “to consider participant perspectives and interests” given “the variable performance LLMs have across contexts, knowledge domains, cultures, and languages.” The authors hope that incorporating these considerations will “empower qualitative researchers to leverage LLMs confidently, and even creatively, for their work.” Tumadóttir, Anna. 2024. “Questions for Consideration on AI & the Commons.” Creative Commons, July 24. https://creativecommons.org/2024/07/24/preferencesignals/. Tumadóttir talks about CC, license, tools, and policy and discusses the way to foster a healthy commons environment in the digital era. She points out how the introduction of the CC license enhanced Creative Commons by giving the creators choice over usage of their works. However, it is now a question if the same healthy commons can be maintained today because of rapid technological development. Preference signals for AI is a notion to give an agent (creator, rights holder, entity of some kind) more flexibility on how they want their work to be used for AI model training. However, the choice is still binary which is offering all or nothing. After consultations, it is found that people want more control over their work and if not, they might not share their work at all. Therefore, it is important to identify the right type of preference signals for it to be useful to benefit the public interest. During this process, we need to keep in mind its effect and variation on cultural heritage, different education sectors, and regions. Moreover, to make the preference signals effective, we need to examine its structure as well as whether legal enforcement is necessary. Yue, Yongjie, Maosong Zheng, Jiahui Liu, Zhimin Wang, and Chenhui Mao. 2025. “A Practical Guide and Assessment on Using ChatGPT to Conduct Grounded Theory Studies: Step-by-Step Reporting of the Data Analysis Process.” Journal of Medical Internet Research 27: e70122. https://doi.org/10.2196/70122. Yue et al. employ a study that entailed researchers using either manual coding or coding assisted by ChatGPT to perform data analysis on a 40,000 word dataset made up of interviews with 8 players of Listen and Play in Jianghu, a Chinese MMORPG designed for the blind. The study’s aim was to “provide detailed guidelines for using ChatGPT in grounded theory within the Chinese context,” “evaluate the effectiveness of ChatGPT coding” in this context, and “explore the broader implications and future directions of ChatGPT in qualitative research.” The article goes into great detail as to the process for generating open code with ChatGPT. The results of the study show that, although manual coding performed slightly better than ChatGPT-assisted coding, the “difference was not statistically significant” in its production of nodes and reference points. However, during axial coding, the categories and subcategories generated through these respective processes was significantly different, with only half of the categories semantically matching between manual and ChatGPT-assisted methods. The authors conclude that, while ChatGPT 4-Turbo “enhanced the diversity and efficiency of coding... it struggled with depth, context, subtle nuances, connections, and coding organization.” Nevertheless Yue et al. see great potential in the application of GenAI-assisted coding in grounded theory, especially as the technology develops. Zhu, Wenhao, Hongyi Liu, Qingxiu Dong, Jingjing Xu, Shujian Huang, Lingpeng Kong, Jiajun Chen, and Lei Li. 2023. “Multilingual Machine Translation with Large Language Models: Empirical Results and Analysis.” Preprint, arXiv, April 10. https://arxiv.org/abs/2304.04675. Zhu et al. perform an empirical study that tests the multilingual translation performance of eight popular LLMs on 102 languages. As the authors explain, this is a particularly difficult task as it requires “semantic alignment between languages.” While LLMs tend to perform surprisingly well at translation, “it is also unclear [sic] how LLM acquires translation ability and which factors affect LLM’s translation ability.” This study in multilingual machine translation (MMT) therefore seeks to answer two questions: how do LLMs perform MMT over massive languages and what factors affect their performance? The results of the study suggest that GPT-4 generally outperforms its competitors but still falls short of Google Translate in some tests. Two key takeaways from this study are that “exemplars in the tail of the prompt have larger influence on an LLM’s behaviour,” meaning that the order in which exemplars are given within a prompt matter, and that cross-lingual translation pairs are particularly helpful exemplars to LLMs. The authors ultimately conclude that an “LLM can acquire translation ability in a resource-efficient way, which indicates [a] promising future of LLM in multilingual machine translation” as the technology evolves. NB: As this research was conducted prior to Google’s implementation of GenAI/LLMs into Google Translate in 2024, it sometimes uses Google Translate as a baseline/comparison point for LLM-powered translation that can be confusing without that context. Forms of Research Output Bozkurt, Aras. 2024. “GenAI et al.: Cocreation, Authorship, Ownership, Academic Ethics and Integrity in a Time of Generative AI.” Open Praxis 16 (1): 1–10. https://doi.org/10.55982/openpraxis.16.1.654. Bozkurt’s essay seeks “to undertake a critical examination of the use of generative AI in academic contexts… examining its implications and exploring the nuances of its integration.” He begins with an assertion of technological determinism, stating “This technological shift… is not just a transient trend but a symbol of the inescapable change… marking the onset of an AI-dominated age and initiating profound and inevitable shifts in our academic and educational paradigms.” The author makes the argument that the advent of AI “requires us to critically reconsider concepts such as cocreation, ownership, and authorship in academic writing processes.” For instance, in exploring the concept of ownership, he briefly notes the ethical concerns of whether those who created the content upon which AI was trained might claim ownership of AI-generated content and explores the notions of AI or even its programmers acting as a co-author. Bozkurt himself takes the explicit stance that GenAI ought not be credited as a co-author. The paper also provides suggestions related to the ethics of AI use and transparency: “In some cases, merely reporting… is inadequate. A more nuanced approach involves providing multilayered statements acknowledging and benchmarking the use of [GenAI], specifying where, when, in which sections, and for what purposes it is employed.” Bozkurt reiterates that responsibility ultimately belongs to the human author, and provides the aiTARAS (Academic Integrity and Transparency in AI-assisted Research and Specification) Framework for this purpose. Bozkurt ends his essay by identifying further problems beyond the scope of the article, including “reimagining of assessment and evaluation,” our overfocus on ChatGPT and English language in this field of study, and the inaccuracy of AI detection software. Colbert-Lewis, Danielle, Lawrence Maminta, Kelly McElroy, Graeme Slaght, and Mark Swartz. 2024. “The Citation Economy as a Site of Extraction for Surveillance Publishing.” Canadian Journal of Academic Librarianship 10 (December): 1–22. https://doi.org/10.33137/cjal-rcbu.v10.43293. Colbert-Lewis et al. (2024) examine the extractive nature of the modern "citation economy," arguing that academic publishing has evolved into a powerful mechanism for surveillance and data commodification. The authors detail how a consolidated group of dominant publishers capture surplus value from the academic lifecycle by extracting free scholarly labour, research data, and personal information. Crucially, these corporations are transitioning from traditional information vendors into technology-driven data brokers. The extracted data is reinvested into proprietary analytics products that are then sold back to universities to evaluate and surveil faculty performance, relying on opaque metrics that threaten academic freedom. Furthermore, the authors highlight the severe ethical compromises inherent in this system, noting that academics inadvertently fuel data infrastructures that these same parent companies sell to external industries, including law enforcement and advertising. To combat the rise of "surveillance publishing," the article concludes with actionable recommendations for scholars and librarians to resist exploitative practices and build ethical infrastructures, alongside a supplementary mini-zine designed to raise awareness about the hidden costs of participating in the citation economy. Frangou, Sophia, Umberto Volpe, and Andrea Fiorillo. 2025. “AI in scientific writing and publishing: A call for critical engagement.” European Psychiatry 68 (1): e98. https://doi.org/10.1192/j.eurpsy.2025.10061. Frangou, Volpe, and Fiorillo provide a succinct summary of the benefits of implementing AI in scientific writing and publishing, before turning to the challenges and risks posed toward the use of such technology. Finally, they end by posing ethical frameworks for the use of AI in both publishing and writing. In publishing, the authors identify tools that can aid in tasks ranging from copyediting to finding reviewers and claim that for both “high-volume publications” and “resource-constrained journals… AI can serve as a force multiplier, expanding what editorial teams can accomplish without compromising the centrality of human discretion and responsibility.” In writing, the article notes that “AI-powered applications offer a suite of tools that, when used judiciously, can enhance the quality, efficiency, and inclusivity of the scholarly communication process,” placing such tools along the trajectory of preexisting aids like EndNote and Zotero, which have also now incorporated AI. Frangou, Volpe, and Fiorillo argue that AI can help scholars identify relevant literature they might otherwise have trouble finding, reduce the linguistic bias encountered by scholars for whom English is not their first language, and aid scholars in navigating the publishing landscape so their work can find the best possible fit. Finally, the authors acknowledge the epistemic and ethical risks of AI: inconsistency and irreproducibility, transparency, informed consent, and data privacy are all core concerns. Key components in the article’s ethical frameworks for publishers and authors include transparency and disclosure, human accountability, and training and skill development. The authors hope that “these principles articulate a shared responsibility for shaping the role of AI in scientific publishing” and AI’s “adoption reinforces the foundational values that give scientific communication its trustworthiness and legitimacy.” Graßhoff, Gerd. 2024. “AI-Reporter: A Path to a New Genre of Scientific Communication.” Preprint, arXiv, July 8. https://arxiv.org/abs/2507.05903. Graßhoff introduces a new tool, the AI-Reporter, that aims “to fundamentally expand scientific communication for the new era…”; according to Graßhoff, the AI-Reporter “creates a sustainable, referenceable, and expandable knowledge base that captures not just content but the living essence of scientific discourse.” In short, Graßhoff proposes a tool that will adapt a scientific presentation into a “public-ready chapter” in about three minutes with “only the author’s consent.” The majority of the article is dedicated to a tech-oriented breakdown of the methodology and workflow of the AI-Reporter, namely the semantic analysis and translation of a recorded slide presentation—consisting of the presentation as a PDF, video recording, and basic metadata—into a publication-ready chapter. Graßhoff’s tool is an attempt to solve what he identifies as a core problem of modern scholarly communication, namely that “knowledge is increasingly presented in dynamic, multimodal formats” that are ultimately ephemeral and content is often lost (14). He aims to refine this tool to enable real-time processing, multilingual support, interactive multimedia components, “optimization for academic disciplines,” and more to this project. He advocates the AI-Reporter as nothing less than “a vision for the future of scientific communication.” Jiang, Jialei. 2024. “When generative artificial intelligence meets multimodal composition: Rethinking the composition process through an AI-assisted design project.” Computers and Composition 74: 102883. DOI: https://doi.org/10.1016/j.compcom.2024.102883. Jiang conducts a study in which GenAI is integrated into students’ composition of multimodal texts, then interviewed on their experience. Jiang’s states her aim as follows: “Through examining writing students use of GenAI tools in multimodal composition, this study seeks to unravel how GenAi technologies influence students’ design choices, problem-solving approaches, and the overall composition process.” In particular, Jiang is interested in “potential opportunities and challenges of incorporating GenAI into students’ multimodal composition process.” Jiang frames the article within composition and writing studies and the theory behind multimodal composition. She builds upon recent pedagogical best practices put forward by Burriss and Leander (2024) that “call upon teachers and educators to engage in the development of an emergent pedagogy of critical AI that “teach[es] about/with AI in emergent, flexible, and speculative ways.” The results of the study emphasize that, while GenAI can streamline workflow and provide examples from which students can build using their own creativity, there were clear limitations to the technology. Of course, students’ realization of these limitations is a success of the course’s pedagogy. The article also provides three case studies of the student compositions as exemplars of the outcome. Jiang concludes with a recommendation “that researchers and educators engage in meaningful dialogues with students about their uses of GenAI during composition practices.” She quotes her own earlier work (Jiang et al. 2024), where she writes that to keep AI out of the classroom “is not only idealistic and impossible... it is actually completely disengaged from the realities of the changing technological landscape students are already facing.” Mehlenbacher, Brad, Ana Patricia Balbon, and Ashley Rose Mehlenbacher. 2024. “Synthetic Genres: Expert Genres, Non-Specialist Audiences, and Misinformation in the Artificial Intelligence Age.” Journal of Technical Writing and Communication, January 31. https://doi.org/10.1177/00472816231226249. Genre theorists Mehlenbacher et al. examine which AI-generated texts are “good enough” to pass as “bona fide” (already a “very good” standard). They argue AI-generated text produces misinformation, pushes apart definitions of “information” and “knowledge,” and necessitates rhetorical understanding. The authors therefore believe “synthetic genres” will emerge, departing in situation, form, and even “the very concept of genre users” based on recursive and “poisoned” outputs. In reviewing literature, the authors spotlight AI-generated disinformation’s suasive (instead of propagandistic or accurate) potential, and argue that AI-generated text can be suasive by appearing both “timely” and “appropriate.” However, this “fraudoscientific” text can never be truly “timely” or “appropriate,” as it is always constrained in past work, and only responds to prompts, not situations. Despite this, AI-generated text generated on specialist subjects can still be effectively suasive—and even when identified as AI—because its authority is difficult to challenge. Mehlenbacher et al. therefore conduct two studies of AI-generated outputs, focusing on how they may be deceptive or detected. They first prompted GPT-3 to emulate rhetorical research abstracts by generating a range of definitions alongside text to emulate research processes. Next, they conducted a genre analysis, generating text on controversial topics and prioritizing suasiveness over ability to pass a theoretical “Genre Turing Test.” Based on their results, Mehlenbacher et al. suggest AI-generated text should be cross examined with specialist texts to identify where they need additional development, presenting a valuable teaching opportunity which supports “a powerful way to introduce the epistemic functions of writing.” Non-specialists should also develop tools to identify the source and quality of content, including diverse genre perspectives. The authors conclude by arguing genre scholars should further consider AI-generated texts and emphasizing the importance of genre users. This is because “genre must be understood in a rhetorical context,” while “synthetic genres” can only act as “statistical simulacrum.” Teaching and Pedagogy Adewumi, Tosin, Lama Alkhaled, Christopher Buck, Sergio Hernandez, Sofia Brilioth, Memoona Kekung, Yasmin Ragimov, and Emily Barney. 2023. “ProCoT: Stimulating Critical Thinking and Writing of Students through Engagement with Large Language Models (LLMs).” Preprint, arXiv, December 15. https://doi.org/10.48550/arXiv.2312.09801. Adewumi et al.'s Probing Chain-of-Thought (ProCoT) methodology represents a pedagogical approach that transforms LLM interaction from potential academic dishonesty into an active learning process centred on critical evaluation and evidence-based reasoning. ProCoT requires students to generate initial outputs using LLMs, then systematically affirm or refute each claim using peer-reviewed references, effectively making the AI output a provisional hypothesis to be tested rather than a final answer to be submitted. This approach directly addresses the section's themes of process transparency (the method makes visible each stage of knowledge construction, from initial AI generation through source verification to synthesis) and ethical reflection (students must grapple with AI limitations including hallucination, bias, and lack of disciplinary nuance). The authors' finding that student outputs using ProCoT were significantly more concise than LLM-generated text (208 vs. 391 words on average) while demonstrating enhanced critical thinking suggests that the method trains students to distill and synthesize rather than accept verbose AI output uncritically. ProCoT leverages AI's epistemological weaknesses as pedagogical strengths—the fact that ChatGPT cannot reliably cite sources becomes an opportunity for students to develop information literacy by finding and evaluating primary literature. The method's anti-cheating design is secondary to its pedagogical value: by requiring iterative engagement with AI outputs and scholarly sources, ProCoT embeds the kind of metacognitive reflection (What did the AI get right? What did it miss? How do I know?) that is of significance to pedagogy in the AI era. The authors' evidence from 65 students across disciplines demonstrates feasibility across contexts, emphasizing cross-disciplinary rather than field-specific approaches to AI-entangled pedagogy. Berry, David M. 2023. “AI, Ethics, and Digital Humanities.” In The Bloomsbury Handbook to the Digital Humanities, edited by James O’Sullivan, 445–57. Bloomsbury Academic. https://recoveryhub.siue.edu/wp-content/uploads/2024/10/AI-Ethics-and-Digital-Humanities-.pdf. Berry emphasizes that AI’s use in the field of Digital Humanities raises major ethical questions despite it being an apparent evolution of the field’s traditional utilization of technology to augment human research capabilities. Berry outlines this issue by comparing the modern rise of AI to Digital Humanities’ early history, highlighting that digital humanists operated as coders until they began to “'black box' the computational aspects of doing digital humanities,” thus enabling non-programmers to participate. Since AI “automating processes might cover over ethical issues by transferring them into the hashtables of the machine-learning system,” Berry argues uncritical reliance on AI tools by inexpert modern humanists risks the “algorithmization” and imposition of hierarchical, quantitative frameworks upon Digital Humanities studies. He further identifies that the field of Digital Humanities is “on the cusp of a new set of packages that will further democratize access to machine learning,” a noble goal, but one which will bring these issues to the forefront. For this reason, Berry argues a reconsideration of ethics within Digital Humanities is desperately required on both organizational and individual levels. Brown, Richard. 2023. The AI Generation: How Universities Can Prepare Students for the Changing World. DEMOS and University of London report. https://demos.co.uk/wp-content/uploads/2023/11/The-AI-Generation-2.pdf. Brown argues AI will inevitably and dramatically impact all fields of education, and therefore encourages universities to seize the opportunity to lead these changes instead of merely reacting to them. His voice is echoed by esteemed organizations DEMOS and the University of London, indicating that this shift in the academic environment is already beginning. Brown argues this shift is required due to the developing capacity of AI for automating low-skilled professional tasks, an ability which risks decimating the available number of entry-level graduate positions. As a result, how employers value the workplace skills taught by current curriculums will change drastically, meaning universities must revolutionize their approach to education. Brown provides a blueprint for how universities can adapt to these changes by championing education’s prioritization of “GRASP” (General Relational, Analytic, Social, and Personal) skills as an alternative, advising universities to focus on teaching high-level and tailored skills to students instead of the generic. Universities can do this by practising “active learning” over traditional lecture formats as well as fostering extracurricular and work experience opportunities for their students. Deng, Ruiqi, Maoli Jiang, Xinlu Yu, Yuyan Lu, and Shasha Liu. 2025. “Does ChatGPT Enhance Student Learning? A Systematic Review and Meta-Analysis of Experimental Studies.” Computers & Education 227. https://doi.org/10.1016/j.compedu.2024.105224. Deng and colleagues' meta-analysis of 69 experimental studies provides empirical grounding for claims about ChatGPT's pedagogical impact, revealing both opportunities and methodological challenges in assessing AI-enhanced learning. Their finding that ChatGPT interventions improve academic performance, affective-motivational states, and higher-order thinking propensities while reducing mental effort speaks directly to how assessment must shift when AI augments cognition—if mental effort decreases but learning outcomes improve, traditional measures of "effort" or "struggle" as proxies for learning may require recalibration. In their critique of current assessment methodologies the authors demonstrate that most studies evaluate only final outputs (essay quality, test scores) rather than learning processes, missing opportunities to examine how students interact with AI, what metacognitive strategies they develop, and whether AI use builds transferable skills or creates dependency. Deng et al.'s four methodological propositions—using complex project-based assessments that reveal process, evaluating long-term rather than novelty effects, prioritizing objective over self-reported measures of higher-order thinking, and employing adequate statistical power—provide a research agenda for pedagogy–AI scholarship that centres process evidence. Their documentation of disciplinary variation (language education dominates current research) while calling for cross-disciplinary synthesis focuses on avoiding discipline-specific approaches. The meta-analysis also reveals gaps in understanding collaboration (few studies examine peer learning dynamics when AI is present) and metacognition (limited research on whether students develop awareness of their own and AI's knowledge boundaries), highlighting the resulting pedagogical transformations. Grigoli, L. Renato. 2023. “Ghosts in the Machine.” American Historical Association 61 (3). https://www.historians.org/wp-content/uploads/2024/07/Perspectives_61N3.pdf. Grigoli dismisses fears of AI generated essays threatening academic integrity by arguing that “if the development of artificial intelligence results in the death of the humanities, then it will be because it will have shown that the emperor has no clothes.” He uses ChatGPT to demonstrate that although AI responses appear competent at a passing glance, they lack valuable substance, simply arranging facts in a way that “tricks the reader into… doing all the analytical work.” Clear critical analysis should be the core component of a successful humanities essay, making well-designed assessments of this field far more reliable than those focused on information retention (like engineering). Therefore, AI is not a real threat, but a valuable pedagogical tool which can demonstrate examples of clean prose and prompt students to critically analyze what constitutes critical analysis. Essentially, if an AI generated answer can pass a test, the real problem was "self-inflicted" by either the examiner or the question itself. Lee, Daniel, and Edward Palmer. 2025. “Prompt Engineering in Higher Education: A Systematic Review to Help Inform Curricula.” International Journal of Educational Technology in Higher Education 22: 16. https://doi.org/10.1186/s41239-025-00503-7. Lee and Palmer's systematic review establishes prompt engineering as an emergent pedagogical practice that integrates with longstanding traditions of scholarly inquiry, questioning, and rhetoric. The authors synthesize multiple frameworks (AIPROMT, CLEAR, CRISPE) that codify prompting as a learnable skill involving role specification, context provision, instruction clarity, and iterative refinement—revealing how interaction with LLMs mirrors Socratic dialogue, research question formulation, and the rhetorical tradition of audience awareness. This work positions "prompting as pedagogy" as a reconfiguration of existing scholarly literacies: students learning to prompt effectively must articulate their information needs precisely, anticipate how language shapes output, and iteratively refine queries based on initial results—all core practices in library research, database searching, and scholarly conversation. The authors document how educators are embedding prompt engineering across disciplines, from technical fields where students prompt AI to generate code or solve equations, to humanities contexts where prompting becomes a form of textual analysis (understanding what linguistic patterns trigger particular AI responses reveals implicit biases and training data influences). Their review also identifies ethical dimensions—well-designed prompts can mitigate AI hallucination and bias, while poorly constructed prompts amplify these issues—making prompt literacy a matter of scholarly responsibility. The authors' call for curricula that teach prompting as interdisciplinary competency aligns with how AI entangles with pedagogy across boundaries, reshaping what counts as fundamental scholarly practice in an AI-augmented knowledge environment. Mai, Dang Thi Thuy, Cuong Van Da, and Nguyen Van Hanh. 2024. “The Use of ChatGPT in Teaching and Learning: A Systematic Review through SWOT Analysis Approach.” Frontiers in Education 9: 1328769. https://doi.org/10.3389/feduc.2024.1328769. Mai and colleagues' systematic SWOT analysis of 51 studies on ChatGPT in education, organized through Biggs' 3P model (Presage-Process-Product), offers comprehensive mapping of how generative AI enters and transforms pedagogical systems at multiple scales—from individual learner characteristics (Presage) through interactive teaching-learning processes (Process) to learning outcomes and assessment products (Product). This multilevel framework reveals that ChatGPT's pedagogical impact cannot be isolated to single moments of use but ripples across entire learning ecologies: at the Presage level, ChatGPT shifts what prior knowledge students need (familiarity with prompting interfaces) and what instructor competencies are required (ability to design AI-aware assignments); at the Process level, it transforms interaction patterns from teacher–student and student–student to include human–AI dialogue that may enhance personalized scaffolding but risks reducing peer collaboration; at the Product level, it necessitates fundamental assessment redesign because traditional evaluation instruments (timed essays, closed-book exams) lose validity when AI can generate competent responses. The authors' characterization of ChatGPT as simultaneously "friend" (enabling personalized learning, reducing educator workload, providing instant feedback) and "foe" (enabling plagiarism, potentially reducing critical thinking, creating over-reliance) mirrors the section's emphasis on entanglement—pedagogy cannot simply embrace or reject AI but must negotiate tensions between efficiency and effort, personalization and depersonalization, augmentation and replacement. Mai et al.'s documentation of how educators are adapting curricula to emphasize creativity and critical thinking (skills less easily automated) while using ChatGPT for routine information provision demonstrates pedagogical evolution toward distinctly human capacities. Their call for assessment reform using complex case studies, portfolios, and process documentation rather than single-sitting exams directly supports the notion of process evidence and metacognition as being central to AI-era pedagogy, positioning assessment not as gatekeeping but as making learning processes transparent and accountable. Oates, Angela, and Donna Johnson. 2025. “ChatGPT in the Classroom: Evaluating Its Role in Fostering Critical Evaluation Skills.” International Journal of Artificial Intelligence in Education 35 (4): 1793–1824. https://doi.org/10.1007/s40593-024-00452-8. Oates and Johnson's empirical study of biomedical science master's students offers insight into how AI functions more effectively as object of critical study than as production tool, revealing pedagogical value in making AI outputs subjects for evaluation rather than endpoints of learning. Their finding that students' marks did not improve when submitting AI-generated essays but did improve when critically evaluating AI outputs demonstrates a fundamental pedagogical principle: learning occurs not in consuming AI-generated content but in interrogating it, comparing it against disciplinary standards, identifying its factual errors and rhetorical limitations, and articulating why human-authored scholarship differs. This shifts pedagogy from "use AI to complete tasks" toward "use AI to understand knowledge construction itself"—students learn about evidence evaluation, citation practices, and argumentation by analyzing where ChatGPT succeeds and fails at these scholarly fundamentals. The authors document that ChatGPT demonstrated structural coherence and grammatical accuracy but lacked the disciplinary depth and synthetic insight expected in graduate work, making visible to students the difference between surface-level fluency and genuine expertise. Their emphasis on user interaction as a variable affecting efficacy underscores that pedagogy must address not just whether AI is present but how students are taught to engage it—prompt design, output evaluation, and integration with human research become learnable scholarly practices. Oates and Johnson's conclusion that students preferred writing their own essays despite AI's availability suggests that when pedagogy centres process transparency and critical evaluation rather than output efficiency, students recognize and value the cognitive work that AI cannot replicate, focusing on pedagogy that treats AI as entangled phenomenon requiring simultaneous use and critique. Pope, Andrew, and Rongqian Ma. 2024. “Exploring Historians’ Critical Use of Generative AI Technologies for History Education.” Proceedings of the Association for Information Science and Technology 61 (1): 1071–73. https://doi.org/10.1002/pra2.1188. Pope and Ma conducted semi-structured interviews of seven history professors to assess their attitudes toward employing GenAI tools in their teaching. Understanding this stance is important because Pope and Ma believe incoherent and mismatched approaches could inhibit scholarly communication and undermine academic integrity. The survey demonstrated historians had major concerns about GenAI’s development, especially regarding plagiarism. Respondents were particularly divided on whether copying an AI’s work was academic misconduct. However, historians had far fewer concerns about using GenAI to "augment" human abilities or tailor support to students—although senior academics maintained substantially greater reservations. The historians that were interviewed concurred that critiquing AI generated text could improve a student’s media literacy. Pope and Ma’s study suggests that historians consistently doubt the quality of AI generated responses even when willing to employ it. Tan, Myles Joshua Toledo, and Nicholle Mae Amor Ttan Maravilla. 2024. “Shaping Integrity: Why Generative Artificial Intelligence Does Not Have to Undermine Education.” Frontiers in Artificial Intelligence 7: 1471224. https://doi.org/10.3389/frai.2024.1471224. Tan and Maravilla provide theoretical grounding for understanding how generative AI can be integrated into pedagogy without compromising academic integrity, arguing that responsible implementation depends on alignment with constructivist learning theory and self-determination theory. The authors position AI not as a threat to authentic learning but as a catalyst that necessitates pedagogical transformation—shifting from transmission models toward environments where students actively construct knowledge using AI as one tool among many. Their framework emphasizes process transparency through explicit discussion of AI's role in knowledge construction and ethical reflection through examining how GAI outputs are generated, what assumptions they embed, and where they may mislead or constrain inquiry. This work addresses pedagogy-as-entanglement by showing how GAI forces educators to make visible the epistemic practices that traditional assessment often left implicit: citation tracing, source evaluation, argument construction, and the iterative nature of scholarly writing. Tan and Maravilla argue that rather than banning AI to preserve integrity, educators should redesign assessment to require demonstration of process—portfolios showing prompt refinement, comparison of AI outputs with human-authored sources, and metacognitive reflection on when and why AI was consulted. Their synthesis of educational theory with practical implementation strategies makes this work useful for understanding how pedagogy must evolve to treat AI as both subject matter (what students must understand about how AI functions) and scaffolding (how AI can support learning when used transparently and reflectively). Yim, Iris Heung Yue, and Jiahong Su. 2025. “Artificial Intelligence (AI) Learning Tools in K-12 Education: A Scoping Review.” British Journal of Educational Technology 56 (1): 169–196. https://doi.org/10.1007/s40692-023-00304-9. Yim and Su's scoping review of AI literacy education in K-12 contexts provides grounding for understanding how pedagogy approaches AI as an object of study, not merely as a tool, revealing how students across age groups learn to understand, critique, and create with AI systems. The authors document how intelligent agents (Google Teachable Machine, Learning ML) and software platforms (Scratch, Python) enable students to engage with AI's underlying mechanisms—training models, observing how data shapes outputs, and experiencing firsthand how algorithmic decision-making operates. This approach treats AI as epistemological phenomenon: students don't just use pretrained models but build simple systems themselves, making visible how AI "learns" and where its learning breaks down (limited training data, biased datasets, inability to generalize beyond training conditions). Yim and Su's synthesis of pedagogical strategies—project-based learning where students create AI applications, human-computer collaboration examining how humans and machines complement each other, and game-based approaches making AI concepts accessible—demonstrates how educators scaffold understanding from concrete manipulation to abstract reasoning about intelligence and automation. The review's attention to cognitive, affective, and behavioural learning outcomes moves beyond narrow skill acquisition to examine how AI literacy shapes students' broader epistemic stance: understanding AI's capabilities and limitations, recognizing when algorithmic solutions are appropriate, and developing critical awareness of AI's social implications. Their emphasis on age-appropriate pedagogies and the importance of unplugged activities (learning AI concepts through physical manipulation before digital implementation) provides models for how pedagogy must adapt to learners' developmental stages when teaching about AI systems. This work's K-12 focus complements higher education studies in this section by showing how foundational AI literacy built through hands-on exploration creates readiness for more sophisticated critical engagement with AI as both tool and object of study in advanced scholarship. Service and Peer Review Checco, Alessandro, Lorenzo Bracciale, Pierpaolo Loreti, Stephen Pinfield, and Giuseppe Bianchi. 2021. “AI-Assisted Peer Review.” Humanities and Social Sciences Communications 8: 25. https://doi.org/10.1057/s41599-020-00703-8. Checco et al.'s study represents crucial early (pre-GenAI) empirical work demonstrating that AI can predict peer review outcomes based on "superficial" manuscript features—readability metrics, formatting consistency, reference list structure, linguistic patterns—with surprising accuracy, training neural networks on 3,300 conference papers to correlate these proxy measures with eventual accept/reject decisions. This finding has profound implications for understanding AI's operational role: the research reveals that much of what passes as peer review judgment operates on detectable patterns (clear writing, proper citation formatting, adherence to disciplinary conventions) that AI can identify and assess, potentially automating routine quality checks and flagging submissions unlikely to meet standards. However, the authors explicitly position their work as supporting "semi-automated" rather than fully automated review, emphasizing AI's role in pre-screening and administrative tasks (matching manuscripts with reviewers based on topic modelling, identifying obvious deficiencies) rather than intellectual evaluation. The paper's most valuable contribution lies in its systematic exploration of ethical implications: algorithmic bias risks (AI might penalize unconventional but innovative work, discriminate against non-native English writers, favour institutional prestige markers), transparency requirements (making AI decision criteria auditable), and the necessity of human oversight to prevent automation from calcifying existing disciplinary hierarchies. By documenting both AI's predictive capabilities on operational dimensions and its inherent limitations on substantive judgment, Checco et al. provide empirical foundation for designing peer review systems where AI handles standardized identification tasks while human expertise remains authoritative for evaluative assessment, directly instantiating the core principle that operational assistance must not migrate into epistemic territory. Hosseini, Mohammad, and Serge P. J. M. Horbach. 2023. “Fighting Reviewer Fatigue or Amplifying Bias? Considerations and Recommendations for Use of ChatGPT and Other Large Language Models in Scholarly Peer Review.” Research Integrity and Peer Review 8 (1): 15. https://doi.org/10.21203/rs.3.rs-2587766/v1. Published early in the ChatGPT era, Hosseini and Horbach's analysis provides a prescient examination of how LLM integration transforms both the pragmatics and ethics of peer review labour. The authors document LLMs' potential to combat reviewer fatigue by automating time-consuming tasks—transforming informal reviewer notes into polished reports, generating structured feedback on manuscript sections, identifying linguistic or formatting issues—thereby potentially expanding the pool of contributors who can participate effectively despite language barriers or time constraints. However, the article's core contribution lies in its still-useful systematic identification of risks that emerge when operational assistance crosses into epistemic territory: LLMs trained on existing literature may amplify disciplinary biases (favouring established paradigms over novel approaches), geographic biases (privileging research contexts well-represented in training data), or methodological orthodoxies (flagging unconventional designs as errors rather than innovations). The authors demonstrate through examples how ChatGPT can generate cynical or biased reviews that violate Mertonian norms of universalism, and how confidentiality breaches can occur when reviewers input manuscript excerpts into external AI platforms without institutional safeguards. Hosseini and Horbach's recommendations—mandatory disclosure of LLM use in reviews, human accountability for all AI-generated content, training in bias recognition, and institutional policies prohibiting upload of confidential materials—have influenced subsequent journal guidelines and provide foundational ethical framework for understanding why guardrails against leakage and disclosure norms become operational necessities rather than optional best practices in AI-entangled peer review systems. Shcherbiak, Anna, Hooman Habibnia, Robert Böhm, and Susann Fiedler. “Evaluating Science: A Comparison of Human and AI Reviewers.” Judgment and Decision Making 19 (2024): e21. https://doi.org/10.1017/jdm.2024.24. In this study, a large-scale field experiment comparing GPT-4 with human reviewers on conference abstracts provides empirical evidence about AI's capabilities and limitations in evaluative judgment. It demonstrates that while AI can approximate human performance on certain classification tasks—identifying "very best" abstracts shows moderate alignment—detailed evaluative assessments reveal persistent gaps, with human-AI agreement comparable to human-human variability, suggesting AI does not systematically outperform baseline reviewer disagreement. Critically, the research shows humans substantially outperform AI at detecting AI-generated versus human-written content, with detection tools like GPTZero exhibiting higher accuracy than GPT-4 itself when evaluating authorship. This finding has direct implications for peer review integrity when AI-generated manuscripts enter the submission pipeline. The authors position AI as effective for prescreening—rapidly filtering submissions for basic quality thresholds, identifying obvious errors, flagging compliance issues—while demonstrating it lacks the contextual understanding necessary for nuanced scientific judgment about significance, impact, or methodological soundness. The paper's methodological rigour in isolating human-versus-AI performance dimensions provides context for understanding where operational assistance (screening) legitimately ends and where human evaluative authority (substantive assessment) must begin, directly addressing the guiding principle that human judgment remains central even in AI-augmented workflows. Lim, Gim Hui, M. L. Tan, V. C. W. Hoe, and D. Koh. 2025. “Generative AI in Peer Review Process for Occupational Health.” Occupational Medicine 75 (5): 242–49. https://doi.org/10.1093/occmed/kqaf051. Lim and colleagues' empirical study comparing Microsoft Copilot, ChatGPT, and Google Gemini against human reviewers on eight occupational health manuscripts provides granular quantitative evidence about AI's operational capabilities and limitations in peer review tasks. The research demonstrates that AI tools significantly outperform humans in providing feedback (mean score 3.44 vs. human baseline, p<0.001) across dimensions of relevance, completeness, accuracy, error identification, and constructiveness—documenting AI's strength in systematic checking tasks like identifying missing citations, flagging methodological inconsistencies, noting formatting errors, and pointing out unclear explanations. However, humans substantially outperform AI in generating actionable recommendations (mean score 3.36, p<0.01), with AI showing particular deficiencies in suggesting substantive revisions, connecting findings to broader literature, or identifying conceptual limitations. This performance asymmetry instantiates the operational/epistemic division: AI excels at identification tasks (what's wrong, what's missing) but struggles with evaluative tasks (how to improve, what matters). The study also quantifies efficiency gains (AI reviews complete in 11 minutes versus 45 for humans) while documenting critical limitations: AI outputs contain fabricated references, generate plausible-sounding but inaccurate technical assertions, and require human verification to prevent propagation of errors. Lim et al.'s work demonstrates that even when AI shows superior performance on specific metrics, its integration into peer review requires careful task decomposition—leveraging speed and comprehensiveness for checking functions while preserving human authority for substantive guidance—making it essential evidence for designing hybrid human-AI workflows where operational assistance enhances rather than replaces evaluative expertise. Sun, Zhuanlan. 2025. “Large Language Models in Peer Review: Challenges and Opportunities.” Scientometrics 130 (10): 5503–5546. https://doi.org/10.1007/s11192-025-05440-w. Sun provides essential mapping of LLM applications across the peer review lifecycle, categorizing five distinct operational roles: checklist assistants for formatting and protocol adherence, reviewer selection aids that match manuscripts with appropriate expertise, feedback generators for preliminary assessments, bias detectors that flag methodological or statistical irregularities, and agents that coordinate multi-stage review workflows. The article systematically examines technical approaches including prompt engineering strategies, model evaluation frameworks, and architectural designs for integrating LLMs into editorial management systems. The author argues that while LLMs excel at standardized operational tasks—checking reference integrity, identifying duplicated content across submission databases, verifying compliance with reporting guidelines—they remain fundamentally limited in assessing research novelty, theoretical contributions, and domain-specific methodological rigour. The analysis emphasizes that current LLM limitations (inadequate scientific validation, domain knowledge gaps, inability to analyze complex datasets, ethical concerns around bias perpetuation) position them as supportive tools within human-led processes rather than autonomous decision-makers. This work is useful in understanding operational versus epistemic divisions of labour in AI-assisted scholarship service and documenting the technical infrastructure through which AI becomes entangled with peer review operations while maintaining clear boundaries around human evaluative judgment. AI and Scholarship Infrastructures Organizational Infrastructures Batool, Amna, Didar Zowghi, and Muneera Bano. 2025. “AI Governance: A Systematic Literature Review.” AI and Ethics 5: 3265–3279. https://doi.org/10.1007/s43681-024-00653-w. Batool and colleagues' (2025) systematic literature review offers a comprehensive mapping of AI governance across multiple levels—team, organization, industry, national, and international. The authors utilize a structured analytical framework that examines who governs (stakeholders and roles), what is governed (data, algorithms, systems), when governance occurs (stages in the AI lifecycle), and how it is implemented (frameworks, tools, policies). The article reveals that current governance practices remain fragmented and inconsistent, with significant gaps at the national and international levels. Applied to the context of higher education, this multi-level perspective is essential for understanding how research institutions must position themselves within broader governance ecosystems. Because organizational governance currently operates without clear external scaffolding, research institutions are often forced to improvise institutional arrangements rather than implement established templates. While the paper is broad in scope, its categorization of governance artifacts illuminates the range of mechanisms available to scholarly organizations: from technical tools like algorithmic auditing to organizational structures like ethics committees. Ultimately, the authors' findings highlight the challenges facing research institutions as they attempt to coordinate AI oversight across distributed, autonomous units—such as libraries implementing discovery tools or IT departments deploying infrastructure—demonstrating why cross-functional coordination is both necessary and difficult when existing governance structures were not designed to span these boundaries. Hadley, Emma, Alan Blatecky, and Megan Comfort. 2025. “Investigating Algorithm Review Boards for Organizational Responsible Artificial Intelligence Governance.” AI and Ethics 5: 2485–2495. https://doi.org/10.1007/s43681-024-00574-8. Hadley and colleagues provide empirical documentation of algorithm review boards (ARBs) as organizational governance mechanisms, interviewing technical contributors across sectors to examine ARB membership, scope, success factors, and limitations. Their findings reveal ARBs as cross-functional bodies integrating diverse expertise (technical specialists, ethicists, domain experts, legal advisors, non-specialist stakeholders) to review AI systems for potential risks and harms, operating alongside other responsible AI approaches like policies, audits, and dedicated governance roles. The study's key insight that institutional review boards alone prove insufficient for algorithm governance and that ARBs function most effectively when integrated with existing organizational processes rather than operating as isolated oversight bodies speaks directly to research institutions' governance challenges. The article's documentation of financial tensions between profit motives and responsible AI costs, while focused on commercial contexts, translates to research institutions facing pressure to adopt AI for efficiency gains while managing ethical risks with constrained resources. Hadley et al.'s call for standardized ARB effectiveness metrics points to the broader challenge of demonstrating governance value in organizations where responsible AI practices compete with other institutional priorities for attention and investment. Janssen, Marijn. 2025. “Responsible Governance of Generative AI: Conceptualizing GenAI as Complex Adaptive Systems.” Policy and Society 44 (1): 38–51. https://doi.org/10.1093/polsoc/puae040. Janssen's conceptual article reframes generative AI governance as managing complex adaptive systems characterized by co-evolution, emergent properties, and non-linear interactions between technical and social elements, directly challenging technology-deterministic approaches that treat AI as static tool requiring one-time organizational accommodation. This systems perspective is essential for understanding why research institutions struggle with AI governance: the article demonstrates how AI systems and organizational contexts mutually shape each other through feedback loops, making governance an ongoing adaptive process rather than implementation of fixed policies. Janssen argues that effective organizational stewardship requires holistic, outward-focused governance attending to how AI systems interact with broader organizational processes, public values, and societal concerns—moving beyond narrow risk mitigation to address joint accountability across system components including people, policies, data, and algorithms. For research institutions embedding AI in knowledge infrastructures, this framing illuminates why isolated departmental responses prove insufficient: cataloging systems using AI for metadata generation, research platforms deploying AI for literature synthesis, and administrative systems using AI for resource allocation together constitute an organizational AI ecosystem whose emergent behaviours cannot be governed through component-level oversight alone. The article's emphasis on selecting and combining appropriate policy instruments into adaptive governance packages provides conceptual foundation for cross-functional coordination mechanisms, suggesting that effective institutional stewardship requires deliberate orchestration of technical standards, organizational procedures, and stakeholder engagement processes that together shape AI's organizational trajectory. Janssen's work underscores that AI governance is not primarily technical implementation challenge but organizational transformation requiring institutions to develop new capabilities for managing sociotechnical co-evolution. Papagiannidis, Emmanouil, Patrick Mikalef, and Kieran Conboy. 2025. “Responsible Artificial Intelligence Governance: A Review and Research Framework.” The Journal of Strategic Information Systems 34 (2): 101885. https://doi.org/10.1016/j.jsis.2024.101885. Papagiannidis and colleagues provide theoretical grounding for understanding how organizations operationalize responsible AI principles through governance structures. Their scoping review synthesizes disparate literature to construct a conceptual framework differentiating between structural practices (formal roles, committees, reporting lines), relational practices (stakeholder engagement, accountability mechanisms), and procedural practices (auditing protocols, monitoring systems) across the AI lifecycle. This tripartite framework is valuable for research institutions navigating the challenge of translating broad ethical commitments into actionable organizational arrangements. The article explicitly addresses how governance antecedents like leadership commitment and regulatory pressure shape implementation, and how governance practices in turn affect outcomes like trust, compliance, and innovation capacity. The authors' critical reflection on responsible AI governance reveals tensions inherent to institutional stewardship: between centralized oversight and distributed expertise, between standardized protocols and context-sensitive judgment, between rapid AI deployment and deliberative ethical review. For research organizations managing AI's entanglement with knowledge infrastructures, this work provides vocabulary and analytical categories for diagnosing governance gaps and designing institutional responses that embed responsibility throughout organizational systems rather than treating ethics as external constraint. Weber, Michael, Martin Engert, Norman Schaffer, Jörg Weking, and Helmut Krcmar. 2023. “Organizational Capabilities for AI Implementation—Coping with Inscrutability and Data Dependency in AI.” Information Systems Frontiers 25: 1549–1569. https://doi.org/10.1007/s10796-022-10297-y. Weber and colleagues identify four organizational capabilities essential for AI implementation, addressing AI's distinctive characteristics that distinguish it from previous information technologies: inscrutability (difficulty predicting probabilistic outputs and explaining decision processes) and data dependency (reliance on high-quality, continuously updated data for system performance). Their capability framework—encompassing AI Project Planning, Co-Development, Data Management, and AI Model Lifecycle Management—provides actionable guidance for research institutions developing internal capacity to steward AI systems. The article's grounding in expert interviews from diverse organizational contexts reveals that AI implementation failures often stem not from technical deficits but from organizational incapacity to manage sociotechnical complexity: inscrutability requires enhanced planning and stakeholder communication practices to align expectations around uncertain outcomes, while data dependency demands robust governance of data quality, provenance, and evolution throughout systems' operational lives. For research organizations, these capabilities map onto critical institutional functions: libraries developing metadata systems must manage data quality for AI cataloguing tools; research offices supporting computational scholarship must plan projects acknowledging unpredictable AI behaviour; IT departments maintaining infrastructure must implement lifecycle management ensuring model performance doesn't degrade as institutional data changes. The article's emphasis on Co-Development capability—bringing together technical specialists, domain experts, and end-users to jointly shape AI systems—speaks directly to cross-functional coordination challenges in research institutions where AI touches multiple organizational domains. Weber et al.'s framework reveals that organizational stewardship requires not just governance structures (committees, policies, review processes) but operational capabilities (planning methods, collaboration practices, technical procedures) embedded in day-to-day institutional work. This capability perspective shifts attention from abstract principles to concrete organizational competencies that determine whether responsible AI rhetoric translates into institutional practice. Wu, Chuhao, He Zhang, and John M. Carroll. 2024. “AI Governance in Higher Education: Case Studies of Guidance at Big Ten Universities.” Preprint, arXiv, September 3. https://doi.org/10.48550/arXiv.2409.02017. Wu and colleagues provide rare empirical documentation of how major research universities translate AI governance principles into institutional practice, examining guidance documents from fourteen Big Ten universities to identify organizational patterns and role-specific strategies. Their analysis reveals three key organizational dimensions: multi-unit governance involving information technology departments, teaching and learning centres, libraries, and research offices operating with distributed authority rather than centralized control; role-specific guidance differentiating expectations for faculty, students, staff, and researchers rather than applying uniform policies; and educational-advisory approaches emphasizing learning and adaptation over compliance enforcement. This case-study evidence illuminates the pragmatic challenges research institutions face when operationalizing AI oversight: how to coordinate across functional silos with different mandates, expertise, and risk tolerances; how to balance flexibility (needed because AI applications vary dramatically across contexts) with consistency (needed to ensure institutional values are upheld); how to position governance as enabling innovation rather than constraining it. The article's documentation of specific institutional mechanisms, such as data classification systems limiting what information can be shared with AI tools, or Socratic guidance approaches that pose questions rather than dictate answers, provides concrete examples of organizational stewardship in action. For understanding AI's organizational entanglement with knowledge infrastructures, this work demonstrates how governance emerges through negotiation among multiple institutional actors, each bringing domain expertise and jurisdictional claims, requiring coordination mechanisms that existing organizational charts may not accommodate. Publishing, Communication, Engagement Infrastructures Bergstrom, Tracy, and Dylan Ruediger. 2024. “A Third Transformation? Generative AI and Scholarly Publishing.” Ithaka S+R. October 30, 2024. https://doi.org/10.18665/sr.321519. Drawing on interviews with leaders from publishers, technology disruptors, academic libraries, and scholarship, Bergstrom and Ruediger map the strategic landscape of AI adoption across the publishing industry. The report identifies a bifurcated future: an incrementalist scenario where AI produces efficiency gains without fundamentally altering industry dynamics, versus a transformative scenario creating disruption comparable to or exceeding previous digital transformations. For search and discovery, most interviewees anticipate heavy AI impact, with tools already expanding capabilities through summarization and chatbot interfaces—potentially disrupting the linear progression from discovery to understanding by introducing AI-enabled synthesis. For peer review, interviewees expressed optimism that AI could address chronic strain through "pre-review" feedback, assistance with copy editing and misconduct detection, and more efficient reviewer identification, while raising concerns about accuracy, confidentiality, and maintaining human judgment in evaluation. The authors document a competitive threat from large technology companies: scenarios where commercial LLMs become default interfaces for accessing scholarly content pose significant challenges to publisher positioning. Content licensing to foundation model developers offers clear monetization paths, but broader revenue implications remain uncertain. The analysis identifies a critical research integrity challenge: ensuring transparent standards for AI usage while upholding provenance, attribution, reproducibility, and transparency in an environment of increasing automation. Smaller publishing organizations may struggle to match larger entities' adaptive capacity, potentially accelerating industry consolidation. Broussard, Meredith. 2023. “The Challenges of AI Preservation.” The American Historical Review 128 (3): 1378–81. https://doi.org/10.1093/ahr/rhad366. Broussard illustrates how the current impermanence of digital storage is producing a nightmare for future historians, as content is hidden away by licensing agreements or simply disappears when its host platform ceases to exist. Earlier ideals of the internet preserving a “complete first draft of history” are far from the reality, and problems only intensify when considering the development of AI, which Broussard compares to the printing press. For example, one could access a record of a physical newspaper published on a given day, but Google Search’s software is constantly changing with no canonical daily version, which means no historian could ever hope to examine a comparable snapshot of an AI. Operations like the Internet Archive and the development of emulation can help diminish losses, but it must be recognized that the digital world is decaying. Buitrago-Ciro, Jairo, Marta Samokishyn, Rachel Moylan, Jonathan Hernández Pérez, Oluwabunmi Bakare-Fatungase, and Carmel Firdawsi. 2025. “Bridging the AI Gap: Comparative Analysis of AI Integration, Education, and Outreach in Academic Libraries.” IFLA Journal 51 (3): 682–702. https://doi.org/10.1177/03400352251325274. Buitrago-Ciro et al. examine the websites of 40 libraries in North America, Western Europe, Latin America, and Africa to explore how they have integrated AI technologies into their services, policies, and outreach efforts. Overall, the authors found a geographic divide in the integration of AI into library services, because almost all North American and European libraries had AI educational outreach activities and resources, while only five (out of 10) Latin American libraries and one (out of 10) African libraries did outreach activities. However, AI integration into library services and development of library AI policies and guidelines was more uncommon across all regions, as less than half of North American and European libraries engaged in these activities, while only one library in Latin America and one in Africa integrated AI into their services and none had specific library AI policies. These regional differences in AI integration stem from each region’s socioeconomic contexts, social inequalities, and technological gaps that present additional challenges for AI adoption in African and Latin American libraries. Gulson, Kalervo N., and P. Taylor Webb. 2023. “Steering the Mind Share: Technology Companies, Policy and Artificial Intelligence Research in Universities.” Discourse: Studies in the Cultural Politics of Education 44 (2): 195–207. https://doi.org/10.1080/01596306.2021.1981828. Gulson and Webb examine how major technology companies are actively reshaping the landscape of artificial intelligence research within universities. Drawing on interviews with computer scientists as part of a broader international project, the authors investigate three key areas: the influence of tech companies on AI research practices, the policy frameworks that foster academia-industry collaborations, and the role of "open science" in facilitating the transfer of academic ideas to the corporate sector. The study highlights that the boundaries between academic and corporate research have become highly porous. Companies like Google do not merely act as external funders; they are deeply embedded within university ecosystems, actively steering research directions. Furthermore, the authors illustrate how national policies, such as Canada’s Pan-Canadian Artificial Intelligence Strategy, legitimize and accelerate this corporate influence by heavily incentivizing industry partnerships. While these collaborations provide universities with crucial resources and drive innovation, Gulson and Webb warn that they also raise significant concerns regarding conflicts of interest and the corporatization of open science. Ultimately, the authors argue that corporate funding creates powerful feedback loops among tech companies, policy-makers, and academics, effectively allowing industry to capture the "mind share" of university research. Kousha, Kayvan, and Mike Thelwall. 2024. “Artificial Intelligence to Support Publishing and Peer Review: A Summary and Review.” Learned Publishing 37 (1): 4–12. https://doi.org/10.1002/leap.1570. Kousha and Thelwall provide a systematic mapping of AI tool deployment across the scholarly publishing pipeline, distinguishing demonstrated capabilities from promotional claims. For journal recommendation, the review documents AI-powered systems, including Springer Nature Journal Suggester, Wiley Journal Finder, IEEE Publication Recommender, and JANE (Journal/Author Name Estimator), that analyze text similarity with previously published articles and reports high accuracy rates for appropriate journal matching. The analysis of initial quality control covers a diverse toolkit for plagiarism detection, robot author detection, methods checking, automated statistical verification, transparency and reproducibility checking and manuscript structure validation. Commercial systems draw on databases to suggest appropriate reviewers; the Natural Science Foundation of China's AI-assisted reviewer recommender for grant applications reports approximately 80% accuracy. However, the review identifies a critical boundary: while AI proves effective for finding reviewers and conducting initial quality checks, its value in performing the actual substantive review process "has not been clearly demonstrated." The synthesis reveals that substantial efficiency improvements are achievable in labour-intensive administrative tasks, while human judgment remains integral to core intellectual evaluations—a distinction essential for understanding where AI integration in publishing will proceed incrementally versus face fundamental obstacles. Ma, Lai. 2024. “Generative AI for Academic Publishing? Some Thoughts About Epistemic Diversity and the Pursuit of Truth.” KULA: Knowledge Creation, Dissemination, and Preservation Studies 7 (1): 1–5. https://doi.org/10.18357/kula.287. Ma (2024) critiques the integration of generative AI in academic publishing, arguing that the drive for automation threatens the central values of epistemic diversity and bibliodiversity. By invoking the "Sokal Hoax," the author illustrates the danger of AI-generated "bullshit"—content that mimics the formal structures of scholarship while remaining devoid of actual meaning. Ma specifically uses Scopus AI as a case study to demonstrate how the platformization of scholarly data creates a feedback loop that reinforces the "Matthew Effect," where established, English-language, and well-resourced publications are disproportionately amplified. This process, Ma warns, leads to "epistemic injustice" by marginalizing non-Western research and niche topics that are underrepresented in training datasets. The paper serves as a vital warning: without robust data surveillance and human-centric safeguards, the AI-accelerated publishing model may exacerbate peer-review crises and citation monopolies, ultimately undermining the public trust and diverse knowledge creation that open scholarship aims to protect. Mitchell, Peta, Michelle Riedlinger, Jake Goldenfein, Aaron Snoswell, Jean Burgess, and Kevin Witzenberger. 2025. “Research Genai: Situating Generative AI in the Scholarly Economy.” AoIR Selected Papers of Internet Research. https://doi.org/10.5210/spir.v2024i0.14006. Mitchell et al. focus upon the oft-overlooked subfield of Generative AI tools known as “RGAI”, research-focused models which they argue represent “complex sociotechnical systems.” Hepp et al.’s definition of AI as a “sensitizing concept” and Watermeyer’s consideration of AI as a “labor accelerator” are used to situate this argument within the “scholarly economy” before the authors expand upon these theories, arguing RGAI should be approached as a form of “platform capitalism” akin to academia.edu. To evidence this stance, the authors examine the highly distinct RGAI models “Consensus” and “Writefull” as case studies. “Consensus” is designed to produce an academic “ConsensusMeter” by combining a proprietary LLM with OpenAI technology, while “Writefull” “aims to simplify… the often challenging task of scholarly writing” by employing custom AI models. Mitchell et al. highlight that although both tools are presented as independent and bespoke, they have major corporate backing. This deeply conflicts with the research principles of open scholarship, so Mitchell et al. aim to develop an index of RGAI which are appropriate for use in research. Pividori, Milton, and Casey S. Greene. 2024. “A Publishing Infrastructure for Artificial Intelligence (AI)-Assisted Academic Authoring.” Journal of the American Medical Informatics Association 31 (9): 2103–13. https://doi.org/10.1093/jamia/ocae139. The Manubot AI Editor implements a concrete technical solution to the provenance problem in AI-assisted scholarship through three integrated components: a Python library, a GitHub Actions workflow, and a prompt generator. This architecture addresses provenance directly: all changes are tracked through version control, creating a clear audit trail distinguishing human-authored from machine-generated text and documenting exactly how AI suggestions were accepted, modified, or rejected. Evaluation proceeded through five case studies using both human and automated assessment. Human evaluators assessed whether revisions preserved original meaning and important details, avoided introducing incorrect information, and maintained correct formatting. Automated "LLM-as-a-Judge" iterative assessment evaluated paragraph pairs across criteria including sentence structure clarity, ease of understanding, and grammatical correctness. The evaluations found that models could grasp complex academic concepts and enhance text quality, with particular effectiveness in text-based sections like introductions and discussions. The human-in-the-loop design, where AI suggestions function as proposed edits rather than direct insertions, mitigates risks of incorrect information while preserving authorial agency over final content. Razack, Habeeb Ibrahim Abdul, Sam T. Mathew, Fathinul Fikri Ahmad Saad, and Saleh A. Alqahtani. 2021. “Artificial intelligence-assisted tools for redefining the communication landscape of the scholarly world.” Science Editing. https://doi.org/10.6087/kcse.244. Razack et al. argue AI will transform publishing into a technology-driven industry, streamlining dissemination of scholarship “for the betterment of humankind.” Exploring current trends via case studies, they emphasize AI’s impact upon human performance over its independent “creative” role. They argue AI could benefit the “prospective” researcher by predicting an article’s “citation impact,” identifying evolving research trends, and locating relevant research journals to publish in, cutting past low-charging and “predatory OA journals” which “malign integrity” by publishing “compromised content.” Meanwhile, “retrospective” editors could employ AI to detect plagiarism, streamline peer review, and format publications. For Razack et al., the rise of AI can therefore be compared to the rise of the typewriter, swift adoption likely promoting "human-machine collaboration" over replacement. Accountability-related Infrastructures Ngulube, Patrick, and Neema Florence Vincent Mosha. 2024. “Integrating Artificial Intelligence-Based Technologies ‘Safely’ in Academic Libraries: An Overview through a Scoping Review.” Technical Services Quarterly 42 (1): 46–67. https://doi.org/10.1080/07317131.2024.2432093. Ngulube and Mosha (2025) conduct a scoping review examining the state of research on ethical issues and perceived risks in AI integration within academic libraries. Analyzing 28 studies published before 2024, the authors reveal that research on safe AI adoption in libraries remains nascent, with significant gaps in theoretical grounding and methodological diversity. The review identifies critical ethical concerns including algorithmic bias, privacy violations, reduced transparency and accountability, and job displacement—yet finds that existing literature disproportionately emphasizes employment risks while neglecting environmental and planetary impacts. The authors note that ethical considerations have not taken centre stage in library and information science research, as evidenced by the scarcity of relevant keywords and abstracts in the literature. A major finding is that most studies employ quantitative methodologies at the expense of qualitative and mixed-methods approaches, and only 18% employ explicit theoretical frameworks. The review underscores that academic libraries must navigate the dual challenge of leveraging AI's operational efficiencies while ensuring ethical standards of inclusivity, accessibility, and fair information use—a task complicated by the absence of established best practices and the need for human-centred approaches to AI governance. “Realising Potential, Supporting Users: IFLA Statement on Copyright and Artificial Intelligence.” n.d. IFLA. Accessed May 1, 2025. https://www.ifla.org/news/realising-potential-supporting-users-ifla-statement-on-copyright-and-artificial-intelligence/. This statement aims to help IFLA member libraries in navigating copyright issues and developing programs relevant to AI, positioning libraries as innovators “uniquely situated to lead” in supporting, training, and utilizing AI. However, restrictions stem from rightsholders, economic and moral copyright laws, and health, safety, and privacy laws. Libraries should therefore be guided by the most appropriate policy sources to mitigate restrictions by ensuring AI tools do not compromise areas of concern. Libraries are further recommended to advise decision-makers to permit mining legally accessed content, address bias by promoting “the widest possible access,” build capacity, awareness, and training for evolving technologies, and monitor AI standards. Governments are recommended to follow existing human rights regulations, avoid using copyright law as a “blunt force tool,” develop ethical AI practices, and fund repositories of AI training data—alongside AI companies themselves. Both rightsholders and vendors are advised to avoid language in contracts which restricts the use of AI or prevents exceptions. Overall, this document aims to promote AI by making as much content available as possible. Werder, Karl, Balasubramaniam Ramesh, and Rongen (Sophia) Zhang. 2022. “Establishing Data Provenance for Responsible Artificial Intelligence Systems.” ACM Transactions on Management Information Systems 13 (4): 1–25. https://doi.org/10.1145/3503488. Werder et al. provide a framework for understanding audit trails in the context of AI and research data management, arguing that robust data provenance is not merely a technical feature but a prerequisite for achieving accountability in AI systems. They detail the requirements for a comprehensive provenance system, capable of tracking data from its origin through various transformations and into model training and deployment. This is directly relevant to research infrastructures, as it outlines how repositories and data management platforms can and must evolve to support AI-driven research. The authors’ multi-layered architectural proposal offers a concrete vision for how a research institution could implement a system to audit data pipelines, identify sources of bias, and ensure regulatory compliance. By connecting provenance to the broader goal of “Responsible AI,” the paper provides the conceptual tools for evaluating whether a research infrastructure’s data governance is sufficient to handle the accountability challenges posed by AI. Audience de Angelis, Luigi, Francesco Baglivo, Guglielmo Arzilli, Gaetano Pierpaolo Privitera, Paolo Ferragina, Alberto Eugenio Tozzi, and Caterina Rizzo. “ChatGPT and the Rise of Large Language Models: The New AI-Driven Infodemic Threat in Public Health.” Frontiers in Public Health 11 (2023): 1166120. https://doi.org/10.3389/fpubh.2023.1166120. De Angelis and colleagues examine the emergence of Large Language Models, particularly ChatGPT, as a novel public health threat through the lens of "AI-driven infodemic." The authors trace the rapid evolution of LLMs from earlier transformer-based models through GPT-3 to ChatGPT, highlighting how each iteration has increased the capacity to generate fluent, authoritative-sounding text at scale. Critically, they identify a fundamental misalignment problem: despite efforts to fine-tune these models through reinforcement learning from human feedback, LLMs remain prone to generating plausible-sounding but factually incorrect or fabricated information—a vulnerability particularly dangerous in medical and health contexts. The paper's central argument is that the unprecedented speed and volume at which LLMs can produce convincing content creates conditions for misinformation spread on a scale previously impossible, especially among non-expert publics who lack the epistemic resources to detect AI-generated falsehoods. The authors emphasize that the inability to reliably detect AI-produced text compounds this threat, undermining public trust in scientific institutions and potentially influencing health-related behaviours and policy decisions. The work demonstrates how LLM-mediated scientific communication poses distinct risks to knowledge integrity and public understanding, requiring urgent development of detection mechanisms, governance frameworks, and media literacy initiatives to protect audiences from AI-driven epistemic harm. Bryant, Rebecca. 2024. “Implementing an AI Reference Chatbot at the University of Calgary Library.” Hanging Together, December 12. https://hangingtogether.org/implementing-an-ai-reference-chatbot-at-the-university-of-calgary-library/. Bryant (2024) details the development and implementation of "T-Rex," an AI-powered reference chatbot at the University of Calgary Library, based on a webinar hosted by the OCLC Research Library Partnership. While the library had offered live chat services since the early 2010s, the COVID-19 pandemic caused a massive surge in demand, peaking at over 3,000 inquiries in a single month. To alleviate this strain, the library analyzed past chat transcripts and determined that 12–14% of inquiries were simple, directional questions suitable for automation. To prevent scope creep during the initial training phase, the development team focused on a core set of fifty common questions. Through continuous refinement, the chatbot has since expanded to encompass over 1,000 custom responses, though its effectiveness remains limited by the availability of existing library webpages or FAQs to draw from. Bryant concludes by summarizing the Calgary team's key insights for institutions developing similar tools: developers must anticipate out-of-scope or non-library questions, recognize that users prefer direct answers over links to webpages, program creative responses for off-topic queries, and prepare for user resistance or non-adoption. Hara, Noriko, Eugene Kim, Shohana Akter, and Kunihiro Miyazaki. 2025. “Exploring the Dynamics of Interaction About Generative Artificial Intelligence Between Experts and the Public on Social Media.” Journal of Science Communication 24 (1): A02. https://doi.org/10.22323/2.24010202. Hara and colleagues conduct an empirical investigation into how experts and the public co-produce knowledge about generative AI on social media, specifically analyzing X (formerly Twitter) discussions. Rather than treating the public as passive consumers of expert knowledge, the authors adopt a Public Engagement with Science (PES) framework that recognizes social media platforms as dynamic arenas where non-experts actively contribute to shaping collective understanding of emerging technologies. Through computational and manual analysis, they identify distinct discussion topics, map the roles that both experts and laypeople play in knowledge production, and examine how engagement metrics correlate with these roles. A critical finding is that the public functions beyond questioners seeking expert guidance, but rather as active knowledge co-producers who share practical insights, challenge claims, and contribute their own perspectives and experiences with GenAI tools. This research demonstrates how AI itself—as a subject of discourse—becomes a site of negotiated meaning-making between differently positioned actors. The work moves beyond traditional one-way dissemination models to reveal the interactive, participatory nature of contemporary science communication, with implications for understanding how diverse audiences collectively construct understanding of transformative technologies and how this co-productive process shapes both public perception and the future development of AI systems. Jaillant, Lise, Olivia Mitchell, Eric Ewoh-Opu, and Maribel Hidalgo Urbaneja. 2025. “How Can We Improve the Diversity of Archival Collections with AI? Opportunities, Risks, and Solutions.” AI & Society 40: 4457–4459. https://doi.org/10.1007/s00146-025-02222-z. Jaillant et al. draw from interviews with 20 academics and archival professionals to explore how AI technologies can help address the lack of diversity in archival collections. The authors highlight that AI tools can assist in automatically detecting racist or inappropriate language in metadata, search large amounts of historical records, and uncover new insights within archival records. However, the authors argue that the deployment of AI should involve close collaboration between librarians, archivists, and developers. The current limited collaboration between developers and libraries limits the ethical and inclusive application of AI in archival collections, especially those with sensitive historical materials. The interviewees also warned against the loss of historical context when using AI, the perpetuation of archival biases, and the loss of control when AI tools are designed and governed by tech companies. To conclude, the authors recommend investing in interdisciplinary AI training programs for archivists, educating AI developers about the unique needs of archives so they can create more relevant solutions, and creating professional guidelines about the best practices for AI use to address the lack of diversity in archival collections. Peters, Uwe, and Benjamin Chin-Yee. 2025. “Generalization Bias in Large Language Model Summarization of Scientific Research.” Royal Society Open Science 12 (4): 241776. https://doi.org/10.1098/rsos.241776. Peters and Chin-Yee provide rigorous, large-scale empirical evidence of a systematic and consequential distortion in LLM-generated scientific summaries: the tendency toward overgeneralization. Testing 10 prominent LLMs—including ChatGPT-4o, ChatGPT-4.5, DeepSeek, LLaMA 3.3 70B, and Claude 3.7 Sonnet—on 4,900 summaries of scientific abstracts and full-length articles, the authors demonstrate that even when explicitly prompted for accuracy, most LLMs consistently produce conclusions broader and more definitive than those warranted by the original research. Critically, they identify three specific mechanisms of overgeneralization: the use of generic statements that obscure quantification, the shift from past to present tense (which expands scope), and the omission of qualifiers and limitations. The findings are stark: LLM-generated summaries were nearly five times more likely than human-authored summaries to contain broad generalizations (odds ratio = 4.85, 95% CI [3.06, 7.70], p < 0.001), with newer models performing worse than earlier ones. This research is foundational for understanding how AI-mediated knowledge communication systematically distorts scientific findings at scale, regardless of user expertise or intent. The work demonstrates a concrete mechanism through which LLM intermediation can alter the epistemic content of research, with particular implications for medical and clinical contexts where overgeneralized conclusions can directly influence policy and patient care. The authors propose mitigation strategies including temperature adjustment and systematic benchmarking, but their findings underscore the profound challenge of ensuring faithful knowledge transmission through AI systems. da Silva Cardoso, Heike, and Vitor Rocio. 2025. “Enhancing Digital Libraries Through NLP and Recommender Systems: Current Trends and Future Prospects with Large Language Models.” In Technology and Innovation in Learning, Teaching and Education: TECH-EDU 2024, edited by Arsénio Reis, José P. Cravino, Leontios Hadjileontiadis, Paulo Martins, Sofia B. Dias, Sofia Hadjileontiadou, and Tassos Mikropoulos, 69–79. Communications in Computer and Information Science 2480. Cham: Springer. https://doi.org/10.1007/978-3-032-02672-9_5. da Silva Cardoso and Rocio propose a practical framework for integrating Natural Language Processing and Large Language Models into academic digital libraries to address the information overload facing researchers, students, and faculty. Recognizing that traditional search methods have become insufficient in the face of exponential publication growth, the authors advocate for AI-driven recommender systems capable of delivering precise, relevant, and personalized literature recommendations. Their contribution centres on an audience-aware design philosophy that positions librarians as essential partners in system development. Rather than treating AI as a purely algorithmic solution, the authors emphasize that librarians' direct interactions with users provide irreplaceable insight into the nuanced and context-dependent needs of diverse academic communities. This human-in-the-loop approach represents a noteworthy methodological stance: AI functions not as a replacement for human expertise but as a tool to augment librarian knowledge and enhance the responsiveness of knowledge discovery systems. The work is particularly salient for understanding how contemporary LLM-based systems can be designed to serve scholarly audiences more effectively while maintaining the institutional wisdom and user-centred perspective that librarians bring to information work. Taneja, Ankit Kumar, and Chandra Tripathi. 2020. “AI-Powered Recommender Systems: Personalization and Bias.” Turkish Journal of Computer and Mathematics Education (TURCOMAT) 11 (1): 1090–1094. https://doi.org/10.61841/turcomat.v11i1.14406. Though published before the generative AI era, Taneja and Tripathi's analysis of AI-powered recommender systems remains foundational to understanding algorithmic bias in information discovery. The authors examine the structural tensions inherent in personalization algorithms—systems designed to enhance user experience by tailoring content suggestions yet simultaneously capable of constructing "filter bubbles" and "echo chambers" around users. By continuously recommending content similar to previous engagement, these algorithms systematically exclude novel, dissenting, or interdisciplinary perspectives, undermining serendipitous discovery essential to genuine scholarly inquiry. Critically, the authors frame algorithmic bias not as a technical glitch but as a structural risk embedded within the information architectures that guide knowledge-seekers, and thus they expose how algorithmic personalization—a feature predating generative AI—already demonstrated the capacity to narrow intellectual diversity and reinforce pre-existing beliefs. This pre-GenAI work provides historical context for understanding how contemporary AI systems have intensified these risks, making it useful for evaluating the ethical imperatives of designing systems that promote intellectual openness rather than epistemic closure. de Winter, J. 2024. “Can ChatGPT Be Used to Predict Citation Counts, Readership, and Social Media Interaction? An Exploration Among 2222 Scientific Abstracts.” Scientometrics 129: 2469–2487. https://doi.org/10.1007/s11192-024-04939-y. De Winter’s study contrasts conventional scientometrics, testing whether a large language model can predict an article’s citation count and altmetric scores (whilst noting these may be unreliable assessors of actual scientific impact) by rating its abstract across “semantically diverse” variables. His case study uses ChatGPT-4 to predictively evaluate 2,222 abstracts from articles published in PLOS ONE, monitoring its accuracy by referencing actual citation counts and altmetrics scores. De Winter employs a custom script to prompt OpenAI’s Application Programming Interface (API) to score each abstract on a scale of zero to one hundred across thirty positive variables and their thirty antonyms, which he then organizes under five categories (e.g., “Quality and Reliability”). As Chat-GPT 4 only produced consistent results at the population level, it was run three additional times to determine averages and bolster reliability of individual scores. This language-based evaluation of article abstracts revealed that ChatGPT-4’s scores better correlate with altmetrics and citation counts than conventional readability scores do. Furthermore, “Novel and Engaging” articles were most likely to be cited, whilst “Accessible and Understandable” articles scored highest in altmetrics. AI and Social Complete Alphabetical List of References