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

Open Access Kember, Sarah, and Amy Brand. 2023. “The Corporate Capture of Open-Access Publishing.” The Chronicle of Higher Education, August 16. https://www.chronicle.com/article/the-corporate-capture-of-open-access-publishing. Kember and Brand highlight that open access policies adopted to increase availability of publicly funded research “may unintentionally contribute to greater consolidation in academic publishing” by encouraging commercial publishers to “value quantity over quality and platforms over people.” Kember and Brand use the history of “corporate capture” as a guide, referencing the development of self-archiving practices into a “neoliberal framework” of “openness at any cost.” Open-access reforms which targeted academic publishing as a “monolithic… bad actor” enabled this by hurting small actors more than major commercial entities, which exploited their vulnerability. Since then, the largest publishers have initiated hybrid subscription models to charge more overall, or “Read-and-Publish” deals that are increasingly expensive, decimate shrinking library budgets, and reinforce inequities in authorship. Kember and Brand argue this process was driven by ignorance, error, and the ideology of “immediate interest,” where short term gains are prioritized. The neoliberal idea that “information wants to be free” is similarly “subjugating all values to those of the market,” as open-access policy “inherits a utopia and risks delivering a dystopia” by making vast amounts of linked data readily accessible to third parties. The answer lies in a network of non-profit, fair-profit, and international organizations, which could turn the “false promise of openness into truly public knowledge.” Luth, Eric. 2025. “The Use of Wikipedia, Wikimedia, and Open Access Content for Artificial Intelligence and Text and Data Mining.” Stockholm IP Law Review 6 (1): 109–138. https://doi.org/10.53292/33313cc8.be33e111. Luth examines how Digital Commons stakeholders such as Wikimedia influence the development of AI models trained on Creative Commons (CC) licensed content. He argues that while these resources are essential for AI training, AI developers’ compliance with license conditions is often uncertain, and some AI developers may bypass legal requirements. Luth highlights two strategies: using CC licenses to protect shared knowledge and collaborating to sustain the ecosystem of open access. By combining these approaches, stakeholders can support high-quality AI outputs while preserving the integrity of the Digital Commons. The article emphasizes that AI’s reliance on shared resources carries ethical and practical responsibilities and shows how open licensing can enhance the reliability and social value of AI. Pooley, Jefferson. 2024. "Large Language Publishing: The Scholarly Publishing Oligopoly's Bet on AI." KULA: Knowledge Creation, Dissemination, and Preservation Studies 7 (1): 1–11. https://doi.org/10.18357/kula.291. Pooley begins with the Microsoft–OpenAI vs. New York Times clash to show how both journalism and scholarly publishing are positioning themselves to charge AI firms for using their content. He notes, however, that journalism and academic publishing differ fundamentally: news organizations pay their writers, while scholarly publishers rely on research produced for free by academics or funded by taxpayers. Pooley criticizes commercial publishers for framing AI companies as “free-riders” even though those publishers themselves profit from vast unpaid scholarly labour. To clarify, the author uses the notion of "surveillance capitalism” (using behavioural data to “feed predictive models” and sell it to customers, i.e., Facebook, Google, Silicon Valley) and states that unlike big tech companies in which “if you do not pay you are the product,” publishing companies like Elsevier use researchers both as product and paying customers in “Surveillance publishing.” This allows the “data cartels” to feed predictive models and turn academics' data into a product, which is quietly sold to tech companies without the academics’ knowledge in a process Pooley calls “academic fracking.” Even open-access work is often restricted or enclosed within proprietary platforms that enable further data extraction. Pooley urges academics to push back now and campaign, before these companies fully extend their surveillance-driven business models into AI and continue monetizing scholarly work and researcher behaviour without accountability. Smith, Regan. 2025. “Licensing of Text for Generative AI: Learnings from Non-AI Licensing Practices.” The Columbia Journal of Law & the Arts 48 (4): 457–512. https://doi.org/10.52214/jla.v48i4.13926. This article provides an overview of the challenges and opportunities of licensing textual content for generative AI, analyzing how existing open-access licenses (CC BY, CC BY-SA, CC0) perform in machine learning contexts. The author argues that most licenses were drafted without anticipating large-scale AI training or synthetic output generation, highlighting gaps in coverage while noting that marketplace solutions (through negotiated deals with publishers or voluntary collective management organizations—CMOs) are already emerging. The piece emphasizes that AI developers and rightsholders can collaboratively experiment with licensing arrangements, scaling compliance and business operations to mitigate risk, rather than relying on rigid, government-mandated frameworks. The author proposes that machine-readable licensing provisions, including TDM opt-out signals and provenance-tracking requirements, should become standard features of scholarly communication infrastructure, and engages with the question of whether OA constitutes a commons governed by norms of reciprocity or a reservoir open to private capture, examining case law around "fair use" defences for training data and the adequacy of attribution mechanisms when source texts are statistically blended beyond recognition. Open Data Bender, Emily M., and Batya Friedman. 2018. “Data Statements for Natural Language Processing: Toward Mitigating System Bias and Enabling Better Science.” Transactions of the Association for Computational Linguistics 6: 587–604. https://doi.org/10.1162/tacl_a_00041. The authors propose data statements as a design intervention to reduce exclusion and bias in language technologies, particularly in NLP, arguing that such documentation can prevent misrepresentation and support more ethical and accurate system development. They define a data statement as a “critical enabling infrastructure” that helps developers understand how results may generalize, how systems should be deployed, and what biases might be present. Their recommended schema includes a detailed, long-form account (covering curation methods, language variety, speaker and annotator demographics, context of speech or text, and recording or text quality) and a short form that should accompany any publication using the dataset. Using two case studies, they demonstrate how this schema works in practice and conclude that without a field-wide commitment to this practice, efforts to confront bias will be severely weakened, whereas widespread adoption would immediately clarify representational limits and, over time, foster more inclusive datasets and value-sensitive research. Boyd, Karen. 2021. “Datasheets for Datasets Help ML Engineers Notice and Understand Ethical Issues in Training Data.” Proceedings of the ACM on Human-Computer Interaction 5 (CSCW2): 1–27. https://doi.org/10.1145/3479582. Boyd explores how documentation frameworks such as “Datasheets for Datasets” function in practice for machine learning engineers, showing that open access to data alone does not ensure ethical or responsible AI development. Through think-aloud sessions with 23 practitioners, Boyd demonstrates that engineers who received a Datasheet identified ethical issues earlier and more frequently than those who did not, proving that documentation can actively shape how practitioners recognize, interpret, and respond to potential ethical issues in training data. Framing the analysis through the concept of ethical sensitivity (recognition, particularization, and judgment), Boyd argues that practices focused on structured interventions could increase ethical awareness, like datasheets that could serve as an infrastructure for ethical engagement, prompting engineers to reflect on dataset provenance, demographic representation, and the broader consequences of data reuse, factors often hidden or overlooked in open data environments. This study challenges the assumption that openness automatically leads to fairer AI systems. By offering empirical evidence of how documentation scaffolds ethical decision-making, the author makes a compelling case for embedding such practices throughout AI, especially as machine learning systems increasingly rely on openly available and ethically complicated data. Carroll, Stephanie Russo, Edit Herczog, Māui Hudson, Keith Russell, and Shelley Stall. 2021. “Operationalizing the CARE and FAIR Principles for Indigenous Data Futures.” Scientific Data 8: 108. https://doi.org/10.1038/s41597-021-00892-0. Carroll et al. discuss how the FAIR (Findable, Accessible, Interoperable, and Reusable) principles for scientific data and CARE (Collective Benefit, Authority to Control, Responsibility, and Ethics) principles for Indigenous Data Governance can be operationalized to allow Indigenous Peoples’ Rights to control their data, especially as big data and open data increase access to large datasets. The authors consider that the two sets of principles can be complimentary, both for scientific data and for data created by or about Indigenous Peoples. Scientific data can benefit from the CARE principles because many scientific datasets contain Indigenous Knowledge and these datasets can be oriented in ways that enhance the wellbeing of people by recording provenance and using appropriate governance models. Similarly, Indigenous Data can benefit from the FAIR principles to make data interoperable, regardless of the source or underlying technology, even if the datasets cannot be made openly accessible. The authors provide some recommendations to operationalize FAIR principles with CARE by making Indigenous data FAIR, raising awareness of the CARE principles among the research community before engaging with Indigenous data, and testing ways to use CARE principles for Indigenous data. Carroll, Stephanie Russo, Ibrahim Garba, Óscar Luis Figueroa Rodríguez, Jarita Holbrook, and Raymond Lovett. 2020. “The CARE Principles for Indigenous Data Governance.” Data Science Journal 19 (1): 43. https://doi.org/10.5334/dsj-2020-043. The authors ground the article in the structural biases embedded in data ecosystems and the ongoing processes of colonization that have produced epistemicide and constrained Indigenous Peoples’ ability to sustain their knowledges. They highlight how inequities operate across digital infrastructures and institutions, while Indigenous knowledge systems already engage with this complexity. The article directly refers to the tension between Indigenous Data Sovereignty and the data demands of AI, machine learning, and big data initiatives, noting that secondary data use and broad data sharing can threaten Indigenous rights, consent, and collective interests. Within this context, the Indigenous-led development of the CARE principles (Collective Benefit, Authority to Control, Responsibility, and Ethics) is discussed as a value-based response to concerns about the secondary use and reuse of Indigenous data. It highlights that open data for AI training is not a merely technical matter, but a political one, where broad data sharing and algorithmic extraction can underscore Indigenous rights, consent, and collective interests. By articulating CARE, authors argue for frameworks that centre people and purpose rather than solely data accessibility or system performance. The paper concludes that reclaiming control of data and data narratives shifts Indigenous Peoples from being subjects and invisible within data systems to self-determining users who shape governance, policy, ethics, and innovation in ways aligned with collective benefit and equity. Heger, Amy, Liz B. Marquis, Mihaela Vorvoreanu, Hanna Wallach, and Jennifer Wortman Vaughan. 2022. “Understanding Machine Learning Practitioners' Data Documentation Perceptions, Needs, Challenges, and Desiderata.” Proceedings of the ACM on Human-Computer Interaction 6 (CSCW2): 1–29. https://doi.org/10.1145/3555760. Heger et al. (2022) examine whether existing data documentation frameworks actually meet the needs of machine learning (ML) practitioners, a crucial issue given that AI models now shape high-stakes decisions across disciplines and depend heavily on the quality and transparency of their training data. The authors argue that while documentation initiatives like Datasheets for Datasets aim to support responsible and transparent AI, little is known about how these tools function in industry settings, where workflows, pressures, and data types differ substantially from academic contexts. Through semi-structured interviews and a documentation exercise with 14 practitioners, they find that current documentation practices are largely ad hoc, narrow in scope, and often disconnected from responsible AI concerns. Participants struggled to understand the purpose of many documentation questions, were unsure of appropriate granularity, and often failed to recognize the ethical implications of dataset characteristics. Their findings reveal a significant gap between responsible AI ideals and everyday ML practices. The study provides evidence that documentation frameworks must be contextual, integrated into existing tools, and supported by actionable guidance if they are to prevent downstream harms. By deriving seven design requirements, the authors offer concrete directions for developing documentation standards that can support safer, more accountable open AI data practices. Huerta, E. A., Ben Blaiszik, L. Catherine Brinson, Kristofer E. Bouchard, D. Díaz, et al. 2023. “FAIR for AI: An Interdisciplinary and International Community Building Perspective.” Scientific Data 10: 487. https://doi.org/10.1038/s41597-023-02298-6. Huerta et al. (2023) survey a wide landscape of US and European initiatives defining the FAIR (Findable, Accessible, Interoperable, Reusable) principles for AI models, arguing that responsible AI requires more than open access, it requires infrastructures that make AI assets understandable, comparable, and genuinely reusable. Their main argument is that FAIR must be reinterpreted for AI because models, software, and workflows are now “first-class” research objects, and only coordinated, community-wide standards can support interpretability, reproducibility, and sustainable open science. By mapping dozens of large, interdisciplinary FAIR-for-AI efforts, the authors show how FAIRness enables cross-disciplinary model reuse, transparent benchmarking, and educational access, ultimately laying groundwork for foundation models built from FAIR, AI-ready datasets. The paper highlights a major gap in current AI openness debates: open access alone does not ensure usable, trustworthy, or equitable AI. The authors demonstrate that without shared metadata standards, integrated repositories, and long-term infrastructure, openness can fragment into silos that impede scientific discovery. Their community-focused perspective invites readers to see FAIR not as a bureaucratic checklist but as an evolving, collaborative project essential for scaling transparent, reusable AI across scientific domains. Taeihagh, Araz. 2025. “Governance of Generative AI.” Policy and Society 44 (1): 1–14. https://doi.org/10.1093/polsoc/puaf001. Taeihagh’s work takes on the evolving governance challenges surrounding generative AI, particularly the way its vast training data demands have injected new urgency into long-running debates over open data governance. Unlike rule-based AI systems designed for fixed, pre-specified tasks, generative models train on expansive datasets and produce novel outputs. The author argues that this reshapes the very terrain of risk. This shift invites distinct concerns: hallucination, jailbreak, exposure of sensitive information, accuracy failures, opacity, and the general difficulty of control. Taeihagh not only details how these risks emerge from the technical properties of generative systems but also proposes a governance framework that maps potential interventions to each category of risk and explores the institutional supports required to make such interventions effective. The framework is especially instructive for today’s AI data-governance debates, where the scale and unpredictability of generative models are compelling policy-makers to rethink what openness should entail, whose interests it advances, and how governance mechanisms can guard against extractive data practices while steering AI development toward public values. Open Source Bostrom, Nick. 2017. “Strategic Implications of Openness in AI Development.” Global Policy 8 (2): 135–148. https://doi.org/10.1111/1758-5899.12403. This article is a cross-section that surveys the notion of openness in AI development before proprietary AI models entered the market, observing the early approaches of competing companies including Google, Facebook, and Microsoft, as well as OpenAI’s origins as a non-profit organization—OpenAI now operates as a capped-profit company governed by a non-profit, functioning effectively as a for-profit public benefit corporation rather than a publicly traded entity. The article focuses on long-term moves toward openness and their implications, but also briefly considers the benefits and problems inherent in short- and medium-term openness: Bostrom speculates that too much openness runs the risk of curbing corporate (and therefore funding) interest, ultimately slowing the rate of AI’s technological progress. However, he also asks: “Supposing openness would speed technical progress and rollout of AI capabilities, would that be socially beneficial?” (137). Bostrom notes a number of concerns ranging from public safety to labour encroachment, but still suggests that openness in AI development would be a net positive that accelerates the rate of technological progress in the short and middle-term. In the article’s long term section, Bostrom considers multiple scenarios in which openness might either help or hinder the safety and progress of a general AI superintelligence. He identifies distinct forms of openness (Science and Source Code; Control Methods and Risk Analysis; Capabilities and Expectations; and Values, Goals, and Governance Structures), addressing the potential value of openness in each, and argues that while openness is generally positive, especially in issues of values and safety, sharing too much knowledge may increase the odds someone will abuse this technology or ignore safety. Cooper, Martin. 2023. “Open Source is Good for AI But, Is AI Good for Open Source?” ITNOW 65 (2): 50–51. https://doi.org/10.1093/combul/bwad062. Cooper shares Terence Eden's perspectives on the evolving relationship between artificial intelligence and the open source movement, ultimately raising ethical and practical concerns about attribution and authorship in the age of AI-generated code. Eden believes that open source promotes technologies like AI, cloud, and the internet, driving efficiency, reducing duplication, and advancing scientific discovery. He raises an issue with AI’s use of open source without proper attribution violating one of open source’s core principles: crediting the original creator. Despite acknowledging that AI can assist developers and make programming more accessible, Eden cautions against viewing AI as a substitute for genuine understanding or creativity. Ultimately, Eden’s article is both a celebration and a warning: open source has empowered AI, but unless attribution and ethical use are safeguarded, AI may erode the very openness and integrity that made its success possible. Guo, Xinwei, Yujun Li, Yafeng Peng, and Xuetao Wei. 2024. “Copyleft for Alleviating AIGC Copyright Dilemma: What-if Analysis, Public Perception and Implications.” Preprint, arXiv, February 19. https://arxiv.org/abs/2402.12216. This article analyzes and summarizes copyright disputes related to Artificial Intelligence Generated Content (AIGC) and isolates three primary stakeholders within these disputes: data owners, who create the content upon which AIGCs have been trained; model owners, who develop the AIGC applications and models (e.g., ChatGPT, Stable Diffusion, etc.) themselves; and model users, who apply prompts to AIGC models to generate content. The authors consider the benefits, challenges, and risks that arise from each given stakeholder making a claim to copyright while considering the potential benefits and challenges of embracing copyleft (i.e., granting free use over material so long as the products of that free use are also under copyleft). In their contemplation of each stakeholder, they recognize the benefits for each group in advocating copyright to content, but also identify harm to other stakeholders, society, progress in developing AIGC, and institutions (e.g., “the copyright registration system”). The authors consider three real-life cases surrounding copyright disputes and envision how copyleft would effectively alleviate the disputes and tension surrounding their different stakeholders. The last half of the article performs a survey that asks users who have a range of familiarity with AIGC a range of questions, including who ought to have copyright claims to a generated image of the Mona Lisa in the style of Vincent Van Gogh. The authors note that their survey indicates consistent ambivalence, misunderstanding, and concern over copyright issues related to AIGC. The survey identified that there is support for applying copyleft to AIGC, and that that support increases the more someone understands the issues surrounding AIGC and copyright. Osborne, Cailean. 2024. “Why Companies "Democratise" Artificial Intelligence: The Case of Open Source Software Donations.” Preprint, arXiv, September 26. https://doi.org/10.48550/ARXIV.2409.17876. Osborne aims to advance the "nascent research agenda on the political economy of open source AI" by studying various corporate AI OSS donations to non-profit foundations and thereby sourcing and explaining corporate incentives behind their "democratization" of AI. He firstly outlines this topic in detail, clarifying that the technology required to access AI is expensive, and developing AI even more so, meaning this "democratization" represents only a lowering rather than removal of barriers. Osborne then considers the possible incentives behind organizational level decisions to democratize AI, as companies seek to competitively dominate industry norms, test and improve their products, enlarge intellectual property, reduce R&D costs, and focus external research into their own AI. Significantly, individual developers of open source software do not have the same incentives but tend to accept commercial participation. On one hand, corporate participation means volunteers will be contributing toward these sources of corporate profit without reward; on the other hand, it risks companies dominating open research spaces and developing a complex intercompany research network of "co-opetition" which individual volunteers cannot enter. Therefore, Osborne’s study is highly relevant to issues of commercialization and open scholarship and especially concerns of “open-washing.” The study’s results highlight the predominantly commercial incentives behind corporate “AI democratisation” efforts, appropriately mirroring its most common form—the democratization of governance. However, Osborne's results also illustrate the significance of the individual level, “bottom-up” movement, suggesting that a significant 38% of decisions to donate stemmed from developers. Despite this, he concludes that “AI democratization” is indeed “a strategic calculation, where companies relinquish control of their OSS project in exchange for assumed benefits,” as “democratisation of governance is treated as a social means for primarily economic and technological ends,” broadly supporting the works of Widder et al. and Srnicek, and encouraging other researchers to bring their attention to this area. Radu, Roxana. 2021. “Steering the Governance of Artificial Intelligence: National Strategies in Perspective.” Policy and Society 40 (2): 178–193. https://doi.org/10.1080/14494035.2021.1929728. This article provides a survey and analysis of the national strategies related to AI implemented by a dozen countries around the world. It is of particular interest to those interested in concrete policies and regulations surrounding AI rather than discourse and rhetoric. Radu identifies a tension between nations claiming to take a leadership role and corporations that are “driving developments in the field.” More specifically, she observes that the vast “majority of binding agreements and voluntary commitments” come from corporations rather than nations or government regulatory bodies, indicating a concerning trend toward self-regulation. Governments have often played a key role in facilitating the conditions in which the development of AI technologies have thrived, even funding the rudimentary research that leads to AI models and applications, and this article notes that, regardless of the level of government regulation, the state is rarely a passive player when it comes to AI. Radu identifies the trend of a “dominant place for industry in the drafting process” and that representatives from industry make up a third of members drafting national strategies as well as prominent positions within those groups. Radu also notes absences: for instance, despite more than seventy countries using AI for surveillance, this issue is conspicuously absent from most strategies. Radu concludes that the common thread among many of these strategies is a hybridity between industry and state that affirms “the deliberate choice for functional indetermination,” resulting in a lack of clarity surrounding “who makes the rules and for how long,” and notes there are very few limitations placed on industry as these strategies instead provide vague ethical guidelines. Schmit, Cason, M. J. Doerr, and J. K. Wagner. 2023. “Leveraging IP for AI Governance.” Science 379 (6633): 646–647. https://doi.org/10.1126/science.add2202. This article proposes a hybrid AI copyright model that combines copyleft licensing and “patent trolling” in order to enforce an ethical approach to using AI models and training datasets. Schmit et al. contend that current legislation that applies to AI models and their use “are insufficient to impel ethical conduct” and require further consideration. The article advocates its hybrid model—Copyleft AI with Trusted Enforcement (CAITE)—as both rigorous and flexible, allowing for both self-reporting and monitoring. Essentially, the system works by creating an ecosystem in which AI models and training data are developed using copyleft, allowing for free use of these materials so long as the product of that use is covered by the same licensing agreement. At the same time, the enforcement of that licensing agreement is transferred to a regulatory entity called the CAITE Host, which acts as a positive counterpart to the patent troll, ensuring that no individual or organization breaches the terms of copyleft, introduces dangerous applications, data, or features, or unethically uses materials under the copyleft agreement. The CAITE Host and its code of ethics are driven by community oversight and buy-in, and the host is regulated by the community. The authors contend that its flexibility and bottom-up structure make CAITE uniquely suited to handling the rapidly developing field of AI and its shifting needs. Once CAITE achieves a critical mass, it would also become increasingly unlikely for AI developers to opt out. The article concludes by proposing a pilot model for CAITE, and strongly suggests that such a framework is far preferable for industry and others than stringent, relatively slow-moving, and inflexible government regulation in isolation. Shanklin, Grant, Cason Schmit, Jacob Pierce, Nicole Martinez-Martin, and Matthew DeCamp. 2025. “The Case for Contextual Copyleft: Licensing Open Source Training Data and Generative AI.” Preprint, arXiv, July 17. https://arxiv.org/abs/2507.12713. This article introduces and describes a new form of content licensing, CCAI (Contextual Copyleft AI), designed to apply copyleft values and ethics to AI content, and should appeal to anyone interested in coding, copyright, and the (mis)use of training data for generative AI. The authors advocate for an approach that preserves the values of the Free and Open Source Software (FOSS) movement by using copyleft as a mechanism for the rights of AI content. The CCAI license requires that the free use and distribution of the software in question, that any work derived from that software must also be licensed under CCAI, and that CCAI software must “include source code or provide a way to access it.” The CCAI draws from and imitates other copyleft licenses but explicitly addresses and outlines the expectations regarding specific components of AI models within its license, applying copyleft and a requirement of openness to training data, code, and parameters. The article contends that whether it is easily demonstrable or not, clear guidelines of when training on publicly available data is and is not okay can act as a deterrent to misuse, and the CCAI license will give developers more agency in how others use their code, accelerate open source AI development by giving open source AI models open source training data only they can use, and limit the practice of open washing—claims that software and AI is open source when it is not (or not fully open source). The authors admit there are security risks that come with open AI models (e.g., vulnerabilities to malicious actors), but they contend that the benefits far outweigh those risks, especially when accompanied by the implementation of the CCAI and complementary regulations. Widder, David Gray, Dawn Nafus, Laura Dabbish, and James Herbsleb. 2022. “Limits and Possibilities for ‘Ethical AI’ in Open Source: A Study of Deepfakes.” In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’22), 2035–46. New York: Association for Computing Machinery. https://doi.org/10.1145/3531146.3533779. Widder et al. assert that significant understudied harms arise from divergent practices of “transparency and accountability” in the open source community. They argue that interventions designed for private companies may be improper for open source contexts, highlighting gaps in current approaches to “Ethical AI.” Their purpose is to reveal these systemic vulnerabilities and propose frameworks that better address the unique social and technical dynamics of open source development. They use an open source Deepfake creation tool as an extreme case to highlight how ethical reasoning intersects with broader cultural frameworks and interview project leaders. In this case study, researchers offer a definition of “Down stream control assumptions,” suggesting that “Ethical AI” research and “design interventions” would be improved by explicitly acknowledging these assumptions and identifying ways to work effectively under weak assumptions of downstream control in open sources contexts. Open Science Ación, Laura, Mariela Rajngewerc, Gregory Randall, and Lorena Etcheverry. 2023. “Generative AI Poses Ethical Challenges for Open Science.” Nature Human Behaviour 7 (12): 1800–1801. https://doi.org/10.1038/s41562-023-01740-4. This article explores the tension between Open Science’s need to make knowledge freely accessible and the rise in GenAI and its tendencies to use any information available to it without necessarily considering how its application of such information (including information made available through Open Science) may do harm. These “frictions,” as the authors describe them, point to Open Science’s need to make data available while ensuring said data is not used to do harm. It is a concise summary of the foundational issues at the heart of these conversations, and great for anyone interested in how GenAI complicates Open Knowledge. The authors propose selective licensing as one approach to curbing harm, but notes that this only protects against harm from direct use (e.g., “open-source code used to build a weapon”) and not an indirect use (e.g., “open-source library use within another algorithm that is in itself harmful”). Ultimately, they stress the need for new governance systems and regulations related to the creation of knowledge that prioritizes both the public good and right to research. Baltes, Sebastian, Florian Angermeir, Chetan Arora, Preetha Chatterjee, Jens Deitenbeck, Daniel Graziotin, Christoph Treude, et al. 2025. “Guidelines for Empirical Studies in Software Engineering Involving Large Language Models.” Preprint, arXiv, August 21. https://arxiv.org/abs/2508.15503. This article develops best practices and guidelines for the use of LLMs (expressly language LLMs, not image, audio, or video) for empirical studies in software engineering (expressly data collection, processing, and analysis). The authors aim to “enable reproducibility and replicability” in the field even when LLMs are in use and “produce more reliable evidence on what LLM can and cannot deliver.” They present what they call a “taxonomy of Study Types” in addition to their guidelines, providing examples both within and beyond software engineering, all aimed at promoting the use of LLMs while promoting open science practices. The article considers LLMs as annotators, judges, synthesizers, and subjects in SE, and the guidelines Baltes et al. present have two tiers, MUST and SHOULD, demonstrating a nuanced understanding of practices without which SE research employing LLMs is rendered ineffective (e.g., declaring an LLM’s usage and role) and those practices that merely make this sort of research more rigorous and accessible (e.g., including full interaction logs if privacy and confidentiality can be ensured). The authors advocate for these practices to better enable researchers in SE to engage with LLMs while practicing Open Science. Of particular interest is the way in which the authors structure their understanding of the field and guidelines, and how they recognize distinct studies require slightly different practices. Biderman, Stella, Hailey Schoelkopf, Quentin Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, et al. 2022. “EleutherAI: Going Beyond ‘Open Science’ to ‘Science in the Open.’” Preprint, arXiv, October 12. https://arxiv.org/abs/2210.06413. This article outlines the inception and progress of EleutherAI, a lab and AI model first founded in 2020 by a group of (for the most part) university affiliated academics and researchers concerned that for-profit companies are the primary developers of GenAI research and models, and that there is a lack of transparency and openness related to these technologies. Those interested in alternative approaches to AI models or a concrete example of Open Science in action should read this article. Because these researchers want to make available a “capable large language model… for public use, study, and scrutiny,” they have conducted all EleutherAI research in the open, making it freely accessible to everyone (1): all research and discussions take place on Discord servers, and projects coordinated and teams assembled in those servers on a grassroots basis. This “public-facing research” leads to a “virtuous cycle” with little turnaround between the conception of ideas and development of projects, and facilitates conversation, collaboration, and community by giving all users, even newcomers, unfettered access to every means of communication in which the group engages. The authors consider the downsides to such an approach as well, citing the issues of relying on volunteer labour, how others might take this public research for their own without proper accreditation to EleutherAI, and the important consideration that perhaps not all things should be made public, but ultimately conclude that such open research is integral to the academic community. Gundersen, Odd Erik, Odd Cappelen, Martin Mølnå, and Nicklas Grimstad Nilsen. 2024. “The Unreasonable Effectiveness of Open Science in AI: A Replication Study.” Preprint, arXiv, December 20. https://doi.org/10.48550/ARXIV.2412.17859. Gundersen et al. examine the extent to which a widely accepted crisis in research reproducibility affects AI research. They do this by conducting a systematic replication study of 30 highly cited AI studies, allowing the authors to identify the problems with irreproducibility. Out of the gate, the replication process could only be attempted with 22 of these studies because their research data needed to be openly available and reproducible without special hardware. Furthermore, the 22 replications were each given 40 working hours of effort, but 12 of these took too long and had to be stopped early, impacting their quality. In result, it was found that out of the 22 replication attempts began only six could be accurately reproduced and five partially reproduced, meaning a 50% success rate. The main problems for reproducibility were identified in six sources: the "source code," "article," "data," "results," and "resources." Notably, each of the six accurate reproductions had code available, and the one other study with available code was a partial success limited by the time constraint, meaning availability of code alongside the data certainly improved the quality of the replication and was the most important factor for success. Therefore, despite limitations of a small sample size and time constraints, Gundersen et al. found that AI research’s reproducibility climbs from 33% accuracy when only data is available to 86% when both code and data are available. This is a significant enough result to empirically demonstrate the benefits of mandating openness and transparency of code as well as data.  Helregel, N. 2025. “AI and Open Science: Implications and Library Practice Recommendations.” Library Trends 73 (3). https://doi.org/10.1353/lib.2025.a961191. Helregel (2025) examines the complex intersection of artificial intelligence and the open science movement, questioning whether current AI technologies possess the necessary transparency to align with open research principles. The article explores multiple dimensions of this relationship, assessing both the potential for AI to actively advance open science goals and the broader impacts of AI adoption on the attitudes and behaviours of researchers and the public. Recognizing the critical role of library and information science professionals in navigating this shifting landscape, Helregel outlines concrete recommendations for institutional library practice. These actionable strategies include prioritizing internal knowledge-building, fostering cross-institutional advocacy, adapting consultation and liaison workflows, navigating complex licensing agreements, and enhancing science communication. Ultimately, the paper positions libraries as vital infrastructure for ensuring that AI integration supports, rather than undermines, the core tenets of transparency, reproducibility, and accessibility in scholarly communication. Hosseini, Mohammad, Serge P. J. M. Horbach, Kristi Holmes, Tony Ross-Hellauer. 2025 “Open Science at the generative AI turn: An exploratory analysis of challenges and opportunities.” Quantitative Science Studies 6: 22–45. https://doi.org/10.1162/qss_a_00337. Hosseini et al. provide an overview of the potential benefits and downsides of incorporating generative AI into different open science practices, as defined by the UNESCO Recommendation on Open Science, particularly open access, open data, open source software, open evaluation, open science infrastructures, open engagement of societal actors, and open dialogue with other knowledge systems. Some common benefits across these practices are that generative AI has the potential to lower the barriers to participation in science, summarize academic literature for non-academic audiences, and simplify processes in coding and data management. However, the authors also warn about the potential risks, including the creation and spread of misinformation, the propagation of hegemonic worldviews, and the monetization of research information. Given these downsides and the high social and ecological costs of generative AI, the authors suggest researchers question if AI tools are the most efficient way of achieving openness in science. The final section of the paper includes further recommendations for institutions, funders, and publishers, such as providing training on the use of generative AI, monitoring its positive and negative impacts, and continuing testing its use for evaluation and assessment purposes. Linåker, Johan, Cailean Osborne, Jennifer Ding, Daniel Puttick, and Richard Nweibo. 2025. “A Cartography of Open Collaboration in Open Source AI: Mapping Practices, Motivations, and Governance in 14 Open Large Language Model Projects.” Preprint, arXiv, September 29. https://arxiv.org/abs/2509.25397. This article engages in a survey and interviews of the practices, motivations, and governance of fourteen open LLMs with the aim of further enabling collaboration throughout the open LLM lifecycle. The essay provides an overview of terms for open AI and major takeaways and challenges for different types of collaboration—types of collaboration include model, data, software, evaluation, non-technical, and compute—at the pre-training, post-training, and post-release stages of the LLM lifecycle. The authors also identify five governance frameworks for open LLM projects ranging from single companies to “non-profit-sponsored grassroots projects.” The authors opine that collaboration does not merely happen with LLMs themselves but the infrastructure and accompanying tools (e.g., datasets, open source frameworks, discussion fora, etc.) needed to produce and maintain them, and more research and attention needs to be given to the ecosystem surrounding LLMs. The article also includes recommendations “to support the global community of practitioners building a more open future for AI” for researchers and developers, AI companies, policy-makers, academic institutions, platform providers, and open source foundations. For instance, they recommend that policy-makers fund infrastructure of public AI, not just specific models, use targeted initiatives to support regional languages, require publicly funded research to use licensing that aligns with open science principles, and insist on a clear definition of what constitutes openness in these cases (36). Papi, Sara, Marco Gaido, Luisa Bentivogli, and Matteo Negri. 2025. “FAMA: The First Large-Scale Open-Science Speech Foundation Model for English and Italian.” Preprint, arXiv, May 28. https://arxiv.org/abs/2505.22759. This article presents an Open Science Speech Foundation Model (SFM) called FAMA, “the first family of large-scale open-science SFMs for English and Italian trained on over 150k hours of exclusively OS-compliant speech data.” FAMA’s results are comparable to its closed counterparts and considerably faster (up to eight times). The authors present a brief summary of recent advancements in automatic speech recognition and speech translation while raising concerns about the closed nature of related applications’ training data and code and pointing to open source alternatives in different domains for inspiration. The majority of the article is dedicated to specifics of architecture, training, and results of FAMA compared to its competitors; however, the end result—that FAMA is competitive with and even outperforms its larger, non-open counterparts like OpenAI’s Whisper in some cases—is an excellent case study in how open science can work in the domain of LLMs and AI. Essential Contexts AI and Social