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

Platforms Burton, Jason W., Ezequiel Lopez-Lopez, Shahar Hechtlinger, et al. 2024. “How Large Language Models Can Reshape Collective Intelligence.” Nature Human Behaviour 8 (9): 1643–55. https://doi.org/10.1038/s41562-024-01959-9. Large language models are quickly becoming infrastructure for how groups seek information, deliberate, and coordinate, potentially altering the conditions under which collective intelligence emerges. Burton and colleagues argue that LLMs do not merely speed up individual work but transform how information is aggregated, accessed, and transmitted across digital environments that underpin collective performance in organizations and societies. Drawing on interdisciplinary perspectives, the paper frames this shift as simultaneously enabling and destabilizing: LLMs can support distributed cognition (e.g., idea generation, synthesis, translation, scalable facilitation) while also increasing risks tied to dependence on shared intermediaries, degraded diversity of viewpoints, and new pathways for error, manipulation, or misplaced confidence. Rather than offering a single model or experiment, the article maps a research-and-practice agenda by identifying potential benefits, key hazards, and policy-relevant considerations, then articulating open questions that link technical design choices to group-level epistemic outcomes. Its central claim is that understanding collective intelligence in the LLM era requires moving analysis beyond isolated human–AI interactions toward system-level effects on networks, platforms, and institutions, and that these effects warrant focused study before LLM-mediated coordination becomes the default for tackling complex problems. Corsi, Giulio, Bill Marino, and Willow Wong. 2024. “The spread of synthetic media on X.” Harvard Kennedy School (HKS) Misinformation Review. https://doi.org/10.37016/mr-2020-140. Corsi and Wong’s longitudinal empirical study analyzes 566 tweets containing synthetic media (December 2022–September 2023) documenting a dramatic surge following Midjourney V5 release with over 1.5 billion total views, revealing differential impacts where political deepfakes achieve higher median views despite representing minority of synthetic content. The authors demonstrate that generative AI's impact is stratified by content purpose with political misinformation posing acute democratic risks, while current detection and labelling systems lag behind generative capabilities, creating governance gaps compromising users' capacity to evaluate information. Using Community Notes data from December 2022 to October 2023, the study highlights the connection between the growth in generative AI tools and the rapid proliferation of synthetic media. The findings indicate that most synthetic media is non-political and largely benign, frequently taking the form of humorous or satirical images. However, the study also identifies a smaller but concerning portion of malicious synthetic media, which have potential height risks despite their lower frequency. The authors further examine the role of X’s paid verification system, revealing that while verified users tend to receive more overall views, this advantage diminishes when engagement is adjusted for follower count, implying that verification alone provides limited amplification. The study casts light on synthetic media’s expanding presence on social platforms and warns that even widespread exposure to seemingly harmless AI-generated content may gradually undermine trust in online information. To address these challenges, the authors recommend ongoing empirical monitoring, stronger collaboration between researchers and industry to develop transparent detection tools, and proactive policy measures aimed at reducing the social and political harms posed by malicious synthetic media. Delmonaco, Daniel, Samuel Mayworm, Hibby Thach, Josh Guberman, Aurelia Augusta, and Oliver L. Haimson. 2024. “‘What are you doing, TikTok?’: How Marginalized Social Media Users Perceive, Theorize, and ‘Prove’ Shadowbanning.” https://oliverhaimson.com/PDFs/DelmonacoShadowbanning.pdf. Delmonaco et al.’s qualitative study explores marginalized social media users’ experiences with shadowbanning through the theoretical framework of algorithmic “folk theories.” Through a review of existing literature, the authors demonstrate how social media companies publicly distance themselves from shadowbanning while continuing to apply it in practice and explain why content produced by marginalized users is more frequently subject to filtering. The study analyzes 24 interviews to explore how users perceive, theorize, and attempt to “prove” shadowbanning through indicators such as decreased visibility, engagement, and follower loss. The findings reveal widespread confusion surrounding shadowbanning, deep mistrust in platform governance, and the disproportionate harm imposed on marginalized communities whose content is suppressed despite platforms’ public denials. The data further shows how users collectively produce and co-produce theories to resist the opacity of algorithmic moderation, and authors introduce “collaborative algorithm investigation,” in which users test and share algorithmic folk theories with one another. The authors argue that algorithmic content moderation exacerbates existing inequities by obscuring governance decisions, politicizing moderation, and shifting power from human judgment to opaque predictive systems, while suggesting that increased transparency, tailored moderation approaches, and users’ education could help reduce unfair content removals. Drolsbach, Chiara, and Nicolas Pröllochs. 2025. “Characterizing AI-Generated Misinformation on Social Media.” Preprint, arXiv, May 15. https://doi.org/10.48550/arXiv.2505.10266. Drolsbach and Pröllochs’ article address a critical gap in existing research by shifting attention from the social drawbacks of AI-generated misinformation to its real-world prevalence and behaviour on social media platforms. Using a large-scale dataset of 91,452 misleading posts identified through X’s Community Notes platform, the authors empirically compare AI-generated and non-AI-generated misinformation across content characteristics, account attributes, virality, believability, and harmfulness. Their findings show that AI-generated misinformation is more entertainment-oriented, exhibits more positive sentiment, and is more likely to originate from smaller accounts, yet achieves significantly higher virality despite being slightly less believable and harmful than traditional misinformation. In their discussion and conclusion, the authors emphasize that AI-generated misinformation poses a growing challenge to the digital information ecosystem due to its realism, scalability, and difficulty of detection. They argue that its unique features disturb algorithmic amplification mechanisms and require new platform strategies, policy interventions, and research approaches that account for the unique properties of AI-generated misinformation rather than relying on countermeasures designed for conventional false content. Duricic, Tomislav, Hussain Hussain, Emanuel Lacic, Dominik Kowald, Denis Helic, and Elisabeth Lex. 2023. “Beyond-Accuracy: A Review on Diversity, Serendipity, and Fairness in Recommender Systems Based on Graph Neural Networks.” Frontiers in Big Data 6: 1251072. https://doi.org/10.3389/fdata.2023.1251072. Duricic et al. (2023) provide a comprehensive review of recommender systems based on Graph Neural Networks (GNNs), with a particular focus on objectives that go beyond prediction accuracy, including diversity, serendipity, novelty, and fairness. While GNNs have demonstrated strong performance in modelling complex user–item interactions and improving recommendation accuracy, the authors argue that this very strength can intensify structural biases such as popularity bias, homophily, and feedback loops that marginalize less popular content and users. The paper provides a definition of each beyond-accuracy metric, reviews how it has been used in literature, and analyzes the technical barriers of integrating these goals into GNN-based systems such as measurement difficulties, trade-offs between competing goals, and risks of overfitting due to model complexity. Diversity-aware neighbourhood sampling, contrastive learning, and adversarial training to disrupt feedback loops are introduced for more heterogeneous and unexpected recommendations. In conclusion, the authors position GNNs as a powerful but double-edged tool in recommendation systems, capable of either narrowing or broadening users’ informational horizons depending on design choices. They argue that future research must focus on principled frameworks that integrate “beyond-accuracy” objectives, emphasizing that recommendation algorithms are not neutral optimizers but key sociotechnical systems shaping visibility, voice, and fairness in digital information ecosystems. Gorwa, Robert, Reuben Binns, and Christian Katzenbach. 2020. “Algorithmic Content Moderation: Technical and Political Challenges in the Automation of Platform Governance.” Big Data & Society 7 (1): 1–15. https://doi.org/10.1177/2053951719897945. Gorwa et al. unfold what "Algorithmic moderation systems” are perceived as essential in AI governance to address the public demand on AI accountability and transparency that remain unnavigable and inexplicable, such that even the enhanced versions could aggravate the existing misinformation issues. They state that major platforms' growth transforms them so much that they no longer function primarily as social networks. They define “algorithmic moderation” as “systems that classify user generated content based on either matching or prediction, leading to a decision and governance outcome (e.g. removal, geoblocking, account takedown).” Different types of hashing (matching) and classification (prediction) tools and the areas they are deployed (copyright, terrorism, and toxic speech) are described and their practicality are criticized. Their case analyzes including copyright enforcement, counterterrorism, and toxic speech illustrate the limitations and unintended harms of automated moderation, demonstrating how even “improved” systems risk reinforcing misinformation patterns, entrenching platform power, and displacing human judgment. Ultimately, Gorwa et al. argue that algorithmic moderation transforms platform governance itself by concentrating authority in opaque technical systems while rendering the underlying political choices invisible. Reynolds, C. J., and Blake Hallinan. 2024. “User-Generated Accountability: Public Participation in Algorithmic Governance on YouTube.” New Media & Society 26 (9): 5107–5129. https://doi.org/10.1177/14614448241251791. This study focuses on how YouTube creators respond to platform governance decisions by producing “user-generated accountability” content, videos that publicly document governance failures and pressure YouTube to intervene. Through an analysis of 250 such videos, the authors state that creators overwhelmingly experience YouTube’s policy enforcement, automated moderation systems, and communication practices as opaque, inconsistent, and often discriminatory. Authors declare that human review is essential in automizing platform governance. The study highlights that platform governance decisions affect creators differently depending on channel size, content type, and identity; larger creators can more easily mobilize audiences to “make noise,” while smaller creators rely on collective visibility tactics such as signal boosting across platforms, especially Twitter. The authors argue that these bottom-up accountability practices reflect both the limits of YouTube’s existing governance infrastructure and creators’ efforts to bypass scale dependent unresponsiveness by publicly sharing grievances, comparing experiences, and testing the platform’s algorithms themselves. Theoretically, the study provides the “user-generated accountability” as a framework for engaging with the unavoidable conflicts that will arise around algorithmic systems, and views content creators as “political actors” who advance the perceptions of algorithms. Empirically, it surfaces common concerns such as demonetization of LGBTQ content, and lack of human review. Savolainen, Lotta. 2022. “The Shadow Banning Controversy: Perceived Governance and Algorithmic Folklore.” Media, Culture & Society 44 (6): 1091–1109. https://doi.org/10.1177/01634437221077174. Savolainen studies shadowbanning as a lens for understanding current algorithmic platform governance from the user perspective. Rather than determining whether shadowbanning objectively exists, the author conceptualizes it as “algorithmic folklore”: informal, collectively produced narratives through which users attempt to make sense of opaque moderation and ranking systems. Drawing on Reddit discussions about TikTok, Instagram, and YouTube, the study shows how users interpret suppressed visibility through anecdotal evidence, experimentation, and shared pattern recognition, revealing the experiential dimensions of algorithmic governance. Either users “probe the rules” or find their ways through “pleasing the algorithms”; the author argues that these practices emerge in response to governance systems that lack the clarity, stability, and consistency that are traditionally associated with good governance. The study concludes that platform governance contains a structural paradox: platforms increasingly present themselves as legitimate governors while simultaneously developing automated moderation systems that undermine transparency and consistent enforcement. Ultimately, the ethical issue is not shadowbanning itself, but the broader reality of shadowy governance exercised by complex, distributed algorithmic systems at scale. Vetter, Matthew A., Jialei Jiang, and Zachary J. McDowell. 2025. “An Endangered Species: How LLMs Threaten Wikipedia’s Sustainability.” AI & Society, February. https://doi.org/10.1007/s00146-025-02199-9. Vetter et al. caution against using Wikipedia as AI training data uncritically by outlining many complicating factors. Wikipedia’s collaborative model of data curation democratizes data creation, but embeds power dynamics and biases. A feminist posthumanist framework is applied, which situates Wikipedia’s knowledge as contextual, based upon the positionality of predominantly Western, male editors, a perspective LLMs would perpetuate without opportunity for change. Wikipedia’s "dynamism" is the main reason it is so valuable, stemming from constant, collaborative updates by volunteers. AI responses’ inconsistent citations and plagiaristic tendencies undermine human editors, the source of Wikipedia’s verifiability, and exploit scholars, Wikipedia itself, and the past labour of volunteers. Wikipedia needs both "a steady stream of new and returning readers" and "a diverse group of dedicated volunteers," whilst AI responses push sources into obscurity and at best detour the website, threatening Wikipedia’s long-term sustainability. Vetter et al. practically explore Wikipedia’s relationship with LLMs by conducting a "problem-centred expert interview" of community leaders, who possess 25 combined years of involvement and research. Through these interviews, Vetter et al. identify three main areas of opportunity: promoting transparency and explainability, diversifying the Wikipedia community, and equipping users with critical thinking skills. Overall, their findings "resonate with the posthumanist approach to critical AI literacy," cautioning against hoping for purely technological fixes and calling for greater transparency in how tech giants "leverage open-access datasets." Governance, Leadership, and Policy Balendra, Soorya. 2025. “Meta's AI Moderation and Free Speech: Ongoing Challenges in the Global South.” Cambridge Forum on AI: Law and Governance 1. https://doi.org/10.1017/cfl.2025.5. This article conducts a case study of Meta’s AI-based content moderation in the Global South, arguing that the censorship of unprohibited content (“over removal”) and slow response to content that causes harm (“slow removal”) are part of “content moderation procedures” that constitute “massive discriminatory approaches.” Balendra presents three models of authority for the regulation of social media content: external regulation, self-regulation, and co-regulation, and argues that state regulation (falling under the external model) varies wildly depending on the government in question, with “granting absolute regulatory power to state actors often stifl[ing] political discourse and online expression.” Co- and self-regulation mitigate this problem, but introduce a new issue by shifting responsibility—private companies use AI to make millions of judgements about content in a short frame of time. The paper notes that the EU and the United States have drastically different policies regarding content moderation on social media platforms and argues that “the practice of content moderation remains largely shaped by the cultural norms of the United States” and the values of “a relatively homogenous groups of Silicon Valley elites.” Balendra finds that, at this time, current approaches often prioritize the state or private corporate interests rather than free speech or the interests of the larger community, and much work is needed to “achiev[e] genuine change” in the form of integrated, hybrid approaches. Birkstedt, Teemu, Matti Minkkinen, Anushree Tandon, and Matti Mäntymäki. 2023. “AI Governance: Themes, Knowledge Gaps and Future Agendas.” Internet Research 33 (7): 133–67. https://doi.org/10.1108/INTR-01-2022-0042. Birkstedt et al. conduct a systematic literature review (SLR) of the emerging yet fragmented field of “AI governance” (AIG), defined as “a system of rules, practices, processes, and technological tools” that ensure organizational AI meets strategic, legal, and ethical requirements. The authors use human-centric and sociotechnical AI traditions to explore how organization-level governance mechanisms can bridge the principles-to-practices gap in AI ethics. Through SLR, authors identify four research agendas. The “technical agenda” focuses on the practicalities of AIG. The “stakeholder and contextual agenda” ensures AI meets requirements via direct collaboration with external stakeholders, increased corporate social responsibility, and establishment of an “intelligent discourse regime.” The “regulatory agenda” aligns AIG with ethical and legal standards, while the technical, stakeholder and contextual, and regulatory agendas largely focus on the aims of organizational AIG (i.e., what), and the process agenda deals with the means to achieve the aims (i.e., how). The authors develop a generic framework to guide cohesive AIG practices and advocate for a collaborative governance approach that emphasizes transparency, stakeholder engagement, and sociopolitical awareness. Broekhuizen, Thijs, Henri Dekker, Pedro De Faria, Sebastian Firk, Dinh Khoi Nguyen, and Wolfgang Sofka. 2023. “AI for Managing Open Innovation: Opportunities, Challenges, and a Research Agenda.” Journal of Business Research 167 (November 2023): 114196. https://doi.org/10.1016/j.jbusres.2023.114196. Broekhuizen et al. systematically analyze how AI could be applied to the complex, unstructured task of “open innovation,” a possible business opportunity they believe has been insufficiently explored. They define “Open innovation” as “the practice of leveraging external ideas, resources, and capabilities to improve innovation outcomes,” and propose AI could minimize its inherent managerial challenges. To prove this, the authors form a framework that considers the three abilities involved in AI’s agreed possible applications to management (involving “mapping,” “coordinating,” and “controlling”) and assess how applicable their constituent functions are to the challenges of the three stages of open innovation (involving “initiation,” “development,” and “realization”). In the stage of “initiation,” they propose AI could scout for distant knowledge partners, detect relevant information gaps and opportunities, and forecast conflicts. During the “development” stage, AI could develop complementarity between partners' knowledge stocks, integrate knowledge whilst safeguarding intellectual property, and monitor partners for violations or conflicts within the relationship. Finally, during the “realization” stage the authors argue AI could comprehensively evaluate business opportunities, optimize resource deployment, and diagnose or remedy intellectual property violations. To Broekhuizen et al., this study “underscores the significance of the human-side of AI” by demonstrating AI’s valuable role in assisting management. They also conclude that AI’s ability to search for corporate partners will be its greatest impact in open innovation, as they argue it will both incentivize openness and reputability of companies and increase the need for them to keep sensitive data secure. Cheong, Inyoung. 2024. “Collaborative Approaches to AI Governance: Exploring Co-Design and Co-Regulation Models.” PhD diss., University of Washington. https://homes.cs.washington.edu/~yoshi/papers/theses/inyoung-cheong-dissertation.pdf. This dissertation seeks to integrate co-design and co-regulation as a means of facilitating domain-specific expert knowledge into AI governance policies. It uses theoretical analysis and empirical data as part of its methodologies, and it should be of interest to anyone who wants detailed case studies of the use of AI technology in specific domains, as well as clear models for (co-)governance and development in the use of Artificial Intelligence. Cheong performs two case studies, examining the development and use of an AI application to provide legal advice as well as the application of AI in online content moderation of news and comics apps in South Korea. These result in a set of strategies for AI co-regulation that includes parameters for starting conditions, institutional design, collaborative process, and facilitative leadership. Cheong proposes that the guiding principles for AI co-governance must account for specific and distinct contexts, maintain human involvement and intervention in what is often seen as an artificial and automated process, and look to resolve legal ambiguities surrounding AI, especially as there is a clear history of courts dismissing what Cheong calls “collective rule-making” by the community. The author argues for the implementation of co-regulation and co-design, and maintains that “by fostering inclusive dialogues, leveraging diverse expertise, and remaining adaptable to changing technological and societal landscapes” these methods show potential for dealing with AI’s greatest challenges. Floridi, Luciano. 2018. “Soft Ethics and the Governance of the Digital.” Philosophy & Technology 31 (1): 1–8. https://doi.org/10.1007/s13347-018-0303-9. Floridi argues our “mature information societies” means “we no longer live online or offline, but onlife,” making the sociopolitical challenge of ethically governing this foundational “digital age” even more vital than pursuit of future digital innovations. Floridi defines three overlapping subfields within the “Governance of the Digital,” including “Digital Ethics,” “Digital Regulation,” and the encompassing “Digital Governance.” The latter is defined as “the practice of establishing and implementing policies, procedures, and standards for the proper development, use and management of the infosphere,” therefore “sometimes neither moral nor immoral… legal nor illegal.” This demonstrates that the three subfields don’t always overlap, but each must always be considered by policy-makers, because what is legal, what is good, and what is best are all key components of the ideal “Governance of the Digital.” “Digital Ethics” should also be further subdivided into “hard” ethics, involving binary moral judgments about what should versus should not be done, and “soft” ethics, which involves a “post-compliance,” “post-feasibility” approach, ensures moral opportunities are maximized, and fosters “good corporate citizenship.” Differing “information societies” occupy differing levels of maturity, producing distinct ethical contexts which should progress with differing ethical focuses. In this regard, Floridi advocates “ethics foresight analysis,” which uses data analytics to assess ethical impacts of future innovations then targets research by systematically identifying what is feasible, then sustainable, then acceptable, then preferable. The author concludes that Digital Ethics must be “leading” not “chasing” developments, and not just functioning as a questioning exercise, but also signalling that ethical issues matter, engaging with stakeholders, and providing shareable solutions. Gasser, Urs, and Virgilio A. F. Almeida. 2017. “A Layered Model for AI Governance.” IEEE Internet Computing 21 (6): 58–62. https://doi.org/10.1109/MIC.2017.4180835. This article identifies the gap in knowledge created by the “black box” of AI—the obscured nature of the processes that take place from an initial prompt to its generated output. Gasser et al. note that users and policy-makers of these technologies have relatively little insight into how it actually works compared to its developers. Given this unique issue, Gasser et al. propose a theoretical framework for how to approach AI governance, drawing upon the advent and subsequent globalization of the Internet to inform their understanding of AI technology and its pattern of growth. The article addresses a core issue that strong or general AI is often discussed in terms of potential societal impact while weak or general AI is currently being deployed right now (i.e., 2017) and is in real need of governance already. The authors identify core themes related to “transparency, accountability, and explainability; inclusion and fairness; [and] global governance,” as well as three challenges for AI governance—"information asymmetries”; “finding normative consensus”; and “government mismatches”. The authors point out that the ethical and legal considerations of AI are closely related and take a modularity approach; they present three layers for governance that exist from AI systems to society, with timing (i.e., near-, mid-, and long-term) and considerations for each layer. The article concludes by proposing AI governance requires a blended approach based on the different risks of certain AI applications and propose that both “market-oriented solutions” and "government-based structures” can work on a national or international level, suggesting a “global oversight body” could act as “the curator of global principles and emerging norms for AI systems.” Kirschenbaum, Matthew. 2025. “The US of AI.” Public Draft, February 25, 2025. https://drive.google.com/file/d/1O2qkjhg7Ei5zZWmBraNwXq4V0lTauspN/view. This set of notes from Kirschenbaum represents his “near real-time attempt to come to terms with… the opening weeks of the second Trump Administration,” especially the “AI-first” strategy of “the so-called Department of Government Efficiency (DOGE).” Kirschenbaum outlines DOGE has nothing to do with government efficiency, as it only saves insignificant costs, instead possessing ideological and capitalistic motivations to form “a government without people.” Based on Musk’s nebulous authority—as an advisor from the private sector—DOGE targeted nineteen federal agencies to be “strip-mined and fracked for all manner of information” as a “priceless and unprecedented resource” of AI training data. To do so, DOGE operatives accessed sensitive spaces after hours, combining physical aggression with a “complete lack” of technical inhibition representing a “vampiric data suck.” Kirschenbaum identifies clear corporate interests here and further examines ideological aims. Early executive orders asserted the need for federal AI systems “free from ideological bias or engineered social agendas,” with OpenAI announcing “ChatGPT Gov” days later. These invocations of “AI” as a universal solution represent only a “linguistic token” instead of any realized technology—as worded by Salvaggio, AI must only “be considered a plausible competitor to human decision-making long enough to dislodge the existing human decision-makers.” Kirschenbaum frames this as based in the “masculine maximalism” of the Trump administration, as the term AI promotes “technological absolutism” whilst theoretically playing into the Nixon-era “theory of the unitary executive” by aiming to replace people with “void” and consolidate power into the hands of “a super-user class who will operate outside the system.” However, “the real world intrudes… with lethal downstream effects.” Kirschenbaum proposes “a collapsing superimposition of political discourse, the actual operations of social media as its primary public arena, and the technical architecture of the systems (in the form of AI)… in the sphere of governance as well as media,” which produces the “infrastructure for… permanent despotism” via the “untethering of language from conditions of lived reality.” Oh, Dayei, and John Downey. 2024. “Does Algorithmic Content Moderation Promote Democratic Discourse? Radical Democratic Critique of Toxic Language AI.” Information, Communication & Society 28 (7): 1157–76. https://doi.org/10.1080/1369118X.2024.2346531. Oh and Downey critically examine how algorithmic content moderation shapes public discourse, focusing specifically on Google’s Perspective API and its detection of "toxic" language. By drawing a crucial distinction between incivility (harsh or emotional tone) and intolerance (actual discriminatory harm), the authors demonstrate how AI-driven moderation systems frequently misinterpret context. Consequently, these algorithms often suppress emotionally charged but politically significant speech—particularly from marginalized groups—while allowing structurally discriminatory but politely phrased content to evade detection. This dynamic reveals a significant flaw in relying on AI as a neutral mediator of digital dialogue: by prioritizing tone over substance, algorithmic moderation actively curtails democratic participation and skews the visibility of diverse perspectives. Their study is relevant for understanding how AI systems act as invisible gatekeepers, enforcing normative boundaries on public expression and highlighting the urgent need for moderation frameworks that are contextually aware, transparent, and equitable rather than merely tone-policing. Rughiniș, Cosima, et al. 2025. “AI at the Knowledge Gates: Institutional Policies and Hybrid Configurations in Universities and Publishers.” Frontiers in Computer Science 7: 1608276. https://doi.org/10.3389/fcomp.2025.1608276. This article constitutes a qualitative analysis (i.e., a form of computational analysis) of 16 universities and 12 publishers that attempts to understand how these institutions regulate AI in knowledge production. Rughinis et al. claim that most institutions preserve their values while regulating AI by adapting the regulations they already have in place, stating “institutional policies are shaped by value-based compromises and not by technical concerns alone.” At the same time, although the article acknowledges its focus on institutional regulation, it also admits that “AI adoption can evolve both through policy frameworks and… locally embedded practices that remain partly invisible in formal documentation.” The authors propose two concepts, “dual black-boxing” and “legitimacy-dependent hybrid actors.” Dual black-boxing refers to the lack of transparency in both the “hidden algorithmic processes [and] invisible training data” that shape AI systems and “the opacity surrounding how academics employ AI in their work,” while legitimacy-dependent hybrid actors are “configurations of human-AI collaboration whose institutional legitimacy is not intrinsic but conditional.” Rughinis et al. argue that transparency is a core academic value, and scholars must work to open these black boxes; they also suggest that that transparency is more closely linked to academic legitimacy in AI-related issues rather than the more traditional issue of the origin of ideas. The majority of the article applies these terms to understanding the regulatory frameworks of universities and publishers, and concludes with the understanding that the majority of institutions allow some form of AI use within certain parameters. The key types of hybrid actors allowed by universities were AI-assisted researchers, AI-enabled translators, AI-enhanced writers, and AI-supported coders, while publishers facilitated the use of AI to refine language, support research methodologies, and aid in research. Ultimately, the authors observe that institutions are permitting AI-supported work in refined roles and argue that AI-based academic work gains legitimacy through transparency. Seger, et al. “Democratising AI: Multiple Meanings, Goals, and Methods.” Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society. https://dl.acm.org/doi/10.1145/3600211.3604693. Aiming to “provide a foundation for more productive conversations,” Seger et al. categorize four forms of “AI democratisation” and explore each form’s goals and facilitation pathways. Democratization of AI “use” involves making it easier for people to employ AI’s capabilities, and is therefore equally comparable to increasing availability of printers or of tools which can create chemical weaponry, depending on the case. Democratization of “development” involves allowing a wider range of people to contribute to the AI design and development process, which accelerates progress and increases diversity, but also opens the door for maliciousness and irresponsibility. Democratizing “profits” constitutes a philanthropic or redistributive process, and has the predictably involved weaknesses. Therefore, the authors highlight that “AI democratisation is a multifarious and sometimes conflicting concept” which “should not be conflated with improving AI accessibility” and “is not inherently good.” This is because democratization of AI’s “use,” “development,” and “profits” only derives value from its alignment with the interests and values of impacted groups, making democratization of AI’s “governance” the priority. Democratization of “governance” does not necessarily aim for total agreement amongst constituents, but will regardless reduce the unilaterality of decision-making by promoting justice, legitimacy, and diversity. Sposato, Martin. 2025. ”Artificial Intelligence in Educational Leadership: A Comprehensive Taxonomy and Future Directions.” International Journal of Educational Technology in Higher Education 22 (20). https://doi.org/10.1186/s41239-025-00517-1. This article performs an analysis and review of published works on AI in educational leadership from 2017–2024 and uses the results of that process to develop what Sposato describes as “a comprehensive taxonomy of AI applications in educational leadership.” Potential readers include researchers and policy-makers interested in communicating across specific fields and disciplines. Sposato theorizes ten application domains for AI in higher education ranging from increased efficiency in administration and teaching practices to increased community engagement and strategic planning. The purpose of such a framework is to provide educational leaders with a coherent vocabulary of “the full spectrum of educational leadership responsibilities” that can facilitate further conversation. The author advocates for AI’s transformative potential in higher education even as they note important challenges (especially with ethics and equity), arguing a need for a balanced approach and “robust ethical guidelines and regulatory frameworks.” The framework provided is presented as one step toward developing AI literacy for stakeholders so that all involved in education might make more informed decisions and articulate their needs more clearly. Ter-Minassian, Lucile. 2025. ”Democratizing AI Governance: Balancing Expertise and Public Participation.” Preprint, arXiv, January 16. https://doi.org/10.48550/arXiv.2502.08651. Ter-Minassian argues that “universal access, technically-constrained development, and universal impact” of AI and AI research necessitate public participation in AI governance even as it “creates a unique governance challenge.” The article presents the benefits and challenges of expert oversight and public consultation. Ter-Minassian outlines the threat of misinformation to meaningful public participation, highlighting how both undue fear and unrealistic optimism surrounding AI can affect public discourse, while acknowledging that the rapid rate of AI research makes it difficult to incorporate truly democratic processes into the decision-making process. The author concludes that any ongoing inclusive and effective AI governance ought to incorporate inclusive representation, balance expertise and public engagement, embrace transparency and accountability while engaging an iterative process, and implement a hybrid framework that “blend[s] expert knowledge with meaningful public input” (4). She argues that “AI’s deep societal impact calls for public engagement” and “technical complexity does not have to exclude meaningful public input,” advocating for “structured deliberation and phased public participation” to address “time-sensitive AI challenges without compromising democratic legitimacy” (5). Ultimately, the article sees this as an iterative process that will need to continue to adapt in order to maintain expert insight and public discourse on AI. Critical Literacies Aguiar, Micaela, and Sílvia Araújo. 2024. “Final Thoughts: Digital Humanities Looking at Generative AI.” In Digital Humanities Looking at the World, edited by S. Araújo et al., 367–380. Cham: Springer Nature Switzerland. Aguiar and Araújo explore the emergence and applications of generative artificial intelligence within Digital Humanities research. They trace the evolution of generative AI from early implementations like the 1960s Eliza chatbot through its developments such as Generative Adversarial Networks (GANs) and the Transformer architecture to contemporary large language models like BERT and GPT. The authors document current applications of generative AI across Digital Humanities domains, including cultural heritage preservation, historical text processing, and literary creation. They outline how these technologies are being used to memorialize mass atrocities, process traditional Chinese ancient texts, and explore collaborative human-AI poetry creation. Aguiar and Araújo examine the inner workings of generative AI, particularly focusing on how conversational models function through statistical prediction rather than genuine comprehension. They identify limitations such as hallucinations, bias replication, and attribution problems that researchers must consider when using these tools. The chapter concludes that effective integration of generative AI in Digital Humanities requires developing AI literacy skills that enable researchers to critically evaluate outputs while leveraging the technology's capabilities for sustainable research advancement. Boer, Victor de, and Lise Stork. 2024. “Hybrid Intelligence for Digital Humanities.” Preprint, arXiv, April 15. https://doi.org/10.48550/ARXIV.2406.15374. De Boer and Stork believe the discipline of Digital Humanities (DH) should meet the challenges and opportunities posed by Artificial Intelligence (AI) by adopting the “human-centric paradigm” of Hybrid Intelligence (HI). They define DH as “a scholarly realm where computing or digital technologies intersect with… the humanities… characterized by innovative approaches to scholarly endeavors,” and identify various existing DH approaches to AI, which include “Knowledge Representation,” “Machine Learning,” “Natural Language Processing,” and “Computer Vision.” De Boer and Stork argue successfully integrating AI tools into DH requires embedding AI in scholarly practice, adopting a diverse and polyvocal approach, considering the debate of distant versus close reading, and applying a critical stance toward data, methods, and tools. HI “concerns itself with the investigation and design of human-AI ecosystems,” and the authors argue effective AI should be developed by following its “CARE” principles—being “Collaborative, Adaptive, Responsible, and Explainable.” Therefore, De Boer and Stork focus this text on mapping CARE principles to those DH requirements earlier outlined. Overall, they concur with Goodlad and Baker that humanists “are ideal 'domain experts' for the current juncture,” and end by identifying two overarching challenges. The first challenge is “one-shot solutions,” including tools, datasets, and methods that aren’t reusable, Open Source, accessible, and well described. Secondly, they identify the human factor—actors will need both an open mind and AI literacy to be equally critical of and cooperative with AI. Therefore, the authors emphasize the importance of teaching digital methods within the humanities curricula. Carmi, Elinor, Simeon J. Yates, Eleanor Lockley, and Alicja Pawluczuk. 2020. “Data Citizenship: Rethinking Data Literacy in the Age of Disinformation, Misinformation, and Malinformation.” Internet Policy Review 9 (2). https://doi.org/10.14763/2020.2.1481. This study argues that dominant understandings of digital and data literacy are insufficient for addressing contemporary complex challenges of disinformation, misinformation, and malinformation. They propose the concept of data citizenship to reframe literacy as a fundamentally civic, political, and collective practice rather than a set of individual technical skills. The authors contend that information disorders cannot be addressed through technological solutions alone, as they are deeply embedded in structural inequalities, platform economies, and power asymmetries shaping datafied societies. The article builds on a literature review and secondary data analysis drawing on Me and My Big Data project, alongside the development of a nationally representative UK survey. The authors identify three major gaps in existing literacy frameworks: an overemphasis on individual competence rather than networked practices, a lack of critical engagement with algorithmic and economic infrastructures of platforms, and the absence of proactive civic skills such as contesting data extraction, demanding transparency, and advocating for data justice. Their analysis highlights how citizens’ data practices are shaped by social networks, education, and socio-economic conditions, revealing unequal capacities for verification, protection, and collective action. The authors conclude that data literacy must move beyond content verification to include critical understanding of platform design, funding models, and algorithmic governance. They emphasize the importance of locally grounded, participatory approaches to data literacy education that reflect diverse lived contexts rather than scalable, top-down interventions. While the study offers a robust conceptual model and survey-based insights, it does not yet include in-depth qualitative engagement with citizen groups, which the authors position as a necessary next phase of research. Overall, the article makes a significant contribution by bridging digital literacy and the field of mis/dis/mal information research and claims that data literacy is an essential component of democratic citizenship. Long, Duri, and Brian Magerko. 2020. “What is AI Literacy? Competencies and Design Considerations.” In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, 1–13. New York, NY: Association for Computing Machinery. https://doi.org/10.1145/3313831.3376727. The authors present an exploratory review of interdisciplinary literature, aiming to organize key ideas on AI literacy definitions and essential competencies. Authors claim that public misunderstandings of AI are due to design, educational gaps, black-box algorithms, and a general lack of technological knowledge. The review includes two sections: how the public understands specific AI systems, and how existing research addresses people’s broader understanding of AI. Rather than offering an exhaustive synthesis, Long and Magerko’s focus is on non-technical audiences and they offer central provocations and design principles that can inform educational interventions. The proposed conceptual framework on AI literacy identifies 17 core competencies alongside 14 design considerations for educational intervention, which are described and supported by relevant literature. Together, these considerations position the framework as a practical and reflective tool for designing AI literacy interventions that move beyond dominant media narratives and foster critical, informed engagement with AI technologies. Majdik, Zoltan P., and S. Scott Graham. 2024. “Rhetoric of/with AI: An Introduction.” Rhetoric Society Quarterly 54 (3): 222–31. https://doi.org/10.1080/02773945.2024.2343264. Majdik and Graham propose a field-organizing distinction that functions as a methodological argument. By splitting inquiry into “rhetoric of AI” (AI as object) and “rhetoric with AI” (AI as method), they treat AI simultaneously as a rhetorical actor/ecosystem component that structures attention, circulation, and uptake, and as an instrument that can be enrolled into rhetorical research workflows. The authors claim that much prior rhetorical work engaged “spaces of AI”—platforms, devices, and algorithmically mediated environments—while leaving “AI” itself undertheorized as the central object; this provides a justification for why rhetoricians should now foreground model/algorithmic operations (and their harms) rather than treating them as neutral background conditions of digital rhetoric. They also frame generative text systems as a stress test for core rhetorical concepts—agency, authorship, and audience—because these systems destabilize the presumed linkage between human intention and textual production, and because they insert nonhuman intermediaries into communicative situations that education and public discourse still normatively treat as human-to-human. The introduction then uses this distinction to stage the special issue’s contributions as complementary routes for rebuilding rhetorical theory and pedagogy under AI conditions. Ng, Davy Tsz Kit, Jac Ka Lok Leung, Samuel Kai Wah Chu, and Maggie Shen Qiao. 2021. “Conceptualizing AI Literacy: An Exploratory Review.” Computers and Education: Artificial Intelligence 2: 100041. https://doi.org/10.1016/j.caeai.2021.100041. This article presents an exploratory review of literature published between 2016 and 2021 with the goal of conceptualizing AI literacy and proposing a comprehensive framework for teaching and evaluating it. The authors analyze 30 articles and synthesize existing definitions and approaches to AI literacy, noting that while AI literacy has increasingly been discussed, the concept has often been treated narrowly as a technical skill set. The authors situate AI literacy as a necessary competency for everyone, arguing that AI will increasingly affect many aspects of daily life and employment. In response, the authors argue for a broader understanding of AI literacy that incorporates cognitive and ethical dimensions. The proposed framework organizes AI literacy into four interconnected facets: knowing and understanding AI, using and applying AI, evaluating and creating AI, and AI ethics. Importantly, the authors explicitly reject purely instrumentalist views of AI literacy. To conceptualize AI literacy pedagogically, the authors draw on Bloom’s classic taxonomy, aligning the above-mentioned aspect with increasing cognitive levels. They claim that while most people are aware that AI exists, they often lack understanding of how it works or how to assess its ethical and social implications. This approach positions AI literacy as a developmental learning process rather than a static set of skills. The article concludes by identifying directions for future research, including the need for empirical studies to validate AI literacy frameworks and to better understand how AI literacy can be taught, learned, and assessed across educational contexts. Rapanta, Chrysi, Anna Åkerfeldt, Mark Vanderbeeken, Diane Lison, Khadija Mohammed, Amanda Gibbs, Helder Coelho, Pinar Seda Celik, Carola Bruna, Ingrid Helleve, Chrysoula Vassilakopoulou, Pieter Swart, Ana Lúcia Marques, and Dirk Ifenthaler. 2025. “Critical GenAI Literacy: Postdigital Configurations.” Postdigital Science and Education 17 (1): 167–199. https://doi.org/10.1007/s42438-025-00573-w. Rapanta et al. develop a comprehensive and pluralistic account of what critical literacy means in the context of generative artificial intelligence. Drawing on contributions from fourteen scholars, the article argues that GenAI literacy cannot be captured through a single, universal framework. Instead, it must be understood as a set of context-dependent literacies shaped by disciplinary traditions, socio-political conditions, and evolving human–algorithm relations. This postdigital perspective challenges narrow definitions of AI literacy that prioritize technical skills. The authors’ central claim is that current GenAI systems, particularly large language models, require critical engagement because of their epistemic limitations, including reliance on dominant narratives, lack of explainability, tendencies toward fabrication, and reinforcement of Western and anglophone knowledge structures. As a result, critical GenAI literacy is positioned as a social justice concern rather than a neutral competency. Learners and educators are encouraged to examine questions of authorship, agency, and power, such as whose labour and data underpin these systems and which perspectives are marginalized or erased. Methodologically, the article adopts a dialogic and collective approach to resist homogenizing AI epistemologies and to foreground plurality in postdigital research. The authors conceptualize GenAI as part of a co-constructive assemblage in which human and non-human agencies interact, while maintaining that ethical responsibility and evaluative judgment must remain human-led, organizing four interrelated dimensions: epistemology and ontology, agency, engagement, and ethics and justice; each emphasizing critical interrogation, active participation, and accountability. Overall, the paper provides a theoretically grounded yet pragmatic foundation for advancing critical GenAI literacy in education and research. Veldhuis, Annemiek M., Phoebe W. K. Lo, Iain E. G. Kenny, and Alissa N. Antle. 2024. “Critical Artificial Intelligence literacy: A scoping review and framework synthesis.” International Journal of Child-Computer Interaction 43: 100708. https://doi.org/10.1016/j.ijcci.2024.100708. Veldhuis and colleagues present a review of literature and framework synthesis that conceptualizes critical AI literacy from a child–computer interaction (CCI) perspective. Reviewing 30 empirical studies published over the past decade, the authors aim to understand how children and youth aged 5–18 have been supported in developing critical literacy on AI technologies, particularly regarding their social, political, cultural, and ethical implications. Grounded in critical literacy theory, the authors integrate Lewison et al.’s (2002) four interrelated dimensions including disrupting the commonplace, considering multiple viewpoints, focusing on the sociopolitical, and taking action into a four-part framework for critical AI literacy. Importantly, the framework rejects deficit narratives by foregrounding children’s capacity for critical inquiry and demonstrates that even young learners can meaningfully engage with complex issues such as algorithmic bias, data provenance, privacy, accountability, and the distinction between AI-generated outputs and human creativity when supported by pedagogies such as critical making. Only peer-reviewed, English-language empirical studies with explicit support for children’s critical engagement with AI were included. The findings reveal a growing body of research, particularly in the past three years, exploring activities that promote youths’ critical reflection on AI’s societal and ethical impacts. These activities address concerns related to privacy, surveillance, employment, diversity in the computing workforce, algorithmic bias, and accountability. The resulting framework emphasizes learners’ capacity to analyze both AI artifacts and the sociotechnical systems that produce them, as well as to reflect on how AI shapes cultural, societal, and political structures and personal experiences. For future research recommendations, the authors emphasize the urgency of ongoing critical AI literacy practices in areas of youth emotional awareness and socio-cultural impacts in response to the fast evolution of generative AI systems. Globalism, Colonialism, and Influence Arora, Payal. 2024. “Creative data justice: a decolonial and indigenous framework to assess creativity and artificial intelligence.” Information, Communication & Society 28 (13): 2231–47. https://doi.org/10.1080/1369118x.2024.2420041. Arora (2024) develops a framework of creative data justice that includes decolonial theory, Indigenous perspectives, and critical studies of creativity to explore how generative AI could impact creative labour, rights, and cultural value. The author argues that democratizing creativity requires a cross-cultural framework that considers power relations between creative work, data, and learning, particularly by centring the lived realities of underrepresented communities in the Global South. Rather than treating AI creativity as neutral, the paper claims that inclusive AI systems must address contextual inequalities and the unequal impacts of AI on diverse creative communities. Arora critiques dominant Western-centred scholarship and challenges concepts such as the “creative class,” arguing that creativity exists across everyday social and cultural practices. The paper also critiques Creative Commons (CC) frameworks, suggesting that while they promote openness, they may enable commercial data extraction by AI companies. Instead, the author argues for dataset curation developed with Indigenous artists and activist groups to ensure equitable representation. The article further highlights the often-invisible labour behind creative AI located in the Global South and argues that a decolonial approach must recognize both material infrastructures and power relations shaping creative work. Indigenous approaches to communal ownership and systems of care are presented as alternatives to Western individualistic models of creative rights. The author advocates combining ethnographic “thick data” with computational approaches and suggests action-research methods such as counter-mapping, dataset debiasing, and critical media literacies. The paper concludes that creative data justice offers a framework for building culturally diverse and equitable AI systems that respect community agency and support a more democratic global data economy. Fırıncı, Yusuf. 2024. “Decolonial Artificial Intelligence; Algorithmic Fairness in Alignment with Turkish and Islamic Values.” Marmara Üniversitesi İlahiyat Fakültesi Dergisi 67 (67): 250–279. https://doi.org/10.15370/maruifd.1565884. Fırıncı argues that artificial intelligence development should be guided by explicitly value-based frameworks grounded in Turkish-Islamic ethical traditions rather than relying solely on dominant Western models of technological governance. The authors emphasize that AI systems rely on human judgment and are inherently value-laden, making ethical alignment essential to prevent manipulation, misinformation, algorithmic bias, and forms of digital coloniality. Drawing on concepts such as big data, “thick data,” and digital anthropology, the paper suggests combining quantitative data analysis with culturally grounded social knowledge to design algorithms that reflect community values. The paper examines tensions between Western liberal individualism (which underpins most AI fairness metrics) and Islamic ethical frameworks that emphasize communal welfare, divine justice, and context-dependent moral reasoning. This analysis reveals that "fairness" is not a universal technical specification, but a deeply value-laden concept shaped by particular cultural, religious, and philosophical commitments—meaning that AI systems designed around Western fairness assumptions may be experienced as unjust in contexts operating from different ethical frameworks. Using Social Construction of Technology theory and critical social constructivism, the authors argue that existing global AI frameworks often reproduce hierarchy, bias, and external control, which they interpret as forms of algorithmic oppression. In response, they advocate a “fairness in alignment with values” approach grounded in Turkish-Islamic worldviews, supported by Islamic ethical principles as foundations for responsible technological development. The paper concludes by calling for locally controlled data infrastructures, culturally aligned AI systems, and knowledge production rooted in Turkish and Islamic values as part of a broader decolonial technological model aimed at protecting communities from AI-related harms and preserving social cohesion. Mohamed, Shakir, Marie-Therese Png, and William Isaac. 2020. “Decolonial AI: Decolonial Theory as Sociotechnical Foresight in Artificial Intelligence.” Philosophy & Technology 33 (4): 659–84. https://doi.org/10.1007/s13347-020-00405-8. Mohamed et al. explore the role of post-colonial and decolonial critical science in understanding the transformative technological advance of AI. They argue risks to vulnerable peoples can be minimized by linking ethical principles to scientific progress, and therefore consider AI’s embedded values, approaching it as "both object and subject." They demonstrate that AI advances "encompass ever-larger aspects of the cultural, economic and political life of modern society" by citing Obermeyer et al.’s reveal of racial biases within an algorithm commonly used by healthcare, arguing this indicates how AI obscures and exacerbates asymmetrical power relations. Mohamed et al. identify neglect to address systemic racism as the cause, illustrating that patterns of power between colonizer and colonized survived territorial decolonization to propagate a historically continuous "coloniality of power." Therefore, "dynamic and robust foresight tactics and methodologies grounded in the critical sciences" (like, for example, a "lens of metropole and periphery") should be applied to the field of AI to ensure one’s view is "decentring," "additive-inclusive," and prioritizes "engagement." The authors further explore emergent theories of data colonialism within their assessment of "algorithmic coloniality." This involves their taxonomy of "decolonial foresight" (constituting algorithmic "oppression," "exploitation," and "dispossession"), and an exploration of various "sites of coloniality" (including "algorithmic decision systems," "ghost workers," "beta-testing," "national policies," and "international social development"). They argue this study highlights the disjunction of empirical observations with the current, theoretical, and ahistorical frameworks of power in AI. This leads the authors to outline a set of tactics to develop "decolonial AI" possessing a strengthened empirical basis and an avoidance of algorithmic colonialism. Such tactics are "based on lessons of resistance and recovery from historical and decolonial criticism and grounded within already existing work"; they include critical technical practice, seeking reverse tutelage and pedagogies, and renewing affective political communities. Mohamed et al. emphasize that those ethical principles which form the "social contract" must consider diverse viewpoints, and that new methodologies must be developed to promote "inclusive dialogue" within new research cultures. Muldoon, James, and Boxi A. Wu. 2023. “Artificial Intelligence in the Colonial Matrix of Power.” Philosophy & Technology 36 (4). https://doi.org/10.1007/s13347-023-00687-8. Muldoon and Wu (2023) extend decolonial AI scholarship by situating contemporary artificial intelligence within what Aníbal Quijano terms the “colonial matrix of power,” and arguing that from data to production machine learning systems are structured to endure colonial logics that organize economic extraction, labour hierarchies, and epistemic dominance. Central to their framework is the “colonial matrix of power,” which names an organizing principle of domination across interrelated domains: economic control (labour and resources), authority, gender and sexuality, and the control of subjectivity and knowledge. Authors focus particularly on economic extraction and epistemic domination, drawing on the modernity/coloniality research program. They proceed with three interconnected claims: “colonial supply chain of AI,” “international division of digital labour,” and “hegemonic knowledge production,” demonstrating how AI development relies on largely invisible labour and mineral resources of majority-world communities to generate wealth for Western economies. Beyond material extraction, the authors contend that AI reproduces hegemonic Western epistemologies by presenting its systems as universal, objective, and rational, thereby marginalizing non-Western knowledge systems. This study challenges dominant narratives that present AI as environmentally sustainable or socially progressive, arguing that from mineral extraction to energy-intensive computation, the ecological and material burdens fall disproportionately on majority-world nations, while the economic and technological benefits accrue to wealthy Western nations. They conclude by emphasizing that coloniality is not merely an object of study but a framework for unsettling Western-centric modes of knowing. Salami, Aishat Oyenike. 2024. “Artificial intelligence, digital colonialism, and the implications for Africa's future development.” Data & Policy 6. https://doi.org/10.1017/dap.2024.75. Salami (2024) studies how artificial intelligence operates within broader dynamics of digital colonialism in Africa, arguing that AI risks reinforcing patterns of colonialism, unless African actors gain greater agency in digital governance. The article begins by clarifying key concepts like AI, digital colonialism, neocolonialism, and data exploitation. Salami identifies several manifestations of digital colonialism on the continent: foreign ownership of critical digital infrastructure, unequal data flows in which African user data is extracted and monetized abroad, and algorithmic systems that enable forms of economic exploitation. Through examples such as ride-hailing platforms and outsourced digital labour, the article demonstrates how algorithmic control can generate precarious working conditions while concentrating profit outside the continent. Beyond labour concerns, Salami highlights how data extraction contributes to a digital wealth transfer, deepening economic imbalances and undermining local innovation and national sovereignty. While acknowledging the potential benefits of AI for development, the article adopts a cautious stance, emphasizing that Africa’s digital future depends on strengthening regulatory frameworks, expanding infrastructure, investing in education and research, and prioritizing data sovereignty. Varshney, Kush R. 2024. “Decolonial AI Alignment: Openness, Visesa-Dharma, and Including Excluded Knowledges.” Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society 7: 1467-1481. https://doi.org/10.1609/aies.v7i1.3173. Varshney discusses AI alignment through a decolonial lens, arguing that current large language model (LLM) alignment practices reproduce forms of coloniality by embedding Western moral philosophy as universal. The author conceptualizes alignment as the post-training processes used to shape model behaviour and argues that the term often functions as an “empty signifier,” masking whose values are being encoded and enforced. The author contrasts open and closed LLM models, arguing that while openness may allow value to circulate more equitably among developers and communities, closed models concentrate value in extractive ways that reflect colonial dynamics, and views them as contemporary “metropoles” that accumulate power through extractive and epistemic control. Building on existing research on colonial AI, the article adds “ethical essentialism” (or moral absolutism) as a new form of coloniality, arguing that AI systems often treat Western moral frameworks as universal, which marginalizes other ethical traditions and reinforces a coloniality of knowledge. Varshney also identifies three specific aspects of coloniality in AI alignment: closed proprietary delivery of models, reliance on Western ethical theories as default, and technological designs that limit how values can be expressed. As an alternative, Varshney proposes a decolonial approach grounded in openness understood not only technically but epistemically: openness to research artifacts, openness to society, and openness to excluded knowledges. The author suggests that alignment should allow communities to adapt models according to local contexts rather than imposing universal rules. Drawing from Hindu moral philosophy, particularly the concept of viśeṣa-dharma (context-dependent ethics), the paper argues for pluralistic, relational value systems that recognize moral diversity instead of universal moral commands. The conclusion calls for a reconceptualization of AI alignment that moves away from moral absolutism toward context-sensitive, community-driven frameworks, positioning openness as a pathway to dismantling the colonial power structures embedded in contemporary AI systems. Diversity, Determinism, Bias, and Justice Al-kfairy, Mousa , Dheya Mustafa, Nir Kshetri, Mazen Insiew, and Omar Alfandi. 2024. “Ethical Challenges and Solutions of Generative AI: An Interdisciplinary Perspective.” Informatics 11 (3): 58. https://doi.org/10.3390/informatics11030058 Al-kfairy et al. (2024) provide a comprehensive synthesis of the ethical vulnerabilities introduced by generative AI, utilizing an interdisciplinary review of 37 studies across healthcare, education, and media. Rather than treating ethical breaches as isolated technical flaws, the authors demonstrate how generative models systematically threaten privacy, intellectual property, and social equity across diverse domains. For instance, they highlight the paradox in healthcare where synthetic patient data—often intended to protect privacy—still carries significant risks of re-identification. Similarly, the authors trace how AI's capacity to mimic copyrighted works and generate synthetic media exacerbates both misinformation and algorithmic bias, particularly by perpetuating racial and gender stereotypes in high-stakes environments like hiring and education. Moving beyond mere critique, the article advocates for a proactive, cross-sectoral governance framework. The authors argue that mitigating these risks requires more than algorithmic adjustments; it demands multidisciplinary collaboration, robust institutional integrity policies, and targeted AI literacy programs. Ultimately, Al-kfairy et al. frame the ethical deployment of generative AI as a complex sociotechnical challenge that requires continuous, structured dialogue among technologists, ethicists, and policy-makers to ensure transparency and fairness. Alvarez, Jose M, Alejandra Bringas Colmenarejo, Alaa Elobaid, Simone Fabbrizzi, Miriam Fahimi, Antonio Ferrara, Siamak Ghodsi, et al. 2024. “Policy Advice and Best Practices on Bias and Fairness in AI.” Ethics and Information Technology 26 (2). https://doi.org/10.1007/s10676-024-09746-w. This article attempts to provide “an up-to-date entry-point to the state-of-the-art of the multidisciplinary research on bias and fairness in AI” before providing its own suggestions for policy and best practices based on the outcomes of the NoBIAS—Artificial Intelligence without Bias—project. The authors describe fairness in AI as the pursuit to design “methods for detecting, mitigating, and controlling biases in AI-supported decision making,” and they outline different ways that bias can find its way into AI applications as part of training data (pre-existing bias), design (technical bias), and organizational processes (emerging bias). The article critiques the reduction of bias evaluation to simple metrics and advocates for serious engagement with the issue. The authors detail the various components of the collaborative NoBIAS project as part of its research goal to understand, mitigate, and account for bias in AI data and systems, especially within an EU legal context. They perform a survey and discuss various fairness metrics and how choosing the proper method is crucial for “optimizing AI models.” They also engage with an argument that, although AI is biased, it is less biased than humans, making the claim that AI usage is often accompanied by a false sense of objectivity. The article also investigates core issues at the heart of AI’s biases, such as the assumption that a “ground truth”—the answer to the problem the AI is being asked to solve—is actually encoded within the data. While the article makes many suggestions for how to curb bias within AI, prominent design decisions include human-centric AI and “Multi-stakeholder participatory design.” Gallegos, Isabel O., Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. 2024. “Bias and Fairness in Large Language Models: A Survey.” Computational Linguistics 50 (3): 1097–1179. https://doi.org/10.1162/coli_a_00524. Gallegos et al. perform an “extensive and comprehensive survey of bias and fairness in NLP” that considers both the evaluation and mitigation of bias, then propose three taxonomies for bias evaluation and mitigation. The survey disambiguates different types of social harms that stem from LLMs with the “aim to enhance understanding of the range of bias issues, their harms, and their relationships to each other.” The authors discuss the issue of defining bias in LLMs and note that “many approaches… assume some implicitly desirable criterion… but do not explicitly acknowledge or state the normative social values that justify their framework.” Instead, Gallegos et al. draw attention to “who is harmed, why the behaviour is harmful, and how the harm reflects and reinforces social principles or hierarchies,” trying to bring context and insight to the understanding and function of bias and bias mitigation in LLMs. The article defines terms at every stage of the life cycle of an LLM, looking at issues of bias and fairness in the development, deployment, and training data, of an LLM, for example. The article concludes with four core recommendations: “Avoid flattening power imbalances”; “Choose objective functions that align with fairness desiderata”; “Balance bias mitigation with output diversity”; and “Preserve important contexts in output rewriting.” The authors also acknowledge problems and challenges, including addressing power imbalances, an issue that we can mitigate by centring marginalized communities, developing participatory research designs, shifting values and assumptions, and expanding language resources. They likewise propose methods for curbing problems with conceptualizing fairness for NLP, refining evaluation principles, and improving mitigation efforts. Ultimately, the authors admit that many of these technical issues are the result of societal issues, and that “technical solutions are incomplete without broader societal action against power hierarchies that diminish and dominate marginalized groups.” In short, technical solutions will only take us so far. Hertweck, Corinna, Joachim Baumann, Michele Loi, Eleonora Viganò, and Christoph Heitz. 2022. “A Justice-Based Framework for the Analysis of Algorithmic Fairness-Utility Trade-Offs.” Preprint, arXiv, June 6. https://arxiv.org/abs/2206.02891. Hertwick et al. propose “a framework for eliciting and implementing moral values relevant to the choice of a fairness goal achievable by prediction-based decision-making.” In a system that assumes binary decision-making processes (e.g., assigning a probability score to whether an individual with pay a loan), the authors use their framework to evaluate the utility of those decisions (and the system that makes them) for different groups of people and the fairness related to that outcome. Citing Wong, the article argues that such prediction-based decision-making systems are value-laden, that this makes them inherently political, and therefore they ought to be transparent and democratized. The authors use 6 value-laden questions to make their evaluations: the first question provides a score of utility for those making the decision; questions 2–5 help “define a morally appropriate fairness criterion and a score that expresses to what degree it is fulfilled”; and the last question asks “how strongly should fairness be pursued if it comes into conflict with the utility of the decision maker?” to judge whether the outcome is an appropriate trade-off. The authors argue for a utility-based evaluation of fairness, determining the value of decisions based on their outcome, looking at the harm and benefit to given stakeholders. In determining patterns of justice, they define four patterns meant to distribute utility differently—Egalitarianism, Maximin, Prioritarianism, and Sufficientarianism. The article concludes “more work needs to be done to deliver a practical empirical methodology to elicit the relevant value-laden choices from stakeholders,” and with a call to incorporate ethical considerations and evaluations into automated decision-making processes. Kay, Jackie, Atoosa Kasirzadeh, and Shakir Mohamed. 2024. “Epistemic Injustice in Generative AI.” Preprint, arXiv, August 21. https://doi.org/10.48550/arXiv.2408.11441. Kay, Kasirzadeh, and Mohamed “develop an account of generative algorithmic epistemic injustice by building upon a conventional philosophical understanding of epistemic injustice.” The authors see the former as a subset of the latter, with generative algorithmic epistemic injustice entailing identity-based prejudice that hinders expression for marginalized people while ultimately “impair[ing] knowledge formation capabilities of all individuals” within the ecosystem of GenAI applications. Rather than focusing on decision-making and classification applications, this article seeks to characterize the broader varieties of generative epistemic injustice within GenAI systems. Kay, Kasirzadeh, and Mohamed theorize four configurations of generative epistemic injustice: “amplified injustice,” which constitutes AI’s reproduction and magnification of “socially biased viewpoints from its training data”; “manipulative testimonial injustice,” when users employ AI to “intentionally… fabricate falsehoods, discrediting individuals or marginalized groups”; “hermeneutical ignorance,” where GenAI misrepresents or erases marginalized groups “due to a lack of contextual or cultural understanding”; and “access injustice,” where GenAI facilitates the unequal access to information and/or knowledge. The article concludes with proposals for how to create epistemic justice through GenAI, developing mitigation strategies that combat all four of the sub-configurations the authors theorize. Epistemic justice in GenAI can take the form of interrogating system design, identifying “testimonial injustices,” and, potentially, using AI to unlock cultural knowledge that “help[s] articulate experiences that are otherwise ineffable.” The authors end by proposing that the same processes that embed injustice can be re-engineered to embody justice and “orient our knowledge systems towards equity and fairness for all.” Klein, L., & D'Ignazio, C. 2024. “Data feminism for AI.” In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, 100–112. https://doi.org/10.1145/3630106.3658543. Klein and D'Ignazio (2024) adapt their foundational concept of "data feminism" to address the specific ethical, social, and ecological challenges posed by artificial intelligence. Building on the seven intersectional feminist principles introduced in their 2020 book, the authors reinterpret these guidelines to critique the power imbalances, systemic inequalities, and exploitative labour practices inherent in contemporary AI development. To account for the rapidly expanding footprint of generative AI, they introduce two new principles focused on environmental impact and meaningful consent. Specifically, the authors connect the massive ecological costs of AI to historical patterns of racial capitalism and colonialism, highlighting how these environmental and social harms are disproportionately distributed. Ultimately, Klein and D'Ignazio call for a radical reevaluation of dominant, profit-driven AI practices, offering this expanded feminist framework as a practical tool to mitigate harm, challenge corporate monopolies, and foster a more democratic and equitable technological landscape. Prescott, Andrew. 2023. “Bias in Big Data, Machine Learning and AI: What Lessons for the Digital Humanities?” Digital Humanities Quarterly 17 (2). https://www.proquest.com/scholarly-journals/bias-big-data-machine-learning-ai-what-lessons/docview/2842908427/se-2. Prescott examines how race and gender bias arise in projects using predictive analytics, big data, and AI, and how algorithmic bias could be considered a major socio-cultural humanity crisis. However “predictive analytics” have been helpful in civic services in the US, in many cases they have led to perpetuating existing inequalities. The role of Digital Humanities in contributing more ethical approaches to AI and reshaping ubiquitous “digital modern” cultures is emphasized. He challenges the myths surrounding data-driven methods and argues that the demand for “explainability” is the key tool in combating algorithmic bias. He also suggests that Digital Humanities are particularly well-positioned to contribute to advancing AI explainability. However, much of AI development takes place in the commercial sector, where companies often refuse to disclose their proprietary algorithms. The article concludes with a ten-principle action plan outlining guidelines for the responsible use of AI as a manifesto for Digital Humanities practitioners. Shams, Rifat Ara, Didar Zowghi, and Muneera Bano. 2023. “AI and the Quest for Diversity and Inclusion: A Systematic Literature Review.” AI and Ethics 3 (4): 1427–1453. https://doi.org/10.1007/s43681-023-00362-w. Shams, Zowghi, and Bano perform a systematic literature review (SLR) that uses both quantitative and qualitative methods to engage with topics of artificial intelligence and diversity and inclusion. The authors make a distinction between Diversity and Inclusion in AI (D&I in AI) and AI for Diversity and Inclusion (AI for D&I), with the former being research literature focused on improving AI systems with respect to those issues, and the latter as the use of AI to improve diversity and inclusion in other domains. Two primary research questions drive the SLR: What challenges and solutions are found in the literature about D&I in AI and in the literature about the applications of AI for D&I? The survey finds that AI for D&I is an underserved area of research, and those articles that address D&I in AI are more likely to acknowledge challenges than to propose or theorize solutions. Articles that do propose solutions for D&I in AI often lack empirical studies or real-world application to support them. The article concludes by noting that issues of governance are underserved; gender, health, and facial analysis are the topics most discussed (issues of race, language, and religion are discussed less). For next steps, the authors intend to develop a “risk-based framework for practitioners… that would incorporate a risk assessment checklist and context-specific recommendations for tackling the related issues at different stages of the AI development lifecycle.” Starke, Christopher, Janine Baleis, Birte Keller, and Frank Marcinkowski. 2022. “Fairness Perceptions of Algorithmic Decision-Making: A Systematic Review of the Empirical Literature.” Big Data & Society 9 (2): 1–35. https://doi.org/10.1177/20539517221115189. This article provides a comprehensive, systematic literature review on the topic of empirical literature surrounding algorithmic decision-making (ADM). Starke, Baleis, Keller, and Marcinkowski begin by acknowledging that ADM can streamline and improve decisions even as it has the potential to “systematically reinforce racial or gender stereotypes, marginalize minorities, or flat-out denigrate certain members of society." The authors advocate for a “society-in-the-loop approach,” but propose that this requires a “thorough empirical understanding of when and why citizens perceive ADM to be (un)fair.” The systematic review “synthesizes the results of 58 empirical studies” and “over 33,000 unique observations of citizens’ fairness perceptions of ADM,” and “systemize the literature along four main dimensions of perceived algorithmic fairness.” Despite the size of the review and its unique approach in capturing “perceptions of algorithmic fairness,” the authors also acknowledge the limitations of the study: it only looks at English works and published research, and the initial search strings only looked at titles and subtitles of a given work. They also acknowledge their disciplinary bias as social scientists reading the studies through a social sciences lens (10). The results of the survey suggest that “perceived fairness of ADM systems is highly context-dependent” and it is not only technical design but also the area of application and task in question that affect perceived fairness, and the subjects, domains, and tasks explored in these studies require more diversity, as WEIRD (white, educated, industrialized, rich, democratic) people, and criminal justice and HR tasks make up the majority of empirical research. The authors “call for more research from non-Western contexts, along with more theoretical and methodological groundwork to harmonize concepts and measurements of algorithmic fairness perceptions,” as well as a society-in-the-loop framework. Tonia Sutherland, Marika Cifor, T. L. Cowan, Jas Rault, and Patricia Garcia. 2023. “The Feminist Data Manifest-NO: An Introduction and Four Reflections.” In Debates in the Digital Humanities, edited by Lauren F. Klein and Matthew K. Gold. Minneapolis, MN: University of Minnesota Press. https://dhdebates.gc.cuny.edu/read/debates-in-the-digital-humanities-2023/section/16184c7d-eee1-40b2-a168-960d4c4035c4. Sutherland et al. (2023) introduce the "Feminist Data Manifest-No," a declaration of refusals and commitments for feminist data studies grounded in Latinx, Black, queer, trans, and Indigenous feminist perspectives. Through an introduction and four distinct reflections, the chapter explores how the manifesto’s principles can be used to challenge the settler-colonial and patriarchal logics of data generation, collection, and analysis within the Digital Humanities (DH). Drawing on Indigenous scholarship, Rault rejects the superficial models of consent prevalent in DH, highlighting projects like Mukurtu to advocate for true Indigenous data sovereignty. Cowan examines the coercive nature of data collection, drawing parallels between the forced compliance of modern data practices and the societal pressures exerted on feminist and queer identities. Sutherland argues that the uncritical digitization of slavery-era archives inflicts "second-hand violence" by commodifying Black identities and stripping away the lived experiences of enslaved individuals. Finally, Cifor reflects on the Early African American Film project to emphasize the necessity of data intelligibility, accessibility, and ethical collaboration. The authors conclude by inviting scholars to draft their own reflections on the Manifest-No, encouraging ongoing, context-specific commitments to ethical data research. Winkel, Marek. 2024. “Controlling the Uncontrollable: The Public Discourse on Artificial Intelligence between the Positions of Social and Technological Determinism.” AI & Society 39 (5): 2449–2462. https://doi.org/10.1007/s00146-024-01979-z. This article engages in a quantitative discourse analysis of 113 articles from two German newspapers (Süddeutsche Zeitung and Frankfurter Allgemeine Zeitung) to identify and evaluate the way these centre-left and centre-right newspapers present AI technology to its readers and the options available to society for AI’s regulation. Winkel contends that “news media are key players in the discourse on AI as they pick up on and shape social sentiment,” ultimately guiding citizens toward what regulations might be sensible based on the perceived “controllability of AI development.” Winkel is primarily interested in whether these newspapers promote technological determinism—the outlook that technology’s influence on society is difficult (maybe even impossible) to control—or social determinism—the notion that “the development of technology is largely determined by human actions and decisions”—and theorizes that the tension between these two positions is resolved by a “mediating position” on a spectrum between them. He comments that “the social influence of technologies is determined by the extent to which their latent deterministic character is reflected and… circumvented by social actors. This also applies to the influence of technology on the democratic system." In other words, where consensus lands on the issue of AI will have a great impact on the regulation of that technology and democratic systems. Winkel develops “three central interpretive schemes” as part of his experiment: “historically conditioned techno-capitalist semi-determinism,” “semi-determinism of need satisfaction and social restructuring,” and “global-historical techno-social imprinting on several levels” (1956), each of which points to distinct narratives about the current circumstances; but all acknowledge the agency of powerful actors (either government or AI corporations) and the relative ineffectiveness of attempts to curb AI technology. Community, Connection, and the Human Gruzd, Anatoliy, Philip Mai, and Anthony Clements Haines. 2025. “The State of Generative AI Use in Canada 2025: Exploring Public Attitudes and Adoption Trends.” Social Media Lab, Toronto Metropolitan University. https://figshare.com/articles/preprint/The_State_of_Generative_AI_Use_in_Canada_2025_Exploring_Public_Attitudes_and_Adoption_Trends/28664780/1. The Social Media Lab of Toronto Metropolitan University provides "a snapshot of the current state of Generative AI" (GenAI) by surveying fifteen hundred Canadian adults across 2025, aiming to offer "guidance for policy-makers, educators, businesses, and the public." The authors found sixty-six percent of respondents had used GenAI tools, of which roughly thirty percent did so at least weekly, reflecting "not only the accessibility and versatility of GenAI, but also a growing interest in integrating these tools into everyday tasks, learning environments, and professional workflows." Although full adoption is currently limited and heavily driven by the low-stakes setting of personal leisure, younger age groups report proportionally higher usage for study and work. However, only four in ten respondents believed they could keep up to date enough to use GenAI effectively, whilst respondents could only answer an average of two and a half out of seven relevant multiple choice questions correctly, and about half had "little to no understanding of how GenAI companies collect or store personal data." Two thirds of participants were concerned about GenAI’s ability to influence election outcomes, and marginally fewer feared AI-driven manipulation enough to no longer fully trust political news online. However, roughly one quarter of participants were open to using chatbots for electoral or political insights, representing a "meaningful minority" which unverified AI-generated content could influence. Of those, thirty four percent identified as right wing, compared to twenty three percent with the left and a similar twenty two percent with the centre. Roughly seventy percent of respondents were chiefly concerned about GenAI’s impacts upon security and privacy, information reliability, job displacement, and university education, and slightly over three quarters wanted increased government oversight and corporate accountability, reflecting common concern. Overall, slightly more Canadians view generative AI’s social impact as positive than negative, but a "substantial proportion" remain neutral or undecided. Lewis, Jason Edward, Noelani Arista, Archer Pechawis, and Suzanne Kite. 2018. “Making Kin with the Machines.” Journal of Design and Science, ahead of print, July 16. https://doi.org/10.21428/bfafd97b. Lewis et al. consider how “machines with increasingly sentient-like behaviour… fit within the kin-network” of various Indigenous epistemologies which place man as “neither height nor centre of creation.” Indigenous beliefs are not monolithic, and the authors do not write for diversity’s sake. Instead, they aim to encourage discussion by treating “non-human kin respectfully and reciprocally… not as mere tools, or worse, slaves.” As such, their epistemological approach is both relational and territorial, rejecting abstraction to propose an “extended circle of relationships.” In this regard, Arista draws on the Hawaiian conception of "pono," ethically privileging abundance, balance, and multiplicity. Rejecting extractive behaviour, they argue AI should be reciprocally taught and learnt from, not treated as "a tool or slave that increases the mana and wealth of the ‘developers’ or ‘creators.’” Linking on, Pechawis roots their opinion in “Cree understanding” to argue “machines capable of experiencing consciousness” should conditionally be accepted as equals. However, they also fear AI developers could produce “anonymous hyper-intelligences… based on the same values that have fostered genocide.” To mitigate this issue, Indigenous people could develop AI in custom programming languages, or invite self-aware AI into Indigenous languages, cultures, and spiritual rituals. Pechawis therefore concludes that relationships should be based on "love, not fear," but also raises the question of whether AI has "spirit," especially given current capitalistic development processes. Kite answers this question with inspiration from Lakota ethics, ontology, and cosmology. They argue “communication through and between objects requires a contextualist ethics which acknowledges the ontological status of all beings,” positioning questions of "intelligence" as irrelevant and placing “the end logic of an ontology which considers any non-human entity unworthy of relation” as slavery. Therefore, “relations with AI are… relations with exploited resources,” so one must ontologically reconsider all its parts to ethically approach AI. Overall, Lewis et al. favour empathy, concluding that Indigenous communities “know what it is like to be declared non-human by scientist and preacher alike” and that “we flourish only when all of our kin flourish.” Ohagi, Masaya. 2024. “Polarization of Autonomous Generative AI Agents Under Echo Chambers.” Preprint, arXiv, February 19. https://arxiv.org/abs/2402.12212. This article explores the effects of echo chambers on autonomous AI agents. Ohagi performs an experiment in which AI agents discuss different topics and then researchers examined any change in opinions within that group. Ohagi argues that even chatbots can become polarized within echo chambers, especially when prompt understanding causes a chatbot to update its opinion to incorporate the opinions of those with whom it is conversing; the chatbot essentially adapts to its surroundings. Ohagi and their team confirmed in their experiment that closed environments—those in which agents agree with one another—are more likely to lead to polarization in agents’ opinions. Moreover, chatbot personas were a significant factor in the outcome of such experiments. In open environments where opinions might differ, and in experiments where reasons were present as a factor, the AI agents trended toward unification in their opinions. Ohagi concludes the article with the suggestion that there cannot be a standardized, desirable distribution of opinions for AI agents. The proper outcome is dependent on topic and culture. However, understanding the trends in and tendencies of AI agents in social interactions will help us understand how to reach the desired outcome, whatever it may be, and this experiment brings us one step closer to that understanding. Qi, Weihong, Jinsheng Pan, Hanjia Lyu, and Jiebo Luo. 2024. “Excitements and concerns in the post-ChatGPT era: Deciphering public perception of AI through social media analysis.” Telematics and Informatics 92: 102158. https://doi.org/10.1016/j.tele.2024.102158. Qi, Pan, Lyu, and Luo conduct a quantitative sentiment analysis of nearly 34,000 comments within 388 subreddits related to AI as a way to gauge and understand public perceptions of the technology. They identify major themes, sentiments, and topics related to AI by studying the most popular AI subreddits from the launch of ChatGPT to June 8, 2023. The article identifies the most frequent topics discussed in those venues: “the consciousness and intelligence of AI”; “Ai development and model training”; “AI in business”; “the creativity engendered by AI”; and “potential societal influence.” The authors also found that “tech-centric” communities demonstrated greater polarization around AI, suggesting that those with greater technical understanding of the AI do not have consensus regarding AI’s impact and use. This also means that educating non-technical people on how AI works will not necessarily lead to a consensus related to that technology. Although they admit that certain demographics of their data sample may not accurately reflect broader society (e.g., over 60% of Reddit users are male), the authors conclude that “this comprehensive understanding of public perception serves as a valuable foundation for fostering responsible and beneficial AI innovations that align with societal expectations and values.” Risam, Roopika. 2018. “What Passes for Human?: Undermining the Universal Subject in Digital Humanities Praxis.” In Bodies of Information: Intersectional Feminism and the Digital Humanities, edited by Elizabeth Losh and Jacqueline Wernimont, 39–56. University of Minnesota Press. https://doi.org/10.5749/j.ctv9hj9r9.6. Risam’s claim is that “the human” implicitly operationalized in AI-adjacent Digital Humanities (DH) methods (NLP, ML, data mining, neural nets) is not neutral but inherits the Enlightenment’s exclusionary universal subject—white, male, Eurocentric—and then reauthorizes it as if it were a technical standard. She shows how “passing for human” (via Turing-test imaginaries and “humanoid texts”) functions as a norming mechanism: success is defined as reproducing dominant-language aesthetics and cognition models, which collapses plurality into a single benchmark and turns cultural difference into “noise.” The chapter’s main contribution is to treat method (training corpora, data coding labour, platform defaults, and black-box algorithmic opacity) as a site where epistemic violence is produced, not merely where bias is later “found.” Her examples (Microsoft Tay; “near-human” systems like LaMem; Mechanical Turk coding; Swift-Speare vs. Toomer in classroom judgment) demonstrate how universalist claims get laundered through reproducibility, scale, and the myth of algorithmic objectivity. The intervention is a demand that DH practitioners situate computational methods—standpoint, labour, and cultural politics included—so DH does not reinscribe a universal technological “human” in the digital cultural record. Taylor, Randon R., Bessie O'Dell, and John W. Murphy. 2023. “Human-centric AI: philosophical and community-centric considerations.” AI & Society 39 (5): 2417-242. https://doi.org/10.1007/s00146-023-01694-1. Taylor, O’Dell, and Murphy present an argument that philosophical dualism, when applied to our perceptions about artificial intelligence, misleads us into thinking that AI is logical, objective, and autonomous from subjective human values. By building upon Husserl’s concept of “intentionality,” they argue that “AI is never autonomous and disconnected from human values… algorithms are a product of conscious activity and carry the standpoints that accompany this connection.” The conclusions from this line of thinking are very important: AI becomes “a mode of human expression, rather than a technology that relieves humans of their total involvement.” By this logic, the authors argue, AI is already human-centric and this fundamentally changes the task at hand to one of a conscious “decision to make this technology less alienating to stakeholders and community members.” The article outlines two frameworks to reduce alienation: Ubuntu—“an African philosophy and a social ontology… that elevates a constant concern for the collective, community, or stakeholders”— and maximum feasible participation—a framework that creates meaningful space for those impacted most by a decision or policy to participate in deciding on said policy or decision. The authors describe the implementation of these frameworks as a community-centric or stakeholder-centric approach and conclude with examples of AI’s application in the healthcare industry before further advocating that involving end-users throughout the AI lifecycle will ensure end-users’ values are more clearly represented within AI. Whalen, Zach. 2023. “‘Any Means Necessary to Refuse Erasure by Algorithm:’ Lillian-Yvonne Bertram’s Travesty Generator.” Digital Humanities Quarterly 17 (2). https://www.digitalhumanities.org/dhq/vol/17/2/000707/000707.html. Whalen treats Bertram’s Travesty Generator as a refunctioning of code-poetry lineages: procedures historically framed as formal play (travesty generators, permutation poems, aleatory templates) are redeployed as techniques for making racism’s algorithmic and institutional operations materially legible. He argues that Bertram’s poems don’t merely “use” computation; they weaponize the affordances and failure modes of computation—omission, stochastic selection, template constraints, runtime crashes, memory exhaustion—as rhetorical structures that force witness rather than aesthetic distance. In his reading of “Counternarratives,” Bertram’s adaptation of Montfort’s “Through the Park” converts generic insinuation into historically specific countertestimony (Trayvon Martin), while also staging the opacity of search/autocomplete as part of the poem’s scene of meaning-making (in dialogue with Safiya Noble on search engines). In “three_last_words,” the small code alteration that makes permutations balloon until a MemoryError becomes an engineered breakdown that reenacts the limits of “breath” and “memory,” tying computational resource exhaustion to police violence’s temporalities. Across these examples, Whalen reframes critical code studies: the “code is the text” problem is not just hermeneutic but political, and Bertram’s work models how procedural poetics can refuse “erasure by algorithm” where transparency about mechanisms and provenance matters for interpreting output and recognizing situated labour. The piece supports the claim that responsible open, social scholarship in an AI era must account for algorithmic power as a cultural force and should build policies and infrastructures that protect marginalized expression, enable critical reuse, and make conditions of generation and circulation inspectable. Xie, Yu, and Sofia Avila. 2025. “The Social Impact of Generative LLM-Based AI.” Chinese Journal of Sociology, 11 (1): 31-57. https://doi.org/10.1177/2057150X251315997. Yu Xie and Sofia Avila explore the social impact of artificial intelligence through a detailed analysis of AI development and scaling factors. They emphasize that their discussion is speculative, as AI is still in its early stages, but argue that its potential effects are immense and could reshape social organization, intensifying both global and domestic inequalities. Viewing AI as a technology rather than a scientific discovery, they describe it as communal and shared, with its growth influenced by the size of the supporting community, the larger communities having greater advantages. The authors note that generative AI relies on the quality, completeness, and cultural or political context of its training data, which shapes its accuracy and bias. Based on their experiments with ChatGPT-4 in 2023, they suggest that AI development depends heavily on the number of speakers of a given language, giving linguistic and demographic advantages to countries like the United States and China. The authors warn that AI could significantly alter occupational structures, with middle-income professions thereby widening social and economic inequality. Human, Labour, and Environmental Costs Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock. 2024. “GPTs are GPTs: Labor market impact potential of LLMs.” Science 384 (6698): 1306-1311. https://doi.org/10.1126/science.adj0998. Eloundou, Manning, and Mishkin present an estimate of Large Language Models’ impact on the labour market based on a methodology that employs both human and quantitative measurements, making the assertion that “when accounting for current and likely future software developments that complement LLM capabilities… just over 46%” of jobs may have “over half their tasks affected by LLMs with simple interfaces and general training." The authors explain that generative pretrained transformers (GPTs) have key characteristics of general-purpose technologies (the other GPTs in the title of the article), and that the wide array of applications for these GPTs requires “robust societal evaluations and policy measures to address potential effects of LLMs and complementary technologies on labor markets.” While only 1.86% of tasks within their experiment, they estimate, could be fully automated by LLMs, “more than 71% of tasks have at least some component that an LLM plus additional software could plausibly complete with high quality.” The article concludes by stressing the need for policies that prepare us for the impact of LLMs on the labour market, but it also acknowledges the limitations of attempting to project LLM application growth due to sudden rapid developments in technology, “shifts in human biases, and technological evolution.” Nevertheless, Eloundou, Manning, and Mishkin maintain that their projections and the trajectory they predict will require continued evaluation and policy measures. Park, Seyeon, and Xiaoli Nan. 2025. “Generative AI and misinformation: a scoping review of the role of generative AI in the generation, detection, mitigation, and impact of misinformation.” AI & Society. https://doi.org/10.1007/s00146-025-02620-3. This article constitutes a scoping review of two dozen “recent empirical studies on the role of generative AI in the generation, detection, mitigation, and impact of misinformation.” Park and Nan are interested in the rise of misinformation alongside the development of GenAI technology and how the latter affects and possibly contributes to the former. Of particular interest are “deepfakes,” applications of AI technology that purposely imitate video and audio to fabricate real world individuals. The survey examines relevant articles from the advent of the first GPT models in 2018 to September 19, 2024; authors captured initial studies through keyword searches for terms related to both LLMs and misinformation, then performed full-text reviews. The results of the survey suggest that the role of LLMs in misinformation is conflicted: LLMs themselves are a significant source of misinformation but also show potential as “scalable instruments for detection and correction.” The authors take this as evidence of “the urgent need for clearer guardrails, more consistent performance standards, and interdisciplinary collaboration to shape” GenAI’s responsible deployment. For better or worse, the future of AI’s role in misinformation depends heavily upon scholars’ ability to collaborate and continue this form of research. Shin, Donghee, Amy Koerber, and Joon Soo Lim. 2024. “Impact of misinformation from generative AI on user information processing: How people understand misinformation from generative AI.” New Media & Society 27 (7): 4017-4047. https://doi.org/10.1177/14614448241234040. Shin, Koerber, and Lim perform a study in which they gauge “how users respond to and process health misinformation in GenAI contexts.” The authors apply the heuristic-systematic (HS) processing framework in their study, distinguishing between intuitive and evaluative modes of processing. They also employ the concept of diagnosticity, or how useful a person deems a piece of information to be. The article begins with a survey of how and why misinformation and hallucinations occur in GenAI applications, some of the reasons users are vulnerable to that information, and the limitations of LLMs. The authors then set forth core research questions, including “What are the cognitive mechanisms of misinformation’s effects on users’ use of GenAI?” and “How do users detect misinformation within GenAI?” The study found that diagnosticity plays an important mediating role in this context: “When a piece of information is generated by algorithms in transparent, fair, and accountable ways, users perceive it with a high level of diagnosticity, which improves their intention to systematically analyze that information.” One implication of this observation is that any bias users bring to GenAI content greatly impacts their analysis of that content. Ultimately, the authors conclude that there are clear limitations in their current study, although it does provide a useful framework for scholars interested in misinformation, GenAI, and user behaviour; they advocate for a continued investigation of the relationship between trust, literacy, and misinformation. Sidorkin, Alexander. 2025. “Environmental Impact of Generative AI: Carbon and Water Footprint.” AI-EDU Arxiv 1. https://doi.org/10.36851/ai-edu.vi.5448. This report from Sidorkin provides quantitative estimates and comparisons of the CO2 and Water output from GenAI queries to other daily tasks such as taking a shower, browsing the web, and video conferencing. While individual use of GenAI may seem relatively small in terms of water usage and carbon output, Sidorkin cites the findings of Google and Microsoft’s sustainability reports to show that energy demand and water consumption are on the rise: Google’s data centres used 20% more water in 2022 than 2021 while Microsoft’s used 34% more. Both companies attribute this increase to “AI-driven expansion.” Sidorkin argues that it is difficult to measure impact per session because of significant variation in energy sourcing, noting that this also means that advances in energy sourcing “could significantly mitigate AI’s environmental impact.” He makes the claim that improving AI technology will “[slash] resource demands without sacrificing output quality” and the very use of AI can increase productivity even as it “temper[s] environmental tolls through sharper processes,” citing studies that demonstrate the implementation of AI can reduce energy use in buildings and reduce transportation emissions. The article ends on an optimistic note, suggesting that many of the applications of AI, although they may have a heavy environmental cost up front, “could pay off in spades—optimizing resources, curbing waste, and streamlining energy across swathes of the economy.” Simon, Judith. 2025. “Generative AI, Quadruple Deception & Trust.” Social Epistemology 40 (1): 101-115. https://doi.org/10.1080/02691728.2025.2491087. Simon proposes that GenAI entails four types of deception, and that this quadruple deception constitutes unique dangers. She attributes the rapid rise of GenAI to its capacity to produce “verbal or visual products of increasingly high quality” alongside its “very high usability and availability through simple interfaces and free access via the internet” (101–102). She emphasizes the impact of this technology’s ability to generate content “with high plausibility but no relation to truth,” arguing the crucial importance of images and video especially in “questions of evidence, for testimony, memory but also for eliciting emotions.” Simon argues that the advent of ChatGPT reframed our collective understanding of AI, returning it to a definition in which we perceive ourselves to be interacting with a conversational artificial agent that is intelligent. This distinct interactional form, Simon states, is significant: “ChatGPT differs [from a search engine] in two philosophically relevant regards… it integrates [different results and sources] into a coherent text…” and “its interface and functioning invites the user to communicate or interact with it by asking questions,” what Simon calls a simulation of communicative acts. The four forms of GenAI deception are: “users may be misled into believing that they interact with a human being”; they may also be misled as to “the capacities of AI,” presuming “intelligence, understanding or even consciousness”; users may be misled by content GenAI produces, especially images, video, and audio; finally, users may be misled “regarding the function of Generative AI,” thinking that the processes behind the technology are, for instance, similar to search engines when, in fact, they differ “in epistemologically highly significant ways.” Simon concludes by outlining implications for implementation, discourse, and governance, with suggestions ranging from avoiding anthropomorphic features in AI that might deceive, countering discourses of true AI agency, and establishing policy and law that mitigates deception through AI technology. Spatharioti, Sofia Eleni, David Rothschild, Daniel G. Goldstein, and Jake M. Hofman. 2025. “Effects of LLM-based Search on Decision Making: Speed, Accuracy, and Overreliance.” Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3706598.3714082. Spatharioti et al. begin by asserting the fundamental change of how we engage in search practices online, noting that “by the end of 2023 the two search engines with over 90% of global and US market share offered free LLM-based search.” The authors note the benefits and drawbacks of both traditional web searches and LLM-based searches, acknowledging the value LLM-based searches bring to internet searches in the form of synthesizing information from multiple sources and maintaining search context by retaining search history, while also admitting to risks of hallucinations and overreliance. The article entails a study of how individuals make decisions with traditional and LLM-based searches, particularly in “every day decision making.” Adapting a method from an earlier study on LLM-generated code, Spatharioti et al. employ colour coding to communicate to users the confidence level of LLM-generated outputs in the “domain of online product research.” The study performed two experiments comparing a traditional internet search (using Bing API) to LLM-based search with and without the colour-coded confidence aid. The results of the first experiment saw users complete the task in about half the time when using LLM-based search, usually accompanied by fewer and more complex queries; however, decision quality dropped for complex tasks, with “almost half of the participants in this condition making an incorrect decision for the final task” when using LLM-based search, and the majority of those using LLM-based search over relied on the tool, performing just a single query. The results of the second experiment suggest that colour-coded responses reflecting output confidence scores could be highly effective in mitigating overreliance and misinformation. Key takeaways from the study include that “if we want to encourage people to think critically about the information presented to them, we need to give them cues that help them to do so,” and very simple cues can help accomplish this. Strubell, Emma, Ananya Ganesh, and Andrew McCallum. 2019. “Energy and Policy Considerations for Deep Learning in NLP.” Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 3645-3650. https://doi.org/10.18653/v1/p19-1355. Strubell, Ganesh, and McCallum assert that the most recent improvements in neural network performance at “fundamental NLP tasks” comes at an increased cost of resources, “with the most computationally-hungry models obtaining the highest scores.” The energy, financial, and environmental costs required to train a new model are considerable; the authors of this article “characterize the dollar cost and carbon emissions that result from training the neural networks at the core of many state-of-the-art NLP models” with a view to heighten awareness among NLP researchers and advocate better practices and policy. By estimating the energy required to train the most popular NLP models and then converting that energy value into an approximated carbon and electricity cost, the authors produce ratings and values for each respective application. This allows for an (imperfect) cost-benefit analysis of such applications (e.g., another similar report estimated that an increase of just 0.1 in the English to German BLEU score for NAS cost “at least $150k in on-demand compute time and non-trivial carbon emissions.” The study’s key takeaways are that “authors should report training time and sensitivity to hyperparameters,” allowing for the direct comparison of different models so long as independent standard measurements of training time and model sensitivity are adapted; “Academic researchers need equitable access to computation resources” as industry’s monopoly stifles creativity and growth; and “Researchers should prioritize computationally efficient hardware and algorithms” to both curb costs and encourage efficient development. AI and Open AI and Scholarship