This just-in-time session follows Google’s announcement in May 2025 of “AI Mode” and OpenAI’s “Deep Search” campaign in July 2025. While AI Overviews, introduced the year before, represented systems for combining autogenerated syntheses with elements of conventional search engines, these new developments threaten to displace traditional search completely, causing rapid and profound changes in the material infrastructure for research and knowledge creation. In conjunction with the increasing preponderance of bot-generated “AI slop” on the internet (e.g. 404 Media 2025), the situation for seeking reliable information online is increasingly precarious. While MLA discussions around generative AI so far have centered on the impact on teaching, writing, creativity, and cognition, this special session focalizes the changing landscape for research and knowledge creation. In particular, it focuses on the core limitations of LLMs with respect to “knowledge”—what historian Johan Fredrikzon (2025) describes as their fundamental “epistemological indifference”—and the implications of those limitations on longstanding processes of research and information sourcing. Although alternative search engines exist, the prevailing research ecosystem encourages students, scholars, and society at large to privilege autogenerated syntheses over links to websites. Our panelists address this urgent issue by outlining the scale of the challenge and the socio-technical and epistemic differences between conventional “search” and the new generative AI paradigm.

The panel continues the MLA’s ongoing conversations on the teaching of critical AI literacies and the burgeoning field of critical AI studies. It is in dialogue with the MLA-CCCC’s task force’s ongoing work on best practices in writing classrooms—alongside adjacent perspectives on teaching critical AI literacies (Critical AI @ Rutgers) and resistance (McIntyre, Fernandes, and Sano-Francini 2025). It likewise addresses PMLA’s Theories and Methods cluster (Kirschenbaum and Raley 2024) on the epistemic challenges of generative AI and the two Critical AI issues devoted to LLMs, generative AI, and the rise of chatbots (Goodlad and Stone 2024, 2025). “AI Mode’”s transformation of the internet as a key site of knowledge-production—in particular its displacement of conventional search—requires immediate attention from MLA members who, as skilled researchers and teachers of research practices, understand the core implications for pedagogy, information literacy, knowledge production, and democratic function.

Co-conveners Xuenan Cao and Daniella Gáti open the session by highlighting the challenges of an “AI Mode” paradigm that prioritizes synthetic content over searchable links. In the first paper, "The New Informational Order: From the Human Editor to the Machinic Agency," Nina Beguš contrasts Google’s original reliance on human curation and expertise with “AI Mode”’s privileging of statistically operationalized concepts of relevance: Beguš argues that scholars of information must become active contributors to the AI information field to move it in the direction of trust and expertise. Tiffany DeRewal’s paper, "What's in a Summary? Understanding and Interrogating How LLM-based Systems Process Sources," provides a much-needed explanation of how autogenerated summaries work—describing the lack of transparency and questionable reliability of this supposed LLM-based affordance. Such obscurity with respect to the source of knowledge and the criteria for machine summarization means that GenAI tools offer knowledge creation as the nearly frictionless product of human interactions with opaque systems. Leslie Allison’s “Stochastic Gatekeeping: How Google AI Mode Reshapes Access to Knowledge” describes how probabilistic text generating systems render the precise ways of creating knowledge unexplainable. Because LLMs work by predicting the most probable responses to user generated queries, they privilege the most common patterns. Such “stochastic gatekeeping” homogenizes content while eroding verifiability and distancing information-seekers from direct access to source material. Overall, the panel thus points to epistemic risks—a narrowing of what counts as knowledge that already correlates with pseudo-science and model collapse (Messeri and Crockett 2025; Shumailov et al 2025). The panel concludes with a response from Lauren Goodlad, the editor of Critical AI and the moderator of audience discussion.

The panel is organized to galvanize audience participation: the quick introductory provocation (7 minutes) sets up the main trajectories of inquiry, followed by three papers (15 minutes each) and a response (7 minutes) in the form of a set of questions for audience uptake and open conversation (16 minutes).

The participants have been selected to represent diversity across geography (Hong Kong, UK, US), linguistic expertise (from East Asian to Slavic), rank (from senior faculty to emerging scholars and teachers), and fields of inquiry relevant to the MLA’s growing engagement with critical AI studies (comparative literature, media studies, information literacy, algorithmic cultures and queer theory).