Annotated text data are central to many empirical applications in political science, yet many annotation tasks involve categories that are sparsely populated. These low-N categories pose problems for both encoder-only and decoder-only (generative) Transformer large language models (LLMs), leading to low performance, biased evaluation and measurement error in downstream analyses. This article introduces the synthetic imputation approach, a targeted method for expanding sparse categories while remaining consistent with the category-defining content of the human-coded examples. The approach uses a generative LLM to produce additional texts for underrepresented categories by conditioning on small, rotating subsets of human-coded examples. Rather than paraphrasing or replacing the original data, synthetic imputation adds informative variation only where data are thin. I evaluate the approach using both encoder-only and decoder-only LLM classifiers across multiple political science applications. The results show that synthetic imputation substantially improves performance and evaluation stability in sparse settings, recovering full-sample performance once a modest number of true observations is available, with limited and predictable overfitting at very low levels of N.

Visual Question Answering (VQA) is the task of answering questions about images, requiring the integration of multimodal input and reasoning. Modular approaches that incorporate logic-based representations into the reasoning component offer clear advantages over end-to-end trained systems, particularly in terms of interpretability. However, adapting or extending these representations when task requirements change can place a significant burden on developers. To address this challenge, we present an approach for distilling rules from Large Language Models (LLMs). Our method prompts an LLM to extend an initial VQA reasoning theory, expressed as an answer-set program, to meet new requirements of the task. Examples from VQA datasets guide the LLM, validate the results, and help correct erroneous rules by leveraging feedback from the ASP solver. We demonstrate that our approach is effective across diverse VQA datasets. Notably, only a few examples are needed to elicit correct rules from LLMs. Our experiments suggest that rule distillation from LLMs is a promising alternative to traditional data-driven rule learning approaches. Under consideration in Theory and Practice of Logic Programming.1

The European Research Council-funded project “STEMMA: Systems of Transmitting Early Modern Manuscript Verse, 1475 to 1700” computationally models the manuscript circulation of early modern poetry to gain macro-level insights into the overlapping contexts in which it was written, read, and copied. To facilitate STEMMA’s ultimate goal of modeling the circulation of manuscript poetry as a large-scale network, the project began by combining and reconciling five of the most authoritative sources of relevant data using staged large language model (LLM)-assisted pipelines. These pipelines were effective on a large proportion of the data, but they struggled in relation to particular features of early modern poetics, including imitation, translation, and unique works. This article considers how general models’ shortcomings in relation to early modern poetry can be considered a feature rather than a bug, pointing up new insights into poetics by defamiliarizing them. It considers productive intersections between literary theory and criticism and recent scholarship on LLMs and suggests that two seemingly very different fields of study address what are, at core, rather similar fundamental questions.

The integration of generative AI into humanities research presents both new opportunities and significant methodological challenges. While generative AI enables more sophisticated textual, visual, and historical analysis than ever before, it also introduces novel methodological risks. Prompts now function as measurement instruments, construct validity is entangled with model fluency, the increasing abstraction and subjectivity of tasks challenge single-answer evaluations, small changes in prompt or model configuration can materially alter substantive conclusions, and model training encodes strong theoretical assumptions prior to human interaction. This article advances a theory-first framework for generative-AI-assisted humanities research that traces how theoretical commitments intervene across the research lifecycle. Drawing on examples from the computational humanities and social sciences, it argues that theory operates as methodological infrastructure that constrains degrees of freedom, supports reproducibility, and enables principled iteration. By reframing generative AI as a site of theoretical inscription, the article outlines practical strategies for building robust and theoretically grounded humanities research in the era of generative AI.

“Conversational” technologies, products, and services are in the headlines more than ever. But what does it mean to be “conversational?” We address this question through the lens of six decades of empirical research in conversation analysis, which has identified and described the foundational structure and interactional machinery of human sociality. We consider not only how tacit notions of “conversationality” manifest in technologies, products, and services, such as role-play, communication training, and chatbots, but also in research methodologies such as focus groups, semi-structured interviews, and laboratory studies – all of which rarely acknowledge how researcher–participant interaction shapes the data collected. Drawing on a range of examples from different institutional settings, we consider how and whether such technologies can or should leverage “conversation” in ways that reproduce “naturalistic” interactions – and ask what might count as “naturalistic” in this context anyway? We argue that if human conversationality becomes a benchmark, then humans themselves will fail tests derived from normative, not empirical, understandings of how social interaction works.

Using the fields of memory studies and digital humanities, this article argues that there has been a shift from more collective and social memory to more personalised and individual memory. This shift, it is argued here, can be conceptualised through the psychoanalytic concept of ‘psychosis’. While the causes of the changes in our patterns of memory have been located in capitalist and neoliberal principles, the effects of the changes in our memory habits might be found in psychosis. From falling in love with machinic AI replicas to indulging in conspiracy theories to acting as if we are social media influencers or backing ourselves to win out in impossible job markets, we are inclined towards personal fantasy, often at the expense of participating in social life. But why do we do this? Why is it easier to believe a farfetched conspiracy theory or wild personal dream than it is to participate socially and collectively in the world we live in? Part of the reason, at least, is found in our increasing habitual reliance on new and emergent technologies. Often presented to us as a brand-new form of Artificial Intelligence, these generative tools are the latest update to a longer pattern in our digital world: the trend of developing ‘relationships’ with algorithms that, to larger and smaller degrees, we come to rely on for habits of cognition and recognition. By affecting our patterns of memory, these technologies produce a kind of isolation that lends itself to individual and fantastical – rather than shared and realist – thinking.

Extant work shows that generative AI such as GPT-3.5 and perpetuate social stereotypes and biases. A less explored source of bias is ideology: do GPT models take ideological stances on politically sensitive topics? We develop a novel approach to identify ideological bias and show that it can originate in both the training data and the filtering algorithm. Using linguistic variation across countries with contrasting political attitudes, we evaluate average GPT responses in those languages. GPT output is more conservative in languages conservative societies (polish) and more liberal in languages used in liberal ones (Swedish). These differences persist from GPT-3.5 to GPT-4. We conclude that high-quality, curated training data are essential for reducing bias.

The integration of Generative Artificial Intelligence (GAI) is reshaping traditional legal pedagogy by introducing new dimensions to research, drafting, and personalized learning. Despite its transformative potential, the application of GAI in legal education remains promising and is accompanied by such challenges as algorithmic bias, privacy concerns, and the generation of fabricated or inaccurate content—issues that continue to divide scholarly opinion on its adoption. With this in mind, the University of Kansas School of Law offered a one-credit, one-hour course on “AI for Lawyers” during the summer of 2025. Along with reproducing the syllabus of the course and some of the professor’s observations, the paper includes an overview of AI and legal pedagogy and some recommendations going forward for integrating AI into law school curricula.

Parametric modeling and generative design hold promise for architecture, yet their reliance on scripting and predefined constraints has often discouraged early-stage exploration. This paper proposes a conversational AI framework to address these challenges, integrating ChatGPT into two workflows: user-driven (Revit+Dynamo) and AI-driven (Grasshopper). By transforming natural-language prompts into Python scripts or Grasshopper definitions, designers can iterate on geometry, materials, and forms without extensive coding. AI-based visualization tools such as Veras provide near-instant feedback, accelerating the loop from concept to refinement. Rather than evaluating a single software tool, this exploration highlights collaboration between architect and AI, demonstrating how large language models can augment design intent, expand the parameter space, and adapt to contextual needs.

The proliferation of Artificial Intelligence (AI) is significantly transforming conventional legal practice. The integration of AI into legal services is still in its infancy and faces challenges such as privacy concerns, bias, and the risk of fabricated responses. This research evaluates the performance of the following AI tools: (1) ChatGPT-4, (2) Copilot, (3) DeepSeek, (4) Lexis+ AI, and (5) Llama 3. Based on their comparison, the research demonstrates that Lexis+ AI outperforms the other AI solutions. All these tools still encounter hallucinations, despite claims that utilizing the Retrieval-Augmented Generation (RAG) model has resolved this issue. The RAG system is not the driving force behind the results; it is one component of the AI architecture that influences but does not solely account for the problems associated with the AI tools. This research explores RAG architecture and its inherent complexities, offering viable solutions for improving the performance of AI-powered solutions.

Large Language Models (LLMs) raises challenges that can be examined according to a normative and an epistemological approach. The normative approach, increasingly adopted by European institutions, identifies the pros and cons of technological advancement. Regarding LLMs, the main pros concern technological innovation, economic development and the achievement of social goals and values. The disadvantages mainly concern cases of risks and harms generated by means of LLMs. The epistemological approach examines how LLMs produce outputs, information, knowledge, and a representation of reality in ways that differ from those followed by human beings. To face the impact of LLMs, our paper contends that the epistemological approach should be examined as a priority: identifying risks and opportunities of LLMs also depends on considering how this form of artificial intelligence works from an epistemological point of view. To this end, our analysis compares the epistemology of LLMs with that of law, in order to highlight at least five issues in terms of: (i) qualification; (ii) reliability; (iii) pluralism and novelty; (iv) technological dependence and (v) relation to truth and accuracy. The epistemological analysis of these issues, preliminary to the normative one, lays the foundations to better frame challenges and opportunities arising from the use of LLMs.

Generative artificial intelligence (AI) has catapulted into the legal debate through the popular applications ChatGPT, Bard, Dall-E and others. While the predominant focus has hitherto centred on issues of copyright infringement and regulatory strategies, particularly in the context of the AI Act, a critical but often overlooked issue lies in the friction between generative AI and data protection laws. The rise of these technologies highlights unresolved tension between safeguarding fundamental protection rights and and the vast, almost universal, of scale of data processing required for machine learning. Large language models, which scrape nearly the whole Internet rely on and may even generate personal data falling under the GDPR. This tension manifests across multiple dimensions, encompassing data subjects’ rights, the foundational principles of data protection and the fundamental categories of data protection. Drawing on ongoing investigations by data protection authorities in Europe, this paper undertakes a comprehensive analysis of the intricate interplay between generative AI and data protection within the European legal framework.

Retrieval-augmented generation (RAG) adds a simple but powerful feature to chatbots, the ability to upload files just-in-time. Chatbots are trained on large quantities of public data. The ability to upload files just-in-time makes it possible to reduce hallucinations by filling in gaps in the knowledge base that go beyond the public training data such as private data and recent events. For example, in a customer service scenario, with RAG, we can upload your private bill and then the bot can discuss questions about your bill as opposed to generic FAQ questions about bills in general. This tutorial will show how to upload files and generate responses to prompts; see https://github.com/kwchurch/RAG for multiple solutions based on tools from OpenAI, LangChain, HuggingFace transformers and VecML.

Since early 2021, food prices in Britain have increased by 30%. Using monthly microdata, researchers have found that frictions in the UK’s new trade relationship with the European Union (EU) play an important part in this inflation. The trade relationship is evolving, with further changes expected in 2024. This article establishes a framework for identifying trade-related inflation in close to real time. Using programming techniques, we collect daily prices of over 100,000 supermarket items, covering 80% of the UK grocery market. We identify 1,200 products from 12 countries with a protected designation of origin (PDO). This allows us to link price changes to individual EU economies. Addressing the predominance of EU PDOs, we employ a large language model to discern product origins from additional web-scraped data, thus broadening our analysis to cover over 67,000 products. Since August 2023, we find that prices for EU-originating food products have increased at a rate of 50% higher than domestically sourced products. This study presents a unique methodological approach to dissecting food sector inflation, which is well-positioned to be used in a policy setting, allowing us to assess the possible impact of impending nontariff barriers at the GB-EU border in 2024.