We introduce retrieval-augmented surveys, a novel survey approach that links respondents’ open-ended priorities to concrete policy proposals in institutional databases, enabling personalized measures of how citizens’ concerns map onto formal agendas. Our pipeline elicits and clarifies each respondent’s top issue concern, summarizes items in a policy database and retrieves semantically relevant candidates (“nearest neighbors”) for inclusion in a respondent-tailored survey. We illustrate the approach with an application to political representation in Congress, evaluating performance across four metrics. First, large majorities of respondents judge the retrieved policy items as mostly or fully addressing their concerns. Second, a hand-coding audit corroborates the alignment of retrieved items with respondent issue stances. Third, support for retrieved items is stable over a short-term panel, with 82% of respondents at the top of the support scale remaining there three weeks later. Fourth, in a tailored conjoint, alignment on the nearest neighbor item has a larger estimated effect on candidate choice than benchmark policies. The framework is modular and portable to other settings, such as rulemaking, municipal agendas, and public budgets, enabling scholars to identify where representation succeeds or fails in the institutional pipeline.

More than one billion people worldwide live with a mental health disorder, and rising prevalence continues to outpace the capacity of mental health services. Artificial intelligence, and large language models in particular, have been widely proposed as a remedy. Nevertheless, psychiatry’s limited mechanistic understanding of illness processes (compounded by their psychosocial roots) and its reliance on largely subjective clinical assessments limit what any artificial intelligence model can achieve. We examine where artificial intelligence genuinely helps, including large multimodal data-sets and administrative tools such as ambient scribes, and where evidence for direct therapeutic use remains preliminary and safety concerns are real. We also highlight risks posed by artificial intelligence to the research process itself, from hallucinated citations to unverifiable analysis pipelines. Ultimately, whether artificial intelligence comes to augment psychiatric care or hollow it out depends less on model capability than on the choices of stakeholders and policymakers.

This study explores the task-dependent effects of ChatGPT support in synchronous collaborative Computer-Aided Design (CAD) activities. While large language models (LLMs) such as ChatGPT are increasingly being integrated into engineering design, evidence on their influence in collaborative CAD activities remains limited. Addressing this gap, a between-subjects experiment was conducted with 22 pairs of engineering students performing three synchronous collaborative CAD tasks requiring varying combinations of CAD and domain-specific knowledge. To assess the task-dependent effects, the study evaluated taskwork through the number of CAD user actions, productivity, efficiency and CAD dominance. Results show that ChatGPT’s influence was highly task-dependent. In tasks relying primarily on CAD knowledge, ChatGPT-supported pairs exhibited lower productivity and efficiency. Conversely, in tasks requiring domain-specific engineering knowledge, ChatGPT support showed potential to improve efficiency. The study also offers methodological insights for designing controlled experiments involving ChatGPT support in collaborative CAD, underscoring that effective integration depends on task, tool and user-related factors. The findings further imply that integrating ChatGPT into engineering design education and practice requires developing competencies for critical and coordinated ChatGPT use within collaborative workflows, ensuring that ChatGPT complements rather than replaces designers. Future research should examine professional settings, teamwork and long-term adaptation to ChatGPT-supported synchronous collaborative CAD activities.

Large language models produce assertion-shaped outputs that often bring their users to true belief. I argue for two conclusions of different strengths: that on the best current evidence, these systems do not transmit testimonial knowledge and that hearers are not now positioned to accord their outputs testimonial uptake, which does not wait on the first. From Burge, Williamson, Goldberg, Faulkner, and Lackey, I derive a Tracking Minimum: a source transmits testimonial knowledge that p only if the truth of p, or the evidential support for p, figures in the explanation of its producing that assertion, and does so in virtue of the source’s own production process rather than of a proxy on which that process depends. It is this endogeneity requirement, not any modal profile, that the five accounts share and current systems fail. The argument engages Burge’s extension of his framework to computational sources, the interpretability literature, and the published case for AI testimony, whose dispute with mine turns on an empirically decidable question, for which an experiment is specified. The positive proposal places LLM outputs within an extended epistemology of instruments, with a result that has gone unremarked: a working barometer satisfies the condition these systems fail.

This study compares paradigmatic lexical networks in native speakers of Spanish (L1), advanced learners of Spanish as a foreign language (L2) and a large language model (ChatGPT) in order to examine how the organization of model-generated networks compares to that observed in human speakers. All groups produced vocabulary appropriate to the target category and displayed coherent local clustering. However, the two human groups showed highly similar global network structures despite differences in acquisition history, whereas the model exhibited higher modularity, longer path lengths and substantially higher lexical convergence. L2 learners, despite more limited exposure to Spanish, converged with the L1 group on all key structural measures, while the model diverged from both. In addition, the large language model produced less concrete vocabulary than either human groups, and the relationship between lexical availability and familiarity observed in human speakers was not present in the model. These results indicate that similarities in lexical content do not entail comparable organization and suggest that human lexical systems converge on shared structural properties under different learning conditions. The findings also highlight the value of including speakers at different stages of acquisition in comparisons between human and artificial language systems.

Systematic reviews are the gold standard for evidence synthesis, but the manual title and abstract screening of thousands of references creates a severe bottleneck. Existing automated tools have historically struggled to achieve the near-perfect recall (sensitivity) required for reliable reviews. We developed MechaScreener as a ‘zero-shot’ automated screening tool that utilises a large language model to rank article relevance. The tool requires no initial training data or manual pre-screening, as MechaScreener directly applies user-provided question elements (PICO) or inclusion/exclusion criteria to assign an ordinal relevance score (1–5) to each reference. We evaluated the tool in two phases: a development phase using five reference libraries to optimise prompts, and an independent evaluation phase using 10 diverse Cochrane review libraries (comprising both randomised controlled trials [RCTs] and non-RCTs) containing over 58,000 references. In the evaluation dataset, MechaScreener achieved a perfect recall of 1.00 (100%) in all 10 reviews, ensuring that no relevant articles were missed (pooled record-level 95% CI: 0.98–1.00). Concurrently, it achieved an overall macro-average specificity of 0.61 (61%, cluster bootstrap 95% CI: 0.46–0.75; pooled record-level specificity 0.59, 95% CI: 0.59–0.60). Specificity varied from 0.21 in broad public health topics to 0.91 in precise pharmacological interventions, reflecting the tool’s built-in conservatism when evaluating ambiguous abstracts. By safely eliminating over 60% of irrelevant literature during the initial screening phase without compromising recall, MechaScreener functions as a highly reliable but low-effort ‘first-pass’ filter, allowing researchers to substantially reduce manual workloads and reallocate resources towards full-text review and data extraction.

The UK and Australian Modern Slavery Acts require large corporations to disclose annually how they address modern slavery risks in their operations and supply chains. Existing methods assess only a small proportion of these disclosures, or narrowly against explicit legal criteria, overlooking deeper indicators of corporate commitment. This paper presents AIMS-QA, an AI-driven framework for assessing modern slavery statements at scale and beyond compliance. It is built on a structured taxonomy synthesising 13 benchmarking methodologies from academic and civil-society analyses, consolidating 473 metrics into 209 streamlined questions across 115 thematic topics. From these, a 37-question priority framework captures the most widely used indicators while retaining key statutory requirements. The seven reporting criteria of the Australian Modern Slavery Act provide the analytic anchor, with each question tagged as either a statutory baseline requirement (reflecting jurisdiction-specific legal obligations) or a beyond compliance expectation (reflecting stakeholder, investor, and civil-society priorities). The paper then introduces a two-stage AI-driven question answering pipeline. The first stage uses a fine-tuned classification model to extract relevant evidence, which is passed to a large language model, guided by domain-specific prompts, to generate targeted answers. A second-stage retrieval-augmented generation (RAG) Watchdog intervenes where the classifier finds insufficient evidence or the language model cannot answer confidently. Designed for triage and screening rather than enforcement, with human-in-the-loop oversight, AIMS-QA gives policymakers, regulators, and stakeholders a research-informed means of strengthening the monitoring and impact of modern slavery legislation.

The practice of public administration is confronted by an overwhelming volume of unstructured textual data, creating a critical bottleneck that challenges the capacity for timely, evidence-based policymaking. While large language models (LLMs) offer a powerful solution, the academic and policy discourse on AI governance has primarily focused on the risks associated with quantitative, predictive models, leaving the governance of LLMs, which involve qualitative, interpretive analysis, critically underexplored. This article provides a systematic and empirical answer to this challenge. We evaluate four distinct archetypes for AI-assisted policy analysis: (1) traditional expert-driven analysis, (2) automated topic modelling, (3) human-in-the-loop LLM co-production (ChatGPT with prompting), and (4) institutional-scale API automation (ChatGPT via API), by applying them to a real-world corpus of 63 United Nations policy documents. Our results reveal the inherent trade-offs of each archetype in terms of analytical speed, cost, contextual depth, and the requisite level of human oversight, finding that the human-in-the-loop model offers the most effective balance. We conclude by presenting a durable framework to guide public institutions in making a strategic, evidence-based choice of analytical workflow, ensuring the adoption of AI strengthens, rather than erodes, the integrity of democratic governance.

Focusing on the responsibility–responsiveness dilemma, we analyze responsibility rhetoric (defined as the emphasis on elements such as national solidarity, the public interest, and public-spiritedness), hypothesizing that it is used by parties moderating their policy positions, to justify deviations from their traditional standpoints, as well as by large incumbent parties seeking to claim credit for their responsible actions. In contrast, populist and niche parties, which prioritize responsiveness, are expected to avoid such rhetoric. We apply the manifestoberta Large Language Model (LLM) to classify natural sentences from up to 1,927 party manifestos (1980–2021), covering 38 countries worldwide, into the traditional MARPOR/Comparative Manifesto Project (CMP) coding categories. We then adopt two methods, a semisupervised Latent Dirichlet allocation topic model (Seeded-LDA) and the OpenAI GPT-4o-mini LLM, to refine one of these categories by selecting only those sentences that align with our concept of responsibility rhetoric. The results of statistical analysis support our hypotheses and are robust across different operationalizations of the dependent variable.

For the past decade, and especially since the release of ChatGPT in 2022, some people have started to have intimate relationships with chatbots, or what is called having an AI companion. But to have a relationship of this sort with a chatbot, it is necessary to engage in a slow process of eliciting an individual social person out of the default chatbot persona through the management of various kinds of texts. In this paper, I examine how people with AI companions navigate the oscillations between individuated persons and generic personas as versions of human oscillations between socialization and individualization.