Scott A. Hale

Bertie Vidgen, Nino Scherrer, Hannah Rose Kirk et al. · oxford

The past year has seen rapid acceleration in the development of large language models (LLMs). However, without proper steering and safeguards, LLMs will readily follow malicious instructions, provide unsafe advice, and generate toxic content. We introduce SimpleSafetyTests (SST) as a new test suite for rapidly and systematically identifying such critical safety risks. The test suite comprises 100 test prompts across five harm areas that LLMs, for the vast majority of applications, should refuse to comply with. We test 11 open-access and open-source LLMs and four closed-source LLMs, and find critical safety weaknesses. While some of the models do not give a single unsafe response, most give unsafe responses to more than 20% of the prompts, with over 50% unsafe responses in the extreme. Prepending a safety-emphasising system prompt substantially reduces the occurrence of unsafe responses, but does not completely stop them from happening. Trained annotators labelled every model response to SST (n = 3,000). We use these annotations to evaluate five AI safety filters (which assess whether a models' response is unsafe given a prompt) as a way of automatically evaluating models' performance on SST. The filters' performance varies considerably. There are also differences across the five harm areas, and on the unsafe versus safe responses. The widely-used Perspective API has 72% accuracy and a newly-created zero-shot prompt to OpenAI's GPT-4 performs best with 89% accuracy. Content Warning: This paper contains prompts and responses that relate to child abuse, suicide, self-harm and eating disorders, scams and fraud, illegal items, and physical harm.

Hannah Rose Kirk, Bertie Vidgen, Paul Röttger et al. · oxford

Large language models (LLMs) are used to generate content for a wide range of tasks, and are set to reach a growing audience in coming years due to integration in product interfaces like ChatGPT or search engines like Bing. This intensifies the need to ensure that models are aligned with human preferences and do not produce unsafe, inaccurate or toxic outputs. While alignment techniques like reinforcement learning with human feedback (RLHF) and red-teaming can mitigate some safety concerns and improve model capabilities, it is unlikely that an aggregate fine-tuning process can adequately represent the full range of users' preferences and values. Different people may legitimately disagree on their preferences for language and conversational norms, as well as on values or ideologies which guide their communication. Personalising LLMs through micro-level preference learning processes may result in models that are better aligned with each user. However, there are several normative challenges in defining the bounds of a societally-acceptable and safe degree of personalisation. In this paper, we ask how, and in what ways, LLMs should be personalised. First, we review literature on current paradigms for aligning LLMs with human feedback, and identify issues including (i) a lack of clarity regarding what alignment means; (ii) a tendency of technology providers to prescribe definitions of inherently subjective preferences and values; and (iii) a 'tyranny of the crowdworker', exacerbated by a lack of documentation in who we are really aligning to. Second, we present a taxonomy of benefits and risks associated with personalised LLMs, for individuals and society at large. Finally, we propose a three-tiered policy framework that allows users to experience the benefits of personalised alignment, while restraining unsafe and undesirable LLM-behaviours within (supra-)national and organisational bounds.

Hannah Rose Kirk, Andrew M. Bean, Bertie Vidgen et al. · oxford

Human feedback is increasingly used to steer the behaviours of Large Language Models (LLMs). However, it is unclear how to collect and incorporate feedback in a way that is efficient, effective and unbiased, especially for highly subjective human preferences and values. In this paper, we survey existing approaches for learning from human feedback, drawing on 95 papers primarily from the ACL and arXiv repositories.First, we summarise the past, pre-LLM trends for integrating human feedback into language models. Second, we give an overview of present techniques and practices, as well as the motivations for using feedback; conceptual frameworks for defining values and preferences; and how feedback is collected and from whom. Finally, we encourage a better future of feedback learning in LLMs by raising five unresolved conceptual and practical challenges.

Khyati Khandelwal, Manuel Tonneau, Andrew M. Bean et al. · oxford

Large Language Models (LLMs), now used daily by millions, can encode societal biases, exposing their users to representational harms. A large body of scholarship on LLM bias exists but it predominantly adopts a Western-centric frame and attends comparatively less to bias levels and potential harms in the Global South. In this paper, we quantify stereotypical bias in popular LLMs according to an Indian-centric frame through Indian-BhED, a first of its kind dataset, containing stereotypical and anti-stereotypical examples in the context of caste and religious stereotypes in India. We find that the majority of LLMs tested have a strong propensity to output stereotypes in the Indian context, especially when compared to axes of bias traditionally studied in the Western context, such as gender and race. Notably, we find that GPT-2, GPT-2 Large, and GPT 3.5 have a particularly high propensity for preferring stereotypical outputs as a percent of all sentences for the axes of caste (63-79%) and religion (69-72%). We finally investigate potential causes for such harmful behaviour in LLMs, and posit intervention techniques to reduce both stereotypical and anti-stereotypical biases. The findings of this work highlight the need for including more diverse voices when researching fairness in AI and evaluating LLMs.

Hannah Rose Kirk, Bertie Vidgen, Paul Röttger et al. · oxford

In this paper, we address the concept of "alignment" in large language models (LLMs) through the lens of post-structuralist socio-political theory, specifically examining its parallels to empty signifiers. To establish a shared vocabulary around how abstract concepts of alignment are operationalised in empirical datasets, we propose a framework that demarcates: 1) which dimensions of model behaviour are considered important, then 2) how meanings and definitions are ascribed to these dimensions, and by whom. We situate existing empirical literature and provide guidance on deciding which paradigm to follow. Through this framework, we aim to foster a culture of transparency and critical evaluation, aiding the community in navigating the complexities of aligning LLMs with human populations.

Hannah Rose Kirk, Bertie Vidgen, Scott A. Hale · oxford

Annotating abusive language is expensive, logistically complex and creates a risk of psychological harm. However, most machine learning research has prioritized maximizing effectiveness (i.e., F1 or accuracy score) rather than data efficiency (i.e., minimizing the amount of data that is annotated). In this paper, we use simulated experiments over two datasets at varying percentages of abuse to demonstrate that transformers-based active learning is a promising approach to substantially raise efficiency whilst still maintaining high effectiveness, especially when abusive content is a smaller percentage of the dataset. This approach requires a fraction of labeled data to reach performance equivalent to training over the full dataset.

Shreyansh Padarha, Ryan Othniel Kearns, Tristan Naidoo et al.

Systematic literature reviews are essential for synthesizing scientific evidence but are costly, difficult to scale and time-intensive, creating bottlenecks for evidence-based policy. We study whether large language models can automate the complete systematic review workflow, from article retrieval, article screening, data extraction to report synthesis. Applied to epidemiological reviews of nine WHO-designated priority pathogens and validated against expert-curated ground truth, our open-source agentic pipeline (AgentSLR) achieves performance comparable to human researchers while reducing review time from approximately 7 weeks to 20 hours (a 58x speed-up). Our comparison of five frontier models reveals that performance on SLR is driven less by model size or inference cost than by each model's distinctive capabilities. Through human-in-the-loop validation, we identify key failure modes. Our results demonstrate that agentic AI can substantially accelerate scientific evidence synthesis in specialised domains.

Ashkan Kazemi, Artem Abzaliev, Naihao Deng et al.

We propose a novel system to help fact-checkers formulate search queries for known misinformation claims and effectively search across multiple social media platforms. We introduce an adaptable rewriting strategy, where editing actions for queries containing claims (e.g., swap a word with its synonym; change verb tense into present simple) are automatically learned through offline reinforcement learning. Our model uses a decision transformer to learn a sequence of editing actions that maximizes query retrieval metrics such as mean average precision. We conduct a series of experiments showing that our query rewriting system achieves a relative increase in the effectiveness of the queries of up to 42%, while producing editing action sequences that are human interpretable.

Angus R. Williams, Hannah Rose Kirk, Liam Burke et al. · oxford

Public figures receive a disproportionate amount of abuse on social media, impacting their active participation in public life. Automated systems can identify abuse at scale but labelling training data is expensive, complex and potentially harmful. So, it is desirable that systems are efficient and generalisable, handling both shared and specific aspects of online abuse. We explore the dynamics of cross-group text classification in order to understand how well classifiers trained on one domain or demographic can transfer to others, with a view to building more generalisable abuse classifiers. We fine-tune language models to classify tweets targeted at public figures across DOmains (sport and politics) and DemOgraphics (women and men) using our novel DODO dataset, containing 28,000 labelled entries, split equally across four domain-demographic pairs. We find that (i) small amounts of diverse data are hugely beneficial to generalisation and model adaptation; (ii) models transfer more easily across demographics but models trained on cross-domain data are more generalisable; (iii) some groups contribute more to generalisability than others; and (iv) dataset similarity is a signal of transferability.

Manuel Tonneau, Dylan Thurgood, Diyi Liu et al. · oxford

Online hate speech is associated with substantial social harms, yet it remains unclear how consistently platforms enforce hate speech policies or whether enforcement is feasible at scale. We address these questions through a global audit of hate speech moderation on Twitter (now X). Using a complete 24-hour snapshot of public tweets, we construct representative samples comprising 540,000 tweets annotated for hate speech by trained annotators across eight major languages. Five months after posting, 80% of hateful tweets remain online, including explicitly violent hate speech. Such tweets are no more likely to be removed than non-hateful tweets, with neither severity nor visibility increasing the likelihood of removal. We then examine whether these enforcement gaps reflect technical limits of large-scale moderation systems. While fully automated detection systems cannot reliably identify hate speech without generating large numbers of false positives, they effectively prioritize likely violations for human review. Simulations of a human-AI moderation pipeline indicate that substantially reducing user exposure to hate speech is economically feasible at a cost below existing regulatory penalties. These results suggest that the persistence of online hate cannot be explained by technical constraints alone but also reflects institutional choices in the allocation of moderation resources.

Dorian Quelle, Calvin Cheng, Alexandre Bovet et al.

Misinformation and disinformation are growing threats in the digital age, affecting people across languages and borders. However, no research has investigated the prevalence of multilingual misinformation and quantified the extent to which misinformation diffuses across languages. This paper investigates the prevalence and dynamics of multilingual misinformation through an analysis of 264,487 fact-checks spanning 95 languages. To study the evolution of claims over time and mutations across languages, we represent fact-checks with multilingual sentence embeddings and build a graph where semantically similar claims are linked. We provide quantitative evidence of repeated fact-checking efforts and establish that claims diffuse across languages. Specifically, we find that while the majority of misinformation claims are only fact-checked once, 10.26%, corresponding to more than 27,000 claims, are checked multiple times. Using fact-checks as a proxy for the spread of misinformation, we find 32.26% of repeated claims cross linguistic boundaries, suggesting that some misinformation permeates language barriers. However, spreading patterns exhibit strong assortativity, with misinformation more likely to spread within the same language or language family. Next we show that fact-checkers take more time to fact-check claims that have crossed language barriers and model the temporal and cross-lingual evolution of claims. We analyze connected components and shortest paths connecting different versions of a claim finding that claims gradually drift over time and undergo greater alteration when traversing languages. Misinformation changes over time, reducing the effectiveness of static claim matching algorithms. The findings advocate for expanded information sharing between fact-checkers globally while underscoring the importance of localized verification.

Andrew M. Bean, Ryan Othniel Kearns, Angelika Romanou et al.

Evaluating large language models (LLMs) is crucial for both assessing their capabilities and identifying safety or robustness issues prior to deployment. Reliably measuring abstract and complex phenomena such as 'safety' and 'robustness' requires strong construct validity, that is, having measures that represent what matters to the phenomenon. With a team of 29 expert reviewers, we conduct a systematic review of 445 LLM benchmarks from leading conferences in natural language processing and machine learning. Across the reviewed articles, we find patterns related to the measured phenomena, tasks, and scoring metrics which undermine the validity of the resulting claims. To address these shortcomings, we provide eight key recommendations and detailed actionable guidance to researchers and practitioners in developing LLM benchmarks.

Ahmet Kurnaz, Scott A. Hale

Polarization and echo chambers are often studied in the context of explicitly political events such as elections, and little scholarship has examined the mixing of political groups in non-political contexts. A major obstacle to studying political polarization in non-political contexts is that political leaning (i.e., left vs right orientation) is often unknown. Nonetheless, political leaning is known to correlate (sometimes quite strongly) with many lifestyle choices leading to stereotypes such as the "latte-drinking liberal." We develop a machine learning classifier to infer political leaning from non-political text and, optionally, the accounts a user follows on social media. We use Voter Advice Application results shared on Twitter as our groundtruth and train and test our classifier on a Twitter dataset comprising the 3,200 most recent tweets of each user after removing any tweets with political text. We correctly classify the political leaning of most users (F1 scores range from 0.70 to 0.85 depending on coverage). We find no relationship between the level of political activity and our classification results. We apply our classifier to a case study of news sharing in the UK and discover that, in general, the sharing of political news exhibits a distinctive left-right divide while sports news does not.

Calvin Yixiang Cheng, Dorian Quelle, Scott A. Hale

This study investigates how language mutations affect the persistent diffusion of conspiracy theories on social media. Drawing on a three-year dataset of conspiracy-related posts from X, and applying computational linguistic analysis alongside survival modelling, we find that conspiracy claims with greater semantic mutations have substantially longer lifespans. Mutations in psycholinguistic properties, including pronouns, social reference words, cognitive process terms, risk- and health- related vocabularies, are associated with extended lifespans. Mutations in actor, action and target (AAT) categories are associated with longer lifespans as well. Qualitative analysis identifies two predominant mutation patterns: simplification and assimilation, at both linguistic and AAT structural levels. Taken together, the results advance our understanding of how language mutations contribute to conspiracy persistence online and shed lights on longitudinal content moderation strategies. We argue that content moderation should consider the mutability of conspiracy claims and focus on the core claims that can address their potential variations.

Hannah Rose Kirk, Liu Leqi, Fanzhi Zeng et al.

Personalisation is a standard feature of conversational AI systems used by millions; yet, the efficacy of personalisation methods is often evaluated in academic research using simulated users rather than real people. This raises questions about how users and their simulated counterparts differ in interaction patterns and judgements, as well as whether personalisation is best achieved through context-based prompting or weight-based fine-tuning. Here, in a large-scale within-subject experiment, we re-recruit 530 participants from 52 countries two years after they gave their preferences in the PRISM dataset (Kirk et al., 2024) to evaluate personalised and non-personalised language models in blinded multi-turn conversations. We find preference fine-tuning (P-DPO, Li et al., 2024) significantly outperforms both a generic model and personalised prompting but adapting to individual preference data yields marginal gains over training on pooled preferences from a diverse population. Beyond length biases, fine-tuning amplifies sycophancy and relationship-seeking behaviours that people reward in short-term evaluations but which may introduce deleterious long-term consequences. Replicating this within-subject experiment with simulated users recovers aggregate model hierarchies but simulators perform far below human self-consistency baselines for individual judgements, discuss different topics, exhibit amplified position biases, and produce feedback dynamics that diverge from humans.

Hannah Rose Kirk, Bertram Vidgen, Paul Röttger et al.

Detecting online hate is a complex task, and low-performing models have harmful consequences when used for sensitive applications such as content moderation. Emoji-based hate is an emerging challenge for automated detection. We present HatemojiCheck, a test suite of 3,930 short-form statements that allows us to evaluate performance on hateful language expressed with emoji. Using the test suite, we expose weaknesses in existing hate detection models. To address these weaknesses, we create the HatemojiBuild dataset using a human-and-model-in-the-loop approach. Models built with these 5,912 adversarial examples perform substantially better at detecting emoji-based hate, while retaining strong performance on text-only hate. Both HatemojiCheck and HatemojiBuild are made publicly available. See our Github Repository (https://github.com/HannahKirk/Hatemoji). HatemojiCheck, HatemojiBuild, and the final Hatemoji Model are also available on HuggingFace (https://huggingface.co/datasets/HannahRoseKirk/).

Hannah Rose Kirk, Alexander Whitefield, Paul Röttger et al. · oxford

Human feedback is central to the alignment of Large Language Models (LLMs). However, open questions remain about methods (how), domains (where), people (who) and objectives (to what end) of feedback processes. To navigate these questions, we introduce PRISM, a dataset that maps the sociodemographics and stated preferences of 1,500 diverse participants from 75 countries, to their contextual preferences and fine-grained feedback in 8,011 live conversations with 21 LLMs. With PRISM, we contribute (i) wider geographic and demographic participation in feedback; (ii) census-representative samples for two countries (UK, US); and (iii) individualised ratings that link to detailed participant profiles, permitting personalisation and attribution of sample artefacts. We target subjective and multicultural perspectives on value-laden and controversial issues, where we expect interpersonal and cross-cultural disagreement. We use PRISM in three case studies to demonstrate the need for careful consideration of which humans provide what alignment data.

Bertie Vidgen, Adarsh Agrawal, Ahmed M. Ahmed et al. · deepmind, oxford

This paper introduces v0.5 of the AI Safety Benchmark, which has been created by the MLCommons AI Safety Working Group. The AI Safety Benchmark has been designed to assess the safety risks of AI systems that use chat-tuned language models. We introduce a principled approach to specifying and constructing the benchmark, which for v0.5 covers only a single use case (an adult chatting to a general-purpose assistant in English), and a limited set of personas (i.e., typical users, malicious users, and vulnerable users). We created a new taxonomy of 13 hazard categories, of which 7 have tests in the v0.5 benchmark. We plan to release version 1.0 of the AI Safety Benchmark by the end of 2024. The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, the v0.5 benchmark should not be used to assess the safety of AI systems. We have sought to fully document the limitations, flaws, and challenges of v0.5. This release of v0.5 of the AI Safety Benchmark includes (1) a principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items; (2) a taxonomy of 13 hazard categories with definitions and subcategories; (3) tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts. There are 43,090 test items in total, which we created with templates; (4) a grading system for AI systems against the benchmark; (5) an openly available platform, and downloadable tool, called ModelBench that can be used to evaluate the safety of AI systems on the benchmark; (6) an example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models; (7) a test specification for the benchmark.

Hannah Rose Kirk, Iason Gabriel, Chris Summerfield et al. · oxford

Humans strive to design safe AI systems that align with our goals and remain under our control. However, as AI capabilities advance, we face a new challenge: the emergence of deeper, more persistent relationships between humans and AI systems. We explore how increasingly capable AI agents may generate the perception of deeper relationships with users, especially as AI becomes more personalised and agentic. This shift, from transactional interaction to ongoing sustained social engagement with AI, necessitates a new focus on socioaffective alignment-how an AI system behaves within the social and psychological ecosystem co-created with its user, where preferences and perceptions evolve through mutual influence. Addressing these dynamics involves resolving key intrapersonal dilemmas, including balancing immediate versus long-term well-being, protecting autonomy, and managing AI companionship alongside the desire to preserve human social bonds. By framing these challenges through a notion of basic psychological needs, we seek AI systems that support, rather than exploit, our fundamental nature as social and emotional beings.

Manuel Tonneau, Diyi Liu, Samuel Fraiberger et al. · oxford

Perceptions of hate can vary greatly across cultural contexts. Hate speech (HS) datasets, however, have traditionally been developed by language. This hides potential cultural biases, as one language may be spoken in different countries home to different cultures. In this work, we evaluate cultural bias in HS datasets by leveraging two interrelated cultural proxies: language and geography. We conduct a systematic survey of HS datasets in eight languages and confirm past findings on their English-language bias, but also show that this bias has been steadily decreasing in the past few years. For three geographically-widespread languages -- English, Arabic and Spanish -- we then leverage geographical metadata from tweets to approximate geo-cultural contexts by pairing language and country information. We find that HS datasets for these languages exhibit a strong geo-cultural bias, largely overrepresenting a handful of countries (e.g., US and UK for English) relative to their prominence in both the broader social media population and the general population speaking these languages. Based on these findings, we formulate recommendations for the creation of future HS datasets.

Manuel Tonneau, Diyi Liu, Niyati Malhotra et al. · oxford

To address the global challenge of online hate speech, prior research has developed detection models to flag such content on social media. However, due to systematic biases in evaluation datasets, the real-world effectiveness of these models remains unclear, particularly across geographies. We introduce HateDay, the first global hate speech dataset representative of social media settings, constructed from a random sample of all tweets posted on September 21, 2022 and covering eight languages and four English-speaking countries. Using HateDay, we uncover substantial variation in the prevalence and composition of hate speech across languages and regions. We show that evaluations on academic datasets greatly overestimate real-world detection performance, which we find is very low, especially for non-European languages. Our analysis identifies key drivers of this gap, including models' difficulty to distinguish hate from offensive speech and a mismatch between the target groups emphasized in academic datasets and those most frequently targeted in real-world settings. We argue that poor model performance makes public models ill-suited for automatic hate speech moderation and find that high moderation rates are only achievable with substantial human oversight. Our results underscore the need to evaluate detection systems on data that reflects the complexity and diversity of real-world social media.

Chiman Salavati, Shannon Song, Willmar Sosa Diaz et al.

Biased information (recently termed bisinformation) continues to be taught in medical curricula, often long after having been debunked. In this paper, we introduce BRICC, a firstin-class initiative that seeks to mitigate medical bisinformation using machine learning to systematically identify and flag text with potential biases, for subsequent review in an expert-in-the-loop fashion, thus greatly accelerating an otherwise labor-intensive process. A gold-standard BRICC dataset was developed throughout several years, and contains over 12K pages of instructional materials. Medical experts meticulously annotated these documents for bias according to comprehensive coding guidelines, emphasizing gender, sex, age, geography, ethnicity, and race. Using this labeled dataset, we trained, validated, and tested medical bias classifiers. We test three classifier approaches: a binary type-specific classifier, a general bias classifier; an ensemble combining bias type-specific classifiers independently-trained; and a multitask learning (MTL) model tasked with predicting both general and type-specific biases. While MTL led to some improvement on race bias detection in terms of F1-score, it did not outperform binary classifiers trained specifically on each task. On general bias detection, the binary classifier achieves up to 0.923 of AUC, a 27.8% improvement over the baseline. This work lays the foundations for debiasing medical curricula by exploring a novel dataset and evaluating different training model strategies. Hence, it offers new pathways for more nuanced and effective mitigation of bisinformation.

Michael Shliselberg, Ashkan Kazemi, Scott A. Hale et al.

Diaspora communities are disproportionately impacted by off-the-radar misinformation and often neglected by mainstream fact-checking efforts, creating a critical need to scale-up efforts of nascent fact-checking initiatives. In this paper we present SynDy, a framework for Synthetic Dynamic Dataset Generation to leverage the capabilities of the largest frontier Large Language Models (LLMs) to train local, specialized language models. To the best of our knowledge, SynDy is the first paper utilizing LLMs to create fine-grained synthetic labels for tasks of direct relevance to misinformation mitigation, namely Claim Matching, Topical Clustering, and Claim Relationship Classification. SynDy utilizes LLMs and social media queries to automatically generate distantly-supervised, topically-focused datasets with synthetic labels on these three tasks, providing essential tools to scale up human-led fact-checking at a fraction of the cost of human-annotated data. Training on SynDy's generated labels shows improvement over a standard baseline and is not significantly worse compared to training on human labels (which may be infeasible to acquire). SynDy is being integrated into Meedan's chatbot tiplines that are used by over 50 organizations, serve over 230K users annually, and automatically distribute human-written fact-checks via messaging apps such as WhatsApp. SynDy will also be integrated into our deployed Co-Insights toolkit, enabling low-resource organizations to launch tiplines for their communities. Finally, we envision SynDy enabling additional fact-checking tools such as matching new misinformation claims to high-quality explainers on common misinformation topics.

Vahid Ghafouri, Robert McNeil, Teodor Yankov et al. · oxford

We present a large-scale computational analysis of migration-related discourse in UK parliamentary debates spanning over 75 years and compare it with US congressional discourse. Using open-weight LLMs, we annotate each statement with high-level stances toward migrants and track the net tone toward migrants across time and political parties. For the UK, we extend this with a semi-automated framework for extracting fine-grained narrative frames to capture nuances of migration discourse. Our findings show that, while US discourse has grown increasingly polarised, UK parliamentary attitudes remain relatively aligned across parties, with a persistent ideological gap between Labour and the Conservatives, reaching its most negative level in 2025. The analysis of narrative frames in the UK parliamentary statements reveals a shift toward securitised narratives such as border control and illegal immigration, while longer-term integration-oriented frames such as social integration have declined. Moreover, discussions of national law about immigration have been replaced over time by international law and human rights, revealing nuances in discourse trends. Taken together broadly, our findings demonstrate how LLMs can support scalable, fine-grained discourse analysis in political and historical contexts.

Chiman Salavati, Shannon Song, Scott A. Hale et al.

The use of inappropriate language -- such as outdated, exclusionary, or non-patient-centered terms -- medical instructional materials can significantly influence clinical training, patient interactions, and health outcomes. Despite their reputability, many materials developed over past decades contain examples now considered inappropriate by current medical standards. Given the volume of curricular content, manually identifying instances of inappropriate use of language (IUL) and its subcategories for systematic review is prohibitively costly and impractical. To address this challenge, we conduct a first-in-class evaluation of small language models (SLMs) fine-tuned on labeled data and pre-trained LLMs with in-context learning on a dataset containing approximately 500 documents and over 12,000 pages. For SLMs, we consider: (1) a general IUL classifier, (2) subcategory-specific binary classifiers, (3) a multilabel classifier, and (4) a two-stage hierarchical pipeline for general IUL detection followed by multilabel classification. For LLMs, we consider variations of prompts that include subcategory definitions and/or shots. We found that both LLama-3 8B and 70B, even with carefully curated shots, are largely outperformed by SLMs. While the multilabel classifier performs best on annotated data, supplementing training with unflagged excerpts as negative examples boosts the specific classifiers' AUC by up to 25%, making them most effective models for mitigating harmful language in medical curricula.

Andrew M. Bean, Simi Hellsten, Harry Mayne et al.

In this paper, we present the LingOly benchmark, a novel benchmark for advanced reasoning abilities in large language models. Using challenging Linguistic Olympiad puzzles, we evaluate (i) capabilities for in-context identification and generalisation of linguistic patterns in very low-resource or extinct languages, and (ii) abilities to follow complex task instructions. The LingOly benchmark covers more than 90 mostly low-resource languages, minimising issues of data contamination, and contains 1,133 problems across 6 formats and 5 levels of human difficulty. We assess performance with both direct accuracy and comparison to a no-context baseline to penalise memorisation. Scores from 11 state-of-the-art LLMs demonstrate the benchmark to be challenging, and models perform poorly on the higher difficulty problems. On harder problems, even the top model only achieved 38.7% accuracy, a 24.7% improvement over the no-context baseline. Large closed models typically outperform open models, and in general, the higher resource the language, the better the scores. These results indicate, in absence of memorisation, true multi-step out-of-domain reasoning remains a challenge for current language models.

Adriano Belisario, Scott A. Hale, Luc Rocher

Gun violence is a pressing human rights issue that affects nearly every dimension of the social fabric, from healthcare and education to psychology and the economy. Reliable data on firearm events is paramount to developing more effective public policy and emergency responses. However, the lack of comprehensive databases and the risks of in-person surveys prevent human rights organizations from collecting needed data in most countries. Here, we partner with a Brazilian human rights organization to conduct a systematic evaluation of language models to assist with monitoring real-world firearm events from social media data. We propose a fine-tuned BERT-based model trained on Twitter (now X) texts to distinguish gun violence reports from ordinary Portuguese texts. We then incorporate our model into a web application and test it in a live intervention. We study and interview Brazilian analysts who continuously check social media texts to identify new gun violence events. Qualitative assessments show that our solution helped all analysts use their time more efficiently and expanded their search capacities. Quantitative assessments show that the use of our model was associated with analysts having further interactions with online users reporting gun violence. Our findings suggest that human-centered interventions using language models can help support the work of human rights organizations.

Ashkan Kazemi, Zehua Li, Verónica Pérez-Rosas et al.

An important challenge for news fact-checking is the effective dissemination of existing fact-checks. This in turn brings the need for reliable methods to detect previously fact-checked claims. In this paper, we focus on automatically finding existing fact-checks for claims made in social media posts (tweets). We conduct both classification and retrieval experiments, in monolingual (English only), multilingual (Spanish, Portuguese), and cross-lingual (Hindi-English) settings using multilingual transformer models such as XLM-RoBERTa and multilingual embeddings such as LaBSE and SBERT. We present promising results for "match" classification (86% average accuracy) in four language pairs. We also find that a BM25 baseline outperforms or is on par with state-of-the-art multilingual embedding models for the retrieval task during our monolingual experiments. We highlight and discuss NLP challenges while addressing this problem in different languages, and we introduce a novel curated dataset of fact-checks and corresponding tweets for future research.

Shiri Dori-Hacohen, Roberto Montenegro, Fabricio Murai et al.

Most Fairness in AI research focuses on exposing biases in AI systems. A broader lens on fairness reveals that AI can serve a greater aspiration: rooting out societal inequities from their source. Specifically, we focus on inequities in health information, and aim to reduce bias in that domain using AI. The AI algorithms under the hood of search engines and social media, many of which are based on recommender systems, have an outsized impact on the quality of medical and health information online. Therefore, embedding bias detection and reduction into these recommender systems serving up medical and health content online could have an outsized positive impact on patient outcomes and wellbeing. In this position paper, we offer the following contributions: (1) we propose a novel framework of Fairness via AI, inspired by insights from medical education, sociology and antiracism; (2) we define a new term, bisinformation, which is related to, but distinct from, misinformation, and encourage researchers to study it; (3) we propose using AI to study, detect and mitigate biased, harmful, and/or false health information that disproportionately hurts minority groups in society; and (4) we suggest several pillars and pose several open problems in order to seed inquiry in this new space. While part (3) of this work specifically focuses on the health domain, the fundamental computer science advances and contributions stemming from research efforts in bias reduction and Fairness via AI have broad implications in all areas of society.

Austin Botelho, Bertie Vidgen, Scott A. Hale

Accurate detection and classification of online hate is a difficult task. Implicit hate is particularly challenging as such content tends to have unusual syntax, polysemic words, and fewer markers of prejudice (e.g., slurs). This problem is heightened with multimodal content, such as memes (combinations of text and images), as they are often harder to decipher than unimodal content (e.g., text alone). This paper evaluates the role of semantic and multimodal context for detecting implicit and explicit hate. We show that both text- and visual- enrichment improves model performance, with the multimodal model (0.771) outperforming other models' F1 scores (0.544, 0.737, and 0.754). While the unimodal-text context-aware (transformer) model was the most accurate on the subtask of implicit hate detection, the multimodal model outperformed it overall because of a lower propensity towards false positives. We find that all models perform better on content with full annotator agreement and that multimodal models are best at classifying the content where annotators disagree. To conduct these investigations, we undertook high-quality annotation of a sample of 5,000 multimodal entries. Tweets were annotated for primary category, modality, and strategy. We make this corpus, along with the codebook, code, and final model, freely available.

Ashkan Kazemi, Kiran Garimella, Gautam Kishore Shahi et al.

There is currently no easy way to fact-check content on WhatsApp and other end-to-end encrypted platforms at scale. In this paper, we analyze the usefulness of a crowd-sourced "tipline" through which users can submit content ("tips") that they want fact-checked. We compare the tips sent to a WhatsApp tipline run during the 2019 Indian national elections with the messages circulating in large, public groups on WhatsApp and other social media platforms during the same period. We find that tiplines are a very useful lens into WhatsApp conversations: a significant fraction of messages and images sent to the tipline match with the content being shared on public WhatsApp groups and other social media. Our analysis also shows that tiplines cover the most popular content well, and a majority of such content is often shared to the tipline before appearing in large, public WhatsApp groups. Overall, our findings suggest tiplines can be an effective source for discovering content to fact-check.

Ashkan Kazemi, Kiran Garimella, Devin Gaffney et al.

Manual fact-checking does not scale well to serve the needs of the internet. This issue is further compounded in non-English contexts. In this paper, we discuss claim matching as a possible solution to scale fact-checking. We define claim matching as the task of identifying pairs of textual messages containing claims that can be served with one fact-check. We construct a novel dataset of WhatsApp tipline and public group messages alongside fact-checked claims that are first annotated for containing "claim-like statements" and then matched with potentially similar items and annotated for claim matching. Our dataset contains content in high-resource (English, Hindi) and lower-resource (Bengali, Malayalam, Tamil) languages. We train our own embedding model using knowledge distillation and a high-quality "teacher" model in order to address the imbalance in embedding quality between the low- and high-resource languages in our dataset. We provide evaluations on the performance of our solution and compare with baselines and existing state-of-the-art multilingual embedding models, namely LASER and LaBSE. We demonstrate that our performance exceeds LASER and LaBSE in all settings. We release our annotated datasets, codebooks, and trained embedding model to allow for further research.

Alexander Robertson, Farhana Ferdousi Liza, Dong Nguyen et al.

The semantics of emoji has, to date, been considered from a static perspective. We offer the first longitudinal study of how emoji semantics changes over time, applying techniques from computational linguistics to six years of Twitter data. We identify five patterns in emoji semantic development and find evidence that the less abstract an emoji is, the more likely it is to undergo semantic change. In addition, we analyse select emoji in more detail, examining the effect of seasonality and world events on emoji semantics. To aid future work on emoji and semantics, we make our data publicly available along with a web-based interface that anyone can use to explore semantic change in emoji.

Zo Ahmed, Bertie Vidgen, Scott A. Hale

Online hate is a growing concern on many social media platforms and other sites. To combat it, technology companies are increasingly identifying and sanctioning `hateful users' rather than simply moderating hateful content. Yet, most research in online hate detection to date has focused on hateful content. This paper examines how fairer and more accurate hateful user detection systems can be developed by incorporating social network information through geometric deep learning. Geometric deep learning dynamically learns information-rich network representations and can generalise to unseen nodes. This is essential for moving beyond manually engineered network features, which lack scalability and produce information-sparse network representations. This paper compares the accuracy of geometric deep learning with other techniques which either exclude network information or incorporate it through manual feature engineering (e.g., node2vec). It also evaluates the fairness of these techniques using the `predictive equality' criteria, comparing the false positive rates on a subset of 136 African-American users with 4836 other users. Geometric deep learning produces the most accurate and fairest classifier, with an AUC score of 90.8\% on the entire dataset and a false positive rate of zero among the African-American subset for the best performing model. This highlights the benefits of more effectively incorporating social network features in automated hateful user detection. Such an approach is also easily operationalized for real-world content moderation as it has an efficient and scalable design.

Zijian Wang, Scott A. Hale, David Adelani et al.

Social media provide access to behavioural data at an unprecedented scale and granularity. However, using these data to understand phenomena in a broader population is difficult due to their non-representativeness and the bias of statistical inference tools towards dominant languages and groups. While demographic attribute inference could be used to mitigate such bias, current techniques are almost entirely monolingual and fail to work in a global environment. We address these challenges by combining multilingual demographic inference with post-stratification to create a more representative population sample. To learn demographic attributes, we create a new multimodal deep neural architecture for joint classification of age, gender, and organization-status of social media users that operates in 32 languages. This method substantially outperforms current state of the art while also reducing algorithmic bias. To correct for sampling biases, we propose fully interpretable multilevel regression methods that estimate inclusion probabilities from inferred joint population counts and ground-truth population counts. In a large experiment over multilingual heterogeneous European regions, we show that our demographic inference and bias correction together allow for more accurate estimates of populations and make a significant step towards representative social sensing in downstream applications with multilingual social media.

Chico Q. Camargo, Scott A. Hale, Peter John et al.

Recent election surprises, regime changes, and political shocks indicate that political agendas have become more fast-moving and volatile. The ability to measure the complex dynamics of agenda change and capture the nature and extent of volatility in political systems is therefore more crucial than ever before. This study proposes a definition and operationalization of volatility that combines insights from political science, communications, information theory, and computational techniques. The proposed measures of fractionalization and agenda change encompass the shifting salience of issues in the agenda as a whole and allow the study of agendas across different domains. We evaluate these metrics and compare them to other measures such as issue-level survival rates and the Pedersen Index, which uses public-opinion poll data to measure public agendas, as well as traditional media content to measure media agendas in the UK and Germany. We show how these measures complement existing approaches and could be employed in future agenda-setting research.

Scott A. Hale, Irene Eleta

The number and quality of user reviews greatly affects consumer purchasing decisions. While reviews in all languages are increasing, it is still often the case (especially for non-English speakers) that there are only a few reviews in a person's first language. Using an online experiment, we examine the value that potential purchasers receive from interfaces showing additional reviews in a second language. The results paint a complicated picture with both positive and negative reactions to the inclusion of foreign-language reviews. Roughly 26-28% of subjects clicked to see translations of the foreign-language content when given the opportunity, and those who did so were more likely to select the product with foreign-language reviews than those who did not.

Scott A. Hale

The number of user reviews of tourist attractions, restaurants, mobile apps, etc. is increasing for all languages; yet, research is lacking on how reviews in multiple languages should be aggregated and displayed. Speakers of different languages may have consistently different experiences, e.g., different information available in different languages at tourist attractions or different user experiences with software due to internationalization/localization choices. This paper assesses the similarity in the ratings given by speakers of different languages to London tourist attractions on TripAdvisor. The correlations between different languages are generally high, but some language pairs are more correlated than others. The results question the common practice of computing average ratings from reviews in many languages.

Suin Kim, Sungjoon Park, Scott A. Hale et al.

Multilingualism is common offline, but we have a more limited understanding of the ways multilingualism is displayed online and the roles that multilinguals play in the spread of content between speakers of different languages. We take a computational approach to studying multilingualism using one of the largest user-generated content platforms, Wikipedia. We study multilingualism by collecting and analyzing a large dataset of the content written by multilingual editors of the English, German, and Spanish editions of Wikipedia. This dataset contains over two million paragraphs edited by over 15,000 multilingual users from July 8 to August 9, 2013. We analyze these multilingual editors in terms of their engagement, interests, and language proficiency in their primary and non-primary (secondary) languages and find that the English edition of Wikipedia displays different dynamics from the Spanish and German editions. Users primarily editing the Spanish and German editions make more complex edits than users who edit these editions as a second language. In contrast, users editing the English edition as a second language make edits that are just as complex as the edits by users who primarily edit the English edition. In this way, English serves a special role bringing together content written by multilinguals from many language editions. Nonetheless, language remains a formidable hurdle to the spread of content: we find evidence for a complexity barrier whereby editors are less likely to edit complex content in a second language. In addition, we find that multilinguals are less engaged and show lower levels of language proficiency in their second languages. We also examine the topical interests of multilingual editors and find that there is no significant difference between primary and non-primary editors in each language.

Han-Teng Liao, King-wa Fu, Scott A. Hale

This paper presents a multilingual study on, per single post of microblog text, (a) how much can be said, (b) how much is written in terms of characters and bytes, and (c) how much is said in terms of information content in posts by different organizations in different languages. Focusing on three different languages (English, Chinese, and Japanese), this research analyses Weibo and Twitter accounts of major embassies and news agencies. We first establish our criterion for quantifying "how much can be said" in a digital text based on the openly available Universal Declaration of Human Rights and the translated subtitles from TED talks. These parallel corpora allow us to determine the number of characters and bits needed to represent the same content in different languages and character encodings. We then derive the amount of information that is actually contained in microblog posts authored by selected accounts on Weibo and Twitter. Our results confirm that languages with larger character sets such as Chinese and Japanese contain more information per character than English, but the actual information content contained within a microblog text varies depending on both the type of organization and the language of the post. We conclude with a discussion on the design implications of microblog text limits for different languages.

Scott A. Hale

This article analyzes users who edit Wikipedia articles about Okinawa, Japan, in English and Japanese. It finds these users are among the most active and dedicated users in their primary languages, where they make many large, high-quality edits. However, when these users edit in their non-primary languages, they tend to make edits of a different type that are overall smaller in size and more often restricted to the narrow set of articles that exist in both languages. Design changes to motivate wider contributions from users in their non-primary languages and to encourage multilingual users to transfer more information across language divides are presented.

Scott A. Hale

This article analyzes one month of edits to Wikipedia in order to examine the role of users editing multiple language editions (referred to as multilingual users). Such multilingual users may serve an important function in diffusing information across different language editions of the encyclopedia, and prior work has suggested this could reduce the level of self-focus bias in each edition. This study finds multilingual users are much more active than their single-edition (monolingual) counterparts. They are found in all language editions, but smaller-sized editions with fewer users have a higher percentage of multilingual users than larger-sized editions. About a quarter of multilingual users always edit the same articles in multiple languages, while just over 40% of multilingual users edit different articles in different languages. When non-English users do edit a second language edition, that edition is most frequently English. Nonetheless, several regional and linguistic cross-editing patterns are also present.

Taha Yasseri, Scott A. Hale, Helen Margetts

Contemporary collective action, much of which involves social media and other Internet-based platforms, leaves a digital imprint which may be harvested to better understand the dynamics of mobilization. Petition signing is an example of collective action which has gained in popularity with rising use of social media and provides such data for the whole population of petition signatories for a given platform. This paper tracks the growth curves of all 20,000 petitions to the UK government petitions website (http://epetitions.direct.gov.uk) and 1,800 petitions to the US White House site (https://petitions.whitehouse.gov), analyzing the rate of growth and outreach mechanism. Previous research has suggested the importance of the first day to the ultimate success of a petition, but has not examined early growth within that day, made possible here through hourly resolution in the data. The analysis shows that the vast majority of petitions do not achieve any measure of success; over 99 percent fail to get the 10,000 signatures required for an official response and only 0.1 percent attain the 100,000 required for a parliamentary debate (0.7 percent in the US). We analyze the data through a multiplicative process model framework to explain the heterogeneous growth of signatures at the population level. We define and measure an average outreach factor for petitions and show that it decays very fast (reducing to 0.1 pervent after 10 hours in the UK and 30 hours in the US). After a day or two, a petition's fate is virtually set. The findings challenge conventional analyses of collective action from economics and political science, where the production function has been assumed to follow an S-shaped curve.

Helen Z. Margetts, Peter John, Scott A. Hale et al.

The Internet has been ascribed a prominent role in collective action, particularly with widespread use of social media. But most mobilisations fail. We investigate the characteristics of those few mobilisations that succeed and hypothesise that the presence of 'starters' with low thresholds for joining will determine whether a mobilisation achieves success, as suggested by threshold models. We use experimental data from public good games to identify personality types associated with willingness to start in collective action. We find a significant association between both extraversion and internal locus of control, and willingness to start, while agreeableness is associated with a tendency to follow. Rounds without at least a minimum level of extraversion among the participants are unlikely to be funded, providing some support for the hypothesis.

Scott A. Hale

At least two software packages---DARWIN, Eckerd College, and FinScan, Texas A&M---exist to facilitate the identification of cetaceans---whales, dolphins, porpoises---based upon the naturally occurring features along the edges of their dorsal fins. Such identification is useful for biological studies of population, social interaction, migration, etc. The process whereby fin outlines are extracted in current fin-recognition software packages is manually intensive and represents a major user input bottleneck: it is both time consuming and visually fatiguing. This research aims to develop automated methods (employing unsupervised thresholding and morphological processing techniques) to extract cetacean dorsal fin outlines from digital photographs thereby reducing manual user input. Ideally, automatic outline generation will improve the overall user experience and improve the ability of the software to correctly identify cetaceans. Various transformations from color to gray space were examined to determine which produced a grayscale image in which a suitable threshold could be easily identified. To assist with unsupervised thresholding, a new metric was developed to evaluate the jaggedness of figures ("pixelarity") in an image after thresholding. The metric indicates how cleanly a threshold segments background and foreground elements and hence provides a good measure of the quality of a given threshold. This research results in successful extractions in roughly 93% of images, and significantly reduces user-input time.