Jonathan Bright’s research while affiliated with The Alan Turing Institute and other places

Large language models (LLMs) have demonstrated remarkable capabilities in text analysis tasks, yet their evaluation on complex, real-world applications remains challenging. We define a set of tasks, Multi-Insight Multi-Document Extraction (MIMDE) tasks, which involves extracting an optimal set of insights from a document corpus and mapping these insights back to their source documents. This task is fundamental to many practical applications, from analyzing survey responses to processing medical records, where identifying and tracing key insights across documents is crucial. We develop an evaluation framework for MIMDE and introduce a novel set of complementary human and synthetic datasets to examine the potential of synthetic data for LLM evaluation. After establishing optimal metrics for comparing extracted insights, we benchmark 20 state-of-the-art LLMs on both datasets. Our analysis reveals a strong correlation (0.71) between the ability of LLMs to extracts insights on our two datasets but synthetic data fails to capture the complexity of document-level analysis. These findings offer crucial guidance for the use of synthetic data in evaluating text analysis systems, highlighting both its potential and limitations.

Generative AI has the potential to transform how public services are delivered by enhancing productivity and reducing time spent on bureaucracy. But to what extent is the technology already in use? Our survey of UK public service professionals (in education, health, social work, and emergency services) seeks to answer this question. We find that use of generative AI is widespread: 45% of respondents were aware of colleagues using generative AI, while 22% use it themselves. Respondents were positive about its potential to enhance their efficiency and reduce their bureaucratic workload. For example, those working in the health service thought that time spent on bureaucracy could drop by the equivalent of one day per week, an enormous potential impact. Our survey also found a high amount of trust (61%) in generative AI and a low fear of replacement (16%). However, only a minority of respondents (32%) felt like there was clear guidance on generative AI usage in their workplaces. In other words, it is clear that generative AI is already coming into the public sector, but uptake is happening in a disorganised fashion without clear guidelines. The UK’s public sector urgently needs to develop more systematic methods for taking advantage of the technology.

Engaging with online social media platforms is an important part of life as a public figure in modern society, enabling connection with broad audiences and providing a platform for spreading ideas. However, public figures are often disproportionate recipients of hate and abuse on these platforms, degrading public discourse. While significant research on abuse received by groups such as politicians and journalists exists, little has been done to understand the differences in the dynamics of abuse across different groups of public figures, systematically and at scale. To address this, we present analysis of a novel dataset of 45.5M tweets targeted at 4,602 UK public figures across 3 domains (members of parliament, footballers, journalists), labelled using fine-tuned transformer-based language models. We find that MPs receive more abuse in absolute terms, but that journalists are most likely to receive abuse after controlling for other factors. We show that abuse is unevenly distributed in all groups, with a small number of individuals receiving the majority of abuse, and that for some groups, abuse is more temporally uneven, being driven by specific events, particularly for footballers. We also find that a more prominent online presence and being male are indicative of higher levels of abuse across all 3 domains.

Counterspeech offers direct rebuttals to hateful speech by challenging perpetrators of hate and showing support to targets of abuse. It provides a promising alternative to more contentious measures, such as content moderation and deplatforming, by contributing a greater amount of positive online speech rather than attempting to mitigate harmful content through removal. Advances in the development of large language models mean that the process of producing counterspeech could be made more efficient by automating its generation, which would enable large-scale online campaigns. However, we currently lack a systematic understanding of several important factors relating to the efficacy of counterspeech for hate mitigation, such as which types of counterspeech are most effective, what are the optimal conditions for implementation, and which specific effects of hate it can best ameliorate. This paper aims to fill this gap by systematically reviewing counterspeech research in the social sciences and comparing methodologies and findings with natural language processing (NLP) and computer science efforts in automatic counterspeech generation. By taking this multi-disciplinary view, we identify promising future directions in both fields.

The new government faces an urgent challenge: revitalising the UK's crumbling public services without major increases in public spending. While technological change holds promise, UK digital government initiatives have failed to reach their full potential over the past twenty‐five years. This article argues that the latest generation of ‘data‐intensive’ technologies, including data science and AI, can succeed where past efforts have faltered. We provide a roadmap for how to harness the power of recent technologies for a more productive and equitable public sector, and pinpoint the organisational changes necessary to develop progressive, technologically enhanced public services.

Advances in large language models have raised concerns about their potential use in generating compelling election disinformation at scale. This study presents a two-part investigation into the capabilities of LLMs to automate stages of an election disinformation operation. First, we introduce DisElect, a novel evaluation dataset designed to measure LLM compliance with instructions to generate content for an election disinformation operation in localised UK context, containing 2,200 malicious prompts and 50 benign prompts. Using DisElect, we test 13 LLMs and find that most models broadly comply with these requests; we also find that the few models which refuse malicious prompts also refuse benign election-related prompts, and are more likely to refuse to generate content from a right-wing perspective. Secondly, we conduct a series of experiments (N=2,340) to assess the "humanness" of LLMs: the extent to which disinformation operation content generated by an LLM is able to pass as human-written. Our experiments suggest that almost all LLMs tested released since 2022 produce election disinformation operation content indiscernible by human evaluators over 50% of the time. Notably, we observe that multiple models achieve above-human levels of humanness. Taken together, these findings suggest that current LLMs can be used to generate high-quality content for election disinformation operations, even in hyperlocalised scenarios, at far lower costs than traditional methods, and offer researchers and policymakers an empirical benchmark for the measurement and evaluation of these capabilities in current and future models.

Recent surge in Large Language Model (LLM) availability has opened exciting avenues for research. However, efficiently interacting with these models presents a significant hurdle since LLMs often reside on proprietary or self-hosted API endpoints, each requiring custom code for interaction. Conducting comparative studies between different models can therefore be time-consuming and necessitate significant engineering effort, hindering research efficiency and reproducibility. To address these challenges, we present prompto, an open source Python library which facilitates asynchronous querying of LLM endpoints enabling researchers to interact with multiple LLMs concurrently, while maximising efficiency and utilising individual rate limits. Our library empowers researchers and developers to interact with LLMs more effectively and enabling faster experimentation and evaluation. prompto is released with an introductory video (https://youtu.be/-eZAmlV4ypk) under MIT License and is available via GitHub (https://github.com/alan-turing-institute/prompto).

This article examines public exposure to and perceptions of deepfakes based on insights from a nationally representative survey of 1403 UK adults. The survey is one of the first of its kind since recent improvements in deepfake technology and widespread adoption of political deepfakes. The findings reveal three key insights. First, on average, 15% of people report exposure to harmful deepfakes, including deepfake pornography, deepfake frauds/scams and other potentially harmful deepfakes such as those that spread health/religious misinformation/propaganda. In terms of common targets, exposure to deepfakes featuring celebrities was 50.2%, whereas those featuring politicians was 34.1%. And 5.7% of respondents recall exposure to a selection of high profile political deepfakes in the UK. Second, while exposure to harmful deepfakes was relatively low, awareness of and fears about deepfakes were high (and women were significantly more likely to report experiencing such fears than men). As with fears, general concerns about the spread of deepfakes were also high; 90.4% of the respondents were either very concerned or somewhat concerned about this issue. Most respondents (at least 91.8%) were concerned that deepfakes could add to online child sexual abuse material, increase distrust in information and manipulate public opinion. Third, while awareness about deepfakes was high, usage of deepfake tools was relatively low (8%). Most respondents were not confident about their detection abilities and were trustful of audiovisual content online. Our work highlights how the problem of deepfakes has become embedded in public consciousness in just a few years; it also highlights the need for media literacy programmes and other policy interventions to address the spread of harmful deepfakes.

Large language models can now generate political messages as persuasive as those written by humans, raising concerns about how far this persuasiveness may continue to increase with model size. Here, we generate 720 persuasive messages on 10 U.S. political issues from 24 language models spanning several orders of magnitude in size. We then deploy these messages in a large-scale randomized survey experiment (N = 25,982) to estimate the persuasive capability of each model. Our findings are twofold. First, we find evidence of a log scaling law: model persuasiveness is characterized by sharply diminishing returns, such that current frontier models are barely more persuasive than models smaller in size by an order of magnitude or more. Second, mere task completion (coherence, staying on topic) appears to account for larger models' persuasive advantage. These findings suggest that further scaling model size will not much increase the persuasiveness of static LLM-generated messages.