Artificial intelligence in the critical synthesis of public health responses to violence

The problem

Policy responses to violence and abuse in the UK are spread across multiple government departments, local authorities, healthcare and education providers, criminal justice agencies, voluntary organisations, and employers. Across these sectors, violence can be defined, framed, and addressed in very different ways. A policy on knife crime may sit in one department while a strategy on coercive control sits in another, each using different language, targeting different populations, and proposing different types of intervention.

This fragmentation makes it difficult for researchers and policymakers to identify where priorities overlap, where gaps exist, or where policies may be inconsistent with each other. Traditional approaches to reviewing this landscape, such as reading through documents manually or producing narrative summaries, are slow and struggle to capture patterns across sectors or track how framing changes over time. With the UK government recently publishing an updated cross-government strategy on violence against women and girls, the demands on researchers and policymakers to rapidly make sense of large and scattered bodies of policy evidence have grown.

This study aimed to describe the development and preliminary exploration of an AI-enabled tool designed to synthesise evidence from violence-related policy documents in the UK.

The solution

The research team compiled a corpus of 343 publicly available UK policy and strategy documents related to violence, drawn from government and third-sector sources. Just over half came from government bodies and the remainder from voluntary and third-sector organisations. Documents were identified through expert review, manual searches of government and organisational websites, and automated collection using web scraping tools. All documents were checked by two members of the research team to confirm relevance.

This collection was used to train an existing AI framework that combines topic modelling, clustering, and a large language model to identify patterns across documents and generate structured summaries in response to questions. The system was then made available through a web-based interface where users could type questions in plain language and receive synthesised answers drawn from the underlying documents.

Users could direct their questions at: (1) government documents only, (2) third-sector documents only, or (3) the full corpus. The interface offered different response styles depending on whether a user wanted a descriptive summary or a more analytical, interpretive output. Responses included references to the specific documents used, along with statistics on how much of the collection was drawn upon, so users could assess the robustness of what they were reading.

The tool was developed and tested between July 2025 and March 2026. 30 violence prevention researchers from disciplines including criminology, epidemiology, economics, psychology, and medicine were invited to use it, alongside eight policymakers and service providers. Participants submitted questions relevant to their own work and provided feedback on accuracy, usefulness, and practical value.

The impact

1. Stakeholders found the tool useful for navigating a fragmented policy landscape
Participants reported that the tool facilitated flexible interrogation of violence-related policy documents and supported identification of recurring framings, sectoral differences, and potential policy siloes.It provided structured starting points for analysis and helped users orient themselves within a large and scattered body of evidence, particularly in the early stages of an inquiry.

2. The tool surfaced cross-sectoral differences that are difficult to spot manually
The system identified meaningful contrasts in how different sectors frame violence. For example, third-sector documents placed greater emphasis on survivors and victim support compared to central government documents. Stakeholders described these findings as consistent with their existing understanding of the field, which gave them confidence in the tool's outputs.

3. Confidence was linked to the tool's bounded, transparent design
Stakeholders reported greater trust in this system than in general-purpose generative AI tools. They attributed this to the system drawing only on the curated collection of documents rather than generating answers from general knowledge. The inclusion of source references and coverage statistics for each response was seen as central to responsible use.

4. The tool is a proof of concept, not a finished product
The project was designed as an exploratory study rather than a formal evaluation. The research team is clear that further work is needed to assess accuracy, examine potential biases, and test the tool's impact on real-world decision-making before wider use.

Lessons learned

The tool provided a simple and easy-to-use interface for interacting with a large collection of policy documents, according to stakeholder feedback. Document management could have been improved to reduce the impact of duplicate or irrelevant documents. A more comprehensive feasibility assessment (including usability metrics) would increase confidence in the tool.