The Role of AI in Digital Forensics: Advancing the Fight Against Violence Against Women with Large Language Models

When a phone becomes the barrier to justice

When a woman handed her smartphone to detectives investigating her domestic abuse complaint ("Force B, Case 36”, [1]), officers promised its return within 24 hours. It took 275 days. By the time forensic analysts provided partial data, key WhatsApp messages, the victim's primary evidence of ongoing coercion, were still awaiting "deep review" [1]. That case was recorded in 2022, and since then, government and policing leaders have committed that no adult rape victim should be without a phone for more than 24 hours [2].

However, even with improvements, the scale of the challenge is stark. In the year ending March 2024, about 2.3 million adults in England and Wales experienced domestic abuse [3]; and in the year ending March 2023, 93% of domestic-abuse-related sexual-offence victims were women [4]. Digital communications saturate these cases: practitioners estimate over 90% of VAWG investigations now have a digital element [5, 6].

As a result, today’s investigations routinely involve months or years of messages across multiple apps (WhatsApp, iMessage, Instagram DMs, Snapchat, etc.). Real police and court files show how quickly volume becomes unmanageable. For example, in November 2020, Dyfed‑Powys Police reported a case in which a man was jailed after bombarding his former partner with threatening messages duplicated across SMS and WhatsApp, with messages found on the victim’s device [7]. And more recently, in August 2024, Sky News reported that Joey Essex had contacted police after receiving more than 100,000 online messages, including threats to his girlfriend, from a single alleged stalker, delivered across public and private platforms [8].

Meanwhile, HMICFRS has warned that policing is unable to keep pace with the increasing volume of digital evidence from devices which has led to considerable backlogs [9, 10]. On top of the workload problem, privacy and trust remain fragile. The government’s End-to-End Rape Review and the ICO’s 2022 opinion criticised disproportionate “digital strip-searches” of complainants [2, 11]. Finally, disclosure law adds complexity: investigators and prosecutors must pursue reasonable lines of inquiry and disclose material that might assist the defence, duties re-emphasised and updated in the Attorney General’s Guidelines (2024). Missed or late disclosure risks collapsed cases and retraumatised victims [12].

So the bottleneck is clear: mountains of messages, strict legal thresholds, and a legitimate need to minimise intrusion, all while moving quickly enough to keep victims engaged.

Can machines read first? Beyond keyword searches

What if an AI-based software could scan, sort and explain conversations at scale, surfacing what matters in hours rather than months, and doing so in a way that is auditable and privacy-conscious? With recent advances in large language models (LLMs) and Natural Language Processing (NLP), that prospect is no longer speculative. We briefly introduce one of the first pilots in UK policing where the Forensics Capability Network (FCN) and Dorset Police have partnered with the University of Warwick and the University of Leeds to actively explore AI triage to support the analysis of digital evidence and thus shrink backlogs and reduce the need to keep victims’ phones [13, 14].

Traditional approaches in policing, such as manual scrolling or keyword searches, often fall short when analysing large volumes of digital evidence. These methods can miss the subtle dynamics of abusive or controlling behaviour, and their results may vary from case to case. Meaning in VAWG-related conversations is highly contextual: a phrase that is harmless in one context may signal coercion or abuse in another. The interpretation of a message shifts depending on cultural background, family structure, personal history, and even the evolving slang or coded language used by individuals as popular culture and TV series such as ‘Adolescence’ (2025) have highlighted for the wider public [15].

To address these limitations, we designed and developed SafeSpeech, a new generation of AI tools for the analysis of textual digital evidence. Unlike traditional systems that simply flag keywords, our approach framework can quickly scan and analyse hundreds of thousands of messages in a few seconds and interpret the broader conversational context, the patterns and dynamics of abusive behaviour embedded in long, multi-platform chats.

Rethinking Context: Our Approach to NLP for VAWG

SafeSpeech is an integrated platform for analysing sexist, abusive, and coercive language at both the message and conversation level. It was built to handle what policing actually faces in VAWG cases: long, intricate chats where meaning is derived from context.

Specifically, SafeSpeech combines fine-tuned classifiers with LLMs to support three key tasks: (1) detecting harmful content at multiple granularities; (2) summarising long, multi-topic exchanges while preserving the coherence of discussions across days, weeks, or months, and focusing on harmful behaviour; and (3) explaining why the model flagged certain content, so that investigators can check, challenge, or disclose the result [17].

Fast and Fine-Grained Sexism Detection.
On the back end, SafeSpeech combines strong specialised models (such as RoBERTa and our novel ensemble “M7-FE” [18]) with recent LLMs like LLaMA, GPTs, and Gemini. On the front end, there is a lightweight browser app where analysts, and researchers in the publicly available version, can upload their own data or test existing benchmarks. Building on our recent work [20], the system also relies on data-augmentation methods to reduce sparsity in under-represented scenarios, and crucially makes operational definitions explicit. This makes it possible to identify, monitor, and potentially mitigate any bias that could otherwise remain hidden in model predictions or data labelling.

Reading conversations, not just messages.
Unlike the vast majority of existing tools in the literature, SafeSpeech does not stop at single-message labels. It first chunks long dialogues into coherent sections (so related moments are grouped even when other chatter interrupts), then runs instruction-based, toxic-aware summarisation that highlights patterns relevant to VAWG, such as escalation, monitoring, intimidation, back-handed “apologies,” etc. In addition, SafeSpeech supports speaker-conditioned summarisation, organising events by the participants involved to make roles and interactions clearer. The result is a compact, navigable storyline that points practitioners to where in the chat those behaviours occur.

Why was a flag raised?
SafeSpeech includes a soft-interpretability layer that visually highlights the phrases most influential to a model's analysis. This technique is a novel application of perplexity-gain analysis, specifically adapted to VAWG-related conversations, and gives reviewers a rapid way to assess whether the system’s focus is appropriate for the case. In the interface, these cues appear as a heatmap, accompanied by a short rationale generated by the LLM, and positioned next to the source messages for easy review [Fig. 1].

Figure 1. Heatmap for a synthetic example from SafeSpeech (public version),

Understanding People.
For research purposes only, the platform can also generate persona cues based on the Big Five traits [19], conditioned on the toxic-aware summaries. The goal is not to label individuals but to support analyst reasoning about behavioural dynamics in conversation, tracking shifts in tone, dominance, or reactivity, while clearly flagging that this is exploratory, not diagnostic.

Deployment and Future Work

The platform is currently being tested on historical cases and within safe, controlled environments. Deployment into live investigations will follow only after exhaustive technical, legal, and ethical validation. This caution is deliberate. VAWG investigations are sensitive and consequential; they demand systems that are robust, auditable, and aligned with victim-centred practice. If the standards are met, the benefits for policing, and, crucially, for the public and victims, could be significant: faster identification of relevant evidence, more focused investigative effort, and case files that present clearer, more coherent narratives to the courts.

In parallel, we are expanding collaborations with forces across the United Kingdom to specialise the system for stalking [21], and to drug-related offences, where complex, coded exchanges in messaging apps often mask supply chains and roles. The final destination is a practical, explainable AI capability that helps officers read first, intrude less, and act sooner, while strengthening the integrity of digital evidence from first report to final adjudication.

Gabriele Pergola
Assistant Professor
Department of Computer Science

---