Digital evidence of VAWG-related behaviours

Digital behaviour is often discussed in terms of clicks, browsing, location traces, transactions, or criminal histories. Those signals matter, but a large proportion of human behaviour is encoded in language. Everyday communications capture demands, threats, apologies, excuses, minimisation, escalation, grooming, and pressure. It has been difficult to measure those patterns at scale without reducing them to keywords or sentiment scores. Newer computational methods make it possible to quantify recurring behavioural cues in language, including subtle forms of coercion and blame shifting, without attempting individual-level forensic linguistics.

A large proportion of human behaviour is encoded in language, and newer computational methods make it possible to quantify recurring behavioural cues at scale.

At herEthical AI, behavioural psychology and computational linguistics are combined to analyse the language of grooming and coercion. The aim is not to replace professional judgement, but to help practitioners see behavioural patterns that are difficult to evidence, labour-intensive to review, and emotionally costly to process, enabling trauma-informed approaches to investigations. 

Why current risk and triage systems struggle to capture lived reality

Risk assessments used across public services are attempting to operate under sustained, high-volume demand. In England and Wales, police recorded 1,350,460 domestic abuse-related incidents and crimes in the year ending March 2025. Under these conditions, assessment tools are used to grade risk quickly, prioritise safeguarding capacity, and route cases into specialist pathways.

The difficulty is, in many cases, most information about a victim’s experience is not captured in a form that can be used, reviewed, and revisited unless the case progresses to prosecution, or reaches a threshold for intervention. This is not a criticism of practitioners. It is a structural effect of triage operating in a fragmented system with limited time and uneven access to information.

From the initial interview, through evidence review, to current rule-based algorithms, complex interactions are compressed into categories that are easier to communicate and act on. That compression can mask the reality of trauma and perpetration. In coercive crimes, the “what happened” is often a longitudinal pattern, not a single incident. If the system mainly captures discrete events, perpetrators can continue harmful behaviour, while victims are repeatedly required to explain what is difficult to articulate, often under pressure.

This challenge is not confined to domestic abuse. Across crimes where coercion is a driving force, including stalking, child sexual exploitation, and coercive relationship-based fraud, the effort required to evidence psychological abuse can outstrip the system’s capacity. In the absence of tangible evidence such as video footage or medical reports, practitioners are left with narratives, fragments, and the digital traces of relationships.

The traces exist, and they should be used

Digital communication is often where coercion becomes visible. Most relationships now have a digital communications component. In relationship and grooming-based scams, for example authorised push payment fraud, the bulk of the interaction may occur through messaging platforms, email, and calls, creating a strong communications record. In post-separation abuse, communications are frequently recorded through co-parenting apps or structured messaging channels. In domestic abuse contexts, victims may have long message histories that capture control, monitoring, isolation, threats, and cycles of apology and escalation.

The question is not whether those traces exist, but whether the system can use them effectively without increasing harm. Reading thousands of messages is time-consuming and costly, but it can also be psychologically difficult. The College of Policing’s guidance on responding to trauma in policing recognises the impact of trauma exposure on staff and the need for structured support, particularly for those in high-exposure roles.

A trauma-informed approach, therefore, should include both victims and the workforce. It should reduce the burden of repeated retelling for victims and reduce the cognitive and emotional load for staff. The College of Policing has also published practice examples where trauma-informed methods at first contact improved engagement and intelligence capture, illustrating that the way information is gathered affects what can be acted on later.

Behaviour detection in communications as investigation support

We are developing an investigation support platform which can detect digitally recorded coercive behaviours and organise them as patterns, with an emphasis on evidence and explainability rather than a single score.

In practical terms, this involves recognising repeated coercive behaviours in communications, linking behaviours to original messages and incidents, supporting search and retrieval across large datasets, and connecting relevant behaviours across documents and channels, so patterns do not remain scattered.

This approach matches how coercion operates. It is rarely one sentence. It is repetition, escalation, and constraint over time. Tools that structure communications evidence as patterns can help investigators and court users see the behavioural picture quickly, while retaining the ability to reference original materials in context.

Tools that structure communications evidence as patterns can help investigators and court users see the behavioural picture quickly, while retaining the ability to reference original materials in context.

APP fraud: behavioural signals before the transaction

In authorised push payment fraud, there are digital behaviour signals in transaction data, device signals, and anomalies in banking behaviour. In relationship-based scams, however, the transaction is often preceded by weeks or months of grooming. By the time money is transferred, many victims are emotionally invested and less likely to seek help early, or to interpret their experience as criminal harm rather than personal misjudgement.

This is why behavioural analysis in communications matters. Transaction monitoring may detect a suspicious payment, but communications-based signals can flag fraudulent cases earlier. They can identify tactics such as urgency, isolation, authority mimicry, manipulation, and pressure. That creates two opportunities: earlier intervention and clearer evidencing once reported.

Once a scam is reported, the same analysis can support investigation by linking behavioural tactics to transaction timelines, witness statements, and other available evidence. It can help build a coherent account of how coercion and manipulation produced the financial outcome.

Domestic abuse: evidencing coercive control earlier, and with less burden

Domestic abuse investigations often rely on interviews, physical evidence, and incident records. Psychological manipulation is harder to spot and harder to prove. Message histories can contain critical evidence, but police rarely have the time to review thousands of messages unless there is confidence that a chargeable case can be made. This creates a structural disadvantage for coercive control, which is defined by pattern and impact rather than a single visible event.

A behaviour detection approach can help by surfacing repeated patterns of manipulation over time, allowing investigators to see frequency, escalation, and clustering. It can also quantify volume and change over time and present that information as supporting context, while retaining underlying messages for evidential review.

However, this still relies on high-quality policing at the interview stage. Trauma-informed first contact increases the chance that victims disclose relevant digital evidence and that it is captured without re-traumatisation. Evidence from policing practice suggests that trauma-informed approaches can improve engagement and intelligence capture, which is directly relevant to evidencing coercive control patterns.

Digital behaviour inside public services: victim-blaming language as an organisational risk

Digital behaviour is not only what offenders do. It is also what institutions do.
Under high workload and repeated exposure to distressing material, practitioners can develop coping responses that distance them from victims. One expression of this is victim-blaming language. Operation Soteria Bluestone has highlighted cultural and operational problems affecting rape investigations, including the need to rebalance investigations towards suspect behaviour rather than victim credibility. The CPS’s Operation Soteria summary also notes that victim-blaming attitudes were observed in responses to domestic abuse, indicating that this is not confined to one offence type.

Police forces have started to address this directly. The College of Policing has published practice examples focused on raising awareness of victim-blaming language and reducing its use, including in child exploitation contexts. Northumbria University has also delivered training to address victim-blaming language, reflecting a practical shift towards measurable cultural change in frontline work.

The problem is that victim-blaming is difficult to measure reliably at scale using manual dip-sampling. It is labour-intensive and hard to sustain when services are stretched. This is an area where automation can support cultural change without pretending to be the change. Our automated victim-blaming detection can quantify prevalence, identify common forms, and locate where issues cluster, enabling supervision, training, and quality improvement to be targeted. It can also help track progress over time, creating feedback loops that are often missing in cultural reform.

This approach also connects to workforce wellbeing. Tools that reduce the burden of reading large volumes of emotionally difficult material can reduce vicarious trauma. In parallel, Police Scotland’s trauma-informed “Act, Don’t React” approach has been associated with measurable operational effects, including lower recorded use of force and fewer officer injuries among trained officers. 

Implications for technology and service design

The argument is not that algorithms are inherently harmful. It is that automation has to match the problem it is introduced into.

If policing and partner agencies are looking to reduce burden, improve service, reduce harm, and allocate resources more effectively, then automating risk assessment as a single score is unlikely to deliver the improvement hoped for. The input data are often incomplete, fragmented, and shaped by disclosure constraints, and single numbers can be operationalised as certainty in contexts where uncertainty is the reality.

Behaviour detection tools sit differently. They can support earlier understanding, earlier evidence capture, and more consistent practice because they operate on the patterns that matter in coercive crimes. They can also reduce the emotional and cognitive burden on victims and practitioners, which is a core part of building services that people trust and return to.

In this sense, digital behaviour is not a peripheral theme. It offers a practical route to designing victim-focused, trauma-informed services that use technology where it genuinely helps: making patterns legible, evidencable, and explainable, while leaving judgement and accountability where they belong, with people.

Image of our investigation tool displaying publicly available data from the LoveFraud02 dataset (Faber, Pamela (2024), “LoveFraud02”, Mendeley Data, V1, doi: 10.17632/kmhvb4x5d8.1) linking a synthetic witness statement event to corresponding extracts in the chat data.