The challenge
Young people routinely encounter violent, misogynistic and abusive material online, but the relationship between that exposure and offline violence is complex. Online content may intensify an existing dispute, normalise harmful attitudes or make an incident more public and persistent without directly causing a particular offence.
London’s Violence Reduction Unit needed a stronger, place-based evidence base showing how these processes operate, which young people face the greatest risks and where earlier intervention may be possible.
The study examined two related but distinct forms of harm: interpersonal violence affecting young people aged 12 to 24, and violence against women and girls associated with misogynistic and gender-based content.
The approach
The project used a mixed-methods design developed by M&C Saatchi World Services, nisien.ai and Shout Out UK.
nisien.ai led the quantitative strand. Over a 12-month period from May 2024 to April 2025, more than 260,000 public posts geolocated to Greater London were collected from X and analysed using automated classifiers. This produced datasets containing 25,509 confirmed misogynistic posts and 10,978 confirmed violence-related posts.
Daily patterns in this material were modelled alongside Metropolitan Police Service records and London Ambulance Service data. The analysis controlled for general platform activity, time trends and differences in the composition of recorded cases.
M&C Saatchi World Services and Shout Out UK led the qualitative strand. This included 22 interviews, eight focus groups and four detailed case studies involving young people, practitioners, teachers, safeguarding professionals and specialist violence-against-women-and-girls services.
Bringing these strands together made it possible to examine both city-wide patterns and the mechanisms through which online activity shaped individual experiences.
The findings
The research describes social media as both a mirror and a motor of harm: reflecting inequalities and conflicts already present in society while making them more visible, persistent and difficult to contain.
The quantitative analysis identified a statistically significant same-day association between the proportion of misogynistic posts and police-recorded violence against women and girls. A ten-percentage-point increase in the share of misogynistic posts was associated with an estimated 3.6% increase in recorded offences that day.
This is an association at city level, not evidence that individual posts caused individual incidents.
No equivalent statistically reliable association was found between publicly visible violence-related content and general violent crime. The qualitative evidence helped explain why: interpersonal disputes often developed through private or semi-private channels that cannot be captured through public-platform data alone.
Interviews and case studies identified pathways including sustained online hostility, audience-driven escalation and the filming and recirculation of incidents. For violence against women and girls, the study highlighted the normalisation of controlling behaviour, image-based abuse, peer sexual objectification and exploitation enabled by platform features.
The outcome
The findings informed recommendations covering safer platform design, digital youth work, peer-led intervention, education, survivor-centred responses and better support for young people facing cumulative risks.
The report also identified the first 24 to 48 hours following a highly visible online incident as a critical window for coordinated intervention.
Following publication, the Mayor of London and the VRU announced almost £5 million for a city-wide programme addressing algorithm-driven online harms. Planned work includes bystander intervention, digital-youth-work approaches, mentoring for girls and young women, positive-masculinity programmes, training for trusted adults and earlier local responses to emerging incidents.
The project extends a central strand of HateLab research: connecting online behaviour with robust offline evidence while avoiding simplistic claims of causation. It demonstrates how computational analysis, lived experience and public-health thinking can be combined to produce evidence that informs prevention and policy.