Unbowed - Tracking technology-facilitated gender-based violence

Law & Policy Platforms & AI Courts Research Regions About us

Category: Research

  • Study shows promise in protecting students from online harm

    Study shows promise in protecting students from online harm

    Most university students surveyed in a pilot study in Myanmar had experienced some form of online harm in the previous year, and a workshop designed to help them recognise and respond to it was associated with less shame and distress and a greater willingness to seek help.

    The pilot was run by Stop Online Harm, a survivor-led group, with researchers from University College London, and the preliminary results were presented at a webinar last week. It examined whether the group’s Hummingbird Workshop changes how students recognise and respond to online harm and specifically technology-facilitated gender-based violence.

    “TFGBV is bad — but there is something schools can do about it,” said Saijai Liangpunsakul, the group’s founder.

    The group describes the findings as promising descriptive results drawn from comparing students at the start and end of the study, rather than the output of a full randomised controlled trial with a counterfactual, and says the next step is to verify them with one.

    “We don’t yet have a lot of RCT evidence on what works to combat online harm, and in particular, technology-facilitated gender-based violence,” said Alexandra Hartman, a professor at UCL. “And that’s why we think that this pilot — and then future work we hope to do — is really important.”

    What the students reported

    At baseline, 62 per cent of participants said they had experienced online harm in the previous year. Stop Online Harm said 27 per cent of all participants reported technology-facilitated gender-based violence specifically, including image-based abuse, sextortion, threats of violence and death threats. Much of it came from unknown or anonymous accounts.

    Being online is central to students’ lives, said Gloria Gennaro, an assistant professor at UCL. “They use it for studying, for work, for social connections, for activism, for all their possible passions in life.”

    Gender shaped the experience. Men and women reported similar overall levels of online harm, but women reported higher rates of technology-facilitated gender-based violence, higher levels of distress and shame when it happened, and were more likely to attribute what happened to them to their gender or sexual identity.

    The workshop

    The Hummingbird Workshop has two parts. The first is a six-episode animated series, each episode around ten minutes, following a character intended to be relatable, and designed to counter myths, reduce shame and victim-blaming, and set out how to get help. The group said the series was built to be engaging and locally relevant down to the characters’ names, clothes and language.

    The second is a small-group session led by a facilitator trained in trauma-informed care, in which students discuss the videos, relate them to their own lives and ask questions.

    How the study was run

    Students were randomly assigned either to early access to the videos or to early access to the videos and the session, and were surveyed at the start, middle and end. All had access to both parts by the end, so that no participant missed the intervention.

    The study passed a university ethics review. Stop Online Harm said it collected only necessary data, asked sensitive questions with care and kept its own support service, the Online Ambulance, available to every student throughout.

    “It has to be part of the whole research process, starting from the beginning to the end,” said Hetty Phu, the group’s programme lead, of the approach to ethics.

    Comparing the start and end of the study, participation was correlated with more help-seeking and with falls in loneliness, victim-blaming attitudes, shame and distress, the group said.

    “The Hummingbird videos and workshop are correlated with several outcomes, including improving well-being and declining levels of shame,” said Sara Merner, a senior policy associate at J-PAL Europe.

    Stop Online Harm said it plans to raise funds for a full randomised controlled trial, extend the work to neighbouring countries and set up a student ambassador programme.

  • Tech abuse linked to partner killings: The UK evidence

    Tech abuse linked to partner killings: The UK evidence

    Technology-facilitated abuse featured in two in five intimate partner homicides in the United Kingdom, according to an analysis of statutory death reviews published by The Observer on Tuesday.

    The newspaper examined every report on an intimate partner killing in the government’s Domestic Homicide Review Library up to April. The library records 647 deaths between 2004 and 2022, of which intimate partner cases account for 75 per cent, or 487 reports. In 85 per cent of those the victim was a woman and the perpetrator a man.

    Domestic Homicide Reviews are statutory multi-agency reviews, carried out when the death of a person over 16 is thought to have resulted from domestic abuse. They are conducted with the participation of the agencies involved, which makes them a different kind of record from survey data or press reports.

    Nicole Jacobs, the domestic abuse commissioner, was quoted as saying the analysis showed the “devastating consequences” of technology abuse and called for stronger responses from police and public services. She said she hears “daily” from survivors who report it and are told by police to “simply stop using it, change the password, or block their perpetrator”.

    “As technology evolves, so do abusers’ tactics. Right now our current risk indicator tools, used by the police and other public services to safeguard victims, are not keeping up with the danger of this abuse.”

    What was found

    The Observer identified four methods of technology-facilitated abuse, and a fifth category in which perpetrators restricted a victim’s access to technology. It reports that many cases featured more than one form. Its findings, in its own terms:

    • Harassment — using technology to bombard victims with communication, send threatening messages, and distribute intimate images without consent — was “the most common category, found in 58% of cases”
    • Monitoring — overtly controlling a victim’s use of devices, or checking their messages — “featured in 40% of all cases”
    • Surveillance — covert use of technology to track or record — “featured in a third”
    • Deception — creating fake accounts, or falsifying messages or evidence — “was present in 15% of cases”
    • Restricting access to technology — destroying devices, changing passwords to lock victims out of accounts, preventing them calling for help — in “44% of cases”

    In five cases, the analysis found technology was used to facilitate or conceal the killing itself: perpetrators falsified messages from victims after their deaths, or used surveillance technology to locate them beforehand.

    Not recorded as a risk

    The finding the analysis returns to is what happened when victims reported it.

    In the case of Alice Ruggles, murdered in Gateshead in 2016 after a campaign of stalking that included her social media being hacked, police advised her to block the perpetrator, turn off her phone and delete her social media accounts. No technology-facilitated abuse was taken into account in the risk assessments carried out into her safety.

    In a Dorset case from 2021, in which the review gave the victim and perpetrator the pseudonyms Daisy and Robert, there was no documented physical violence between them. She reported his messages to police one lunchtime and was asked to confirm that she had told him to stop and had blocked his number. Her report was placed in a queue. No officer had attended when, 23 hours later, he entered her home and killed her.

    What the specialists said

    Emma Pickering, head of technology-facilitated abuse and economic empowerment at the charity Refuge, said the findings were likely “the tip of the iceberg”, and that officers “can be slow or reluctant to obtain evidence directly from perpetrators’ devices”. Police forces, she said, need to be resourced “to investigate tech-facilitated crime and secure digital evidence in a timely and effective manner”.

    Refuge research in 2021 found that one in three women in the UK had experienced online abuse, one in six of them at the hands of a partner or former partner — which the charity puts at almost two million women. Its own client figures show a 258 per cent rise in technology-facilitated abuse between 2018 and 2022. Pickering said Refuge has since seen a rise in reports involving wearable technology, hidden cameras and microphones.

    Jen Reed, head of policy at University College London’s gender and tech research lab, said the findings showed technology had “transformed the way perpetrators commit abuse”. Such abuse is “increasingly normalised, with many people not recognising this form of abuse early enough”, she said, and survivors who seek help “are not always taken as seriously as they may be for other forms of abuse”.

    The data stops in 2022

    The most recent death in the published reviews occurred in 2022. Pickering noted that technology has moved on since, “including the unregulated development of AI”.

    The analysis therefore predates the period in which AI-generated sexual imagery became widely available — and predates every development this publication has reported this month.

    Six days earlier, a scoping review published in PLOS One had found that evidence on escalation from online abuse to offline violence “remains extremely limited”, identifying a single study tracing that progression among the research it examined for sub-Saharan Africa. The two documents measure different things in different places. Read together, they suggest the evidence base is thin because few have looked.

  • Sexual imagery was 96 per cent of deepfake files researchers found

    Sexual imagery was 96 per cent of deepfake files researchers found

    Sexualised imagery made up 96 per cent of 3.46 million synthetic files recorded across 821 deepfake attacks worldwide in the first half of 2026, according to a report by deepfake detection company Resemble AI.

    One in six of the total attacks detected involved non-consensual intimate imagery of adults or children. But those 137 incidents produced nearly all of the material.

    An incident in its dataset is one reported event no matter the size. Resemble AI said 87 per cent of all the files it counted — about 3.01 million — came from one tool, Grok. The Center for Countering Digital Hate puts the number of sexualised images produced during the Grok episode at three million. Set against roughly 3.32 million sexualised files in total, that event accounts for about nine in ten.

    Still images predominate

    A still image appears in 90.5 per cent of the 137 incidents rather than other forms of content.

    Resemble AI says that is due to the minimal effort required. It characterises this category as impulsive and without repeat targets: the opposite of political disinformation.

    Who was targeted

    Four in five of these attacks involved women or a mixed group when gender could be determined — 39 of 49 incidents. Thirty-six identified women only, three women and men, and two men only, with eight unspecified in the coverage. The report notes the first male victims appearing in its data, through extortion and fandom cases.

    The single largest event was down to Grok. On 29 December, X gave every user a one-click image-editing button powered by Elon Musk’s chatbot. xAI put the feature behind a paywall on 9 January and restricted “undressing” prompts on 14 January.

    Citing the Center for Countering Digital Hate, the report puts output at three million sexualised images — 190 a minute across an 11-day window, and 84 times the combined output of the five largest dedicated deepfake websites. Around 23,000 appeared to depict children: one in every 130 images, or one every 41 seconds. The New York Times separately estimated 1.8 million posts likely containing sexualised images of women.

    What the report says about its own limits

    The dataset is built from 1,760 news reports, so it counts reported attacks rather than attacks. And sexual imagery is the category most likely to surface late: 17 per cent of child sexual abuse material incidents were documented more than 90 days after the event, against 3 per cent of every other category. All four cases that took more than 250 days to surface reached the public record only through court proceedings.

    The company also calculates that the 14,915 documented victims of sexual imagery represent $2.24 billion in potential liability under the U.S. statute allowing $150,000 per victim.

  • Online violence is pushing Kenyan women out of politics, UN study finds

    Online violence is pushing Kenyan women out of politics, UN study finds

    Online violence is driving women out of politics in Kenya, and the patterns that disfigured the last general election are already repeating ahead of the next one, a UN Women study has found.

    It is happening in a country where a man convicted of sharing a woman’s intimate images without her consent faces a maximum compensation order of KES 200,000 ($1,500). A man who pirates music can be fined up to KES 5 million, 25 times more. UN Women calls the gap “a statement of values encoded in statute”, one that says “women’s bodily integrity, dignity and psychological well-being are valued well below commercial intellectual property”. The compensation cap comes from the Computer Misuse and Cybercrimes Act 2018, the fine from the Copyright Act.

    The Cost of Silence, published last month by UN Women’s Kenya country office under its DigiKen programme, describes itself as the first comprehensive assessment of technology-facilitated gender-based violence in the country. It says its chapter on political participation offers the most detailed picture yet of the effect on Kenyan women in public life, at a moment when the 2027 general election is approaching and the 2022 patterns are, on the testimony of multiple respondents, “already repeating and intensifying”.

    How candidates were attacked

    Researchers documented coordinated campaigns that combined online character assassination with physical intimidation. Online attacks on women candidates were in some cases followed by organised disruption of their rallies, which the report says suggests digital attack networks were coordinating with operatives on the ground.

    AI-generated deepfakes of women candidates were circulated inside their own constituencies, largely through community WhatsApp groups the candidates could not join and had no means of correcting.

    Analysis by Pollicy and the National Democratic Institute, cited in the study, found 55.7 per cent of the Facebook accounts of women candidates received online violence during the 2022 election. Eighty per cent of women candidates stayed active online throughout that campaign, against 93 per cent of men.

    Eight moderators for a continent

    The report attributes to its own interviews a figure it does not source elsewhere: that Meta has eight content moderation staff covering the whole of Africa, which it describes as a continent of 1.4 billion people speaking more than 50 languages. It calls that “not merely a resource allocation problem” but “a governance design problem”.

    It also records participants describing platforms rolling back moderation between 2023 and 2025 — content once removed being reinstated, human teams replaced by automated systems the report says cannot read Sheng or Swahili or the cultural context of Kenyan gendered abuse, and trust and safety teams cut entirely. The timing, it says, “coinciding with the approach of Kenya’s 2027 electoral cycle, is a cause for specific concern”.

    What participants said

    The study is qualitative: 22 key informant interviews and focus groups in Nairobi, Kisumu and Mombasa. Participants are identified by role only, under the consent protocol agreed with them.

    One survivor described a former partner posting her intimate photographs in 2018 to a Facebook group tied to her home village. “Everyone had access. Messages and calls came from every direction. It was a very dark, lonely period.”

    Another respondent said: “The online-to-offline continuum is not theoretical. It is the story of how women are dying in this country.”

    A third pointed at the vacuum the report wants filled. “We have policies in the country, but there is no policy that addresses TFGBV directly. Nobody knows what to call it.”

    The first of its seven recommendations is a standalone legal framework with an official definition, closing gaps in the Sexual Offences Act. Amendments to that Act, reported by the Daily Nation on 4 August, would create offences of making or sharing sexually explicit deepfakes, sextortion, stalking and voyeurism — two decades after it was passed, and two years before Kenyans vote again.

  • Nearly nine in ten Australian women in politics report online abuse

    Nearly nine in ten Australian women in politics report online abuse

    Nearly 90 per cent of Australian women who hold elected office or are seeking it have experienced online abuse or intimidation linked to their political participation, according to a survey published on Wednesday by Women for Election, an organisation that trains women to stand for public office.

    The survey is the organisation’s own research, commissioned in July 2026.

    It is one of the few Australian datasets that asks women in politics directly about AI-generated abuse. Sixty-five per cent of respondents said they had personally experienced or witnessed AI-generated abuse, deepfakes or manipulated content targeting women in public life, and 98 per cent said they were concerned artificial intelligence would make the problem worse over the next five years.

    Half now question whether they should have stood

    Eighty-two per cent of respondents said they had changed their behaviour because of online abuse. Of those, 58 per cent limited their use of social media, 36 per cent increased security measures and 34 per cent took time away from campaigning or public life. Half said that, knowing what they now know, they were unsure they would run again or would not have run at all. Sixty-four per cent said abuse had made them less likely to stand again, seek higher office or encourage other women to stand.

    Forty-five per cent said family members or campaign supporters had been targeted because of their political involvement.

    Only 47 per cent of those who experienced abuse reported it, and of those, 69 per cent were dissatisfied with the response they received. Women for Election’s media release states that 110 women were surveyed.

    Helen Blackburn, mayor of Bundaberg Regional Council in Queensland, told Australian Associated Press she had faced a campaign of AI-generated fakes across several platforms in the past two months, including a parody series made after she announced her council’s budget. One item featured the Grim Reaper, she said, followed by users writing that they hoped she would die and that “all I need is two bullets and I’ll sort it out”. Blackburn said she had considered resigning.

    The global picture

    Women for Election sets its findings alongside research led by Julie Posetti of City St George’s, University of London, for UN Women. That study, Tipping point: The chilling escalation of online violence against women in the public sphere, published in December 2025 and drawn from a 2025 survey of women journalists, human rights defenders, activists and other public communicators, found that 70 per cent had experienced online violence in the course of their work and that more than 40 per cent reported offline harm linked to online abuse.

    Among women journalists and media workers specifically, the share who linked offline attacks, abuse or harassment to online violence rose from 20 per cent in a 2020 UNESCO survey, also led by Posetti, to 42 per cent in 2025 — more than double.

    Women for Election is calling for greater accountability from social media and AI companies, dedicated protections for candidates during election periods, and a national action plan.

    Australia already has a statutory route. The eSafety Commissioner can require platforms to remove adult cyber abuse, but the threshold is narrow: the material must target one particular Australian adult, must be menacing, harassing or offensive, and must meet a test of serious harm judged by what an ordinary reasonable person would conclude rather than by the effect on the person targeted.

  • US poll shows huge concern over AI sexual images of women

    US poll shows huge concern over AI sexual images of women

    Two-thirds of American adults say they are concerned about artificial intelligence being used to create sexual imagery of women, according to a poll its authors describe as the first to put a number on public opinion about it.

    The survey found 67 per cent registered some or high concern about AI being used to create sexual imagery of women, and 71 per cent about AI deepfakes that sexualise children. It was published on 4 August by Women Who Tech, a non-profit network of women-led startups, and Rad Campaign, a digital agency that works for non-profits.

    Both organisations campaign on this issue and the poll is their own.

    Allyson Kapin, founder of Women Who Tech and a founding partner at Rad Campaign, said the results showed “policymakers and Big Tech companies are lagging far behind their own constituents and customers on this issue”. She said there was “a clear bipartisan majority for safeguards against sexual deepfakes of women and children”, adding: “Victims have waited long enough.”

    Respondents who reported “some” concern are counted alongside those reporting “high” concern, so the headline numbers measure the breadth of unease.

    Fieldwork ran from 13 to 18 June 2026, seven weeks before publication, among 1,015 adults aged 18 and over. Responses were collected online and weighted. The authors give a Bayesian confidence interval of 3.5 for 1,000 interviews, which they describe as roughly equivalent to a margin of error of plus or minus 3.1 points at 95 per cent confidence.

    The survey followed the period in which X was flooded with AI-generated sexual images produced by Grok, the chatbot built by Elon Musk’s xAI, depicting mostly women and children.

    Two bills are waiting in the House of Representatives. The DEFIANCE Act, which would allow victims of sexual deepfakes to sue those who make and share them, passed the Senate unanimously in January. The ENFORCE Act, which would let prosecutors charge AI-generated child sexual abuse material as they would any other, has also cleared the Senate.

    The TAKE IT DOWN Act, which requires platforms to remove non-consensual intimate images within 48 hours of a valid request, took effect on 19 May. The poll’s authors say it is too early to judge its effect.