New 7,230-person survey finds 7% of adults report AI-generated nonconsensual intimate imagery. Key findings on who’s affected and what helps.
A new peer-reviewed survey of 7,230 adults across Australia, the UK, and the US finds that 7.0% of respondents — roughly 1 in 14 — report having experienced AI-generated image-based sexual abuse (AI-IBSA): someone nonconsensually creating, threatening to share, or sharing an AI-generated intimate image of them. The study, published in the International Journal of Human–Computer Interaction in September 2026, is among the first large-scale efforts to measure this specific harm rather than digital image manipulation broadly, and its findings carry weight for parents, platforms, and policymakers trying to gauge how widespread this problem has become.
What They Found
The survey, fielded in mid-2024 with a demographically weighted sample, found that 15.1% of all respondents reported some form of image-based sexual abuse involving digitally altered images (AI-generated or otherwise), while 7.0% specifically involved AI-generated content. Broken down by behavior: 4.3% reported someone created a nonconsensual AI image of them, 3.6% reported a threat to share one, and 4.4% reported an image actually being shared. The UK had the highest AI-IBSA rate (9.3%), followed by Australia (7.7%) and the US (6.5%). When images were shared, social media posts (39.4%) and private messages (32.9%) were the most common distribution routes.
The most counterintuitive finding: men reported AI-IBSA victimization at more than twice the rate of women (10.2% vs. 3.8%, odds ratio 2.28). That cuts against the common public assumption that this is overwhelmingly a problem targeting women — an assumption shaped by high-profile cases of celebrity deepfakes, which are indeed disproportionately made of women. The researchers offer an explanation: their survey measures victimization people are aware of — sextortion attempts, peer pranks, images someone found out about — which may disproportionately involve men and boys as visible targets. Women-targeted images made for private sexual gratification may circulate quietly without the subject ever learning they exist, making that harm largely invisible to a survey instrument like this one. Other demographic patterns were more expected: people under 35 had over 3.5 times higher odds of victimization than older adults, BIPOC respondents had 1.70 times higher odds than white respondents, and people with disabilities had 1.89 times higher odds. LGBTQ+ respondents overall had 1.66 times higher odds, though this was concentrated among LGBTQ+ respondents 35 and older (3.01 times higher odds versus older non-LGBTQ+ respondents), with little difference among younger respondents.
Despite lower reported rates, women who were victimized reported a broader range of harms and more severe emotional impact — 42.6% reported feeling depressed versus 25.7% of men, and 42.4% reported fear for their safety versus 18.1% of men. Men were more likely to report being physically hurt in relation to the incident (8.2% vs. 1.3%) and more likely to report neutral or positive reactions. Most victim-survivors (90.3%) reported at least one negative feeling, and most (69.7%) both reported the incident somewhere and disclosed it to someone for support — though researchers identified shame, a sense of futility, and minimization (“it’s not even real”) as recurring barriers to seeking help.
What This Means
I think the gender finding here is worth sitting with rather than resolving too cleanly, because it complicates a framing that’s become fairly standard in AI-safety commentary: that deepfake abuse is fundamentally a women’s-safety issue. That framing isn’t wrong, exactly — the researchers themselves note it’s likely correct for a huge, largely invisible category of harm. Women’s images used for private sexual gratification may never surface to the woman herself, meaning a self-report survey like this one structurally cannot capture that harm at all. What the survey can measure is the harm people become aware of, and there, men and boys show up more often — plausibly because sextortion scams and peer-network pranks are more likely to be discovered by the target than content made in private for someone else’s use.
That distinction matters practically, not just academically. If platforms and lawmakers design interventions around “the visible problem” — the sextortion attempts, the peer harassment, the cases that get reported — they may be building for the wrong distribution of harm, missing the much larger population of women whose images are being used without their ever finding out. The researchers’ own numbers on reporting tools reinforce this: 33% of victim-survivors in this study didn’t even know StopNCII.org — a hash-matching tool that lets people proactively flag their images for detection — existed as an option, and only 43% of those who tried it found it helpful. Hash-matching also has an inherent limitation the paper flags: it requires the victim to have the actual image to hash, which is useless for someone who never learns an image exists.
Newer approaches like reverse facial-recognition tools (e.g., Digital Dignity) and provenance standards like C2PA and SynthID at least attempt to address discovery and verification, respectively, but neither solves the fundamental asymmetry between visible and invisible harm that this survey’s own numbers seem to be describing. I’d also flag, per the source’s own disclosure, that three of the four authors are Google/YouTube employees and the research was funded by an unrestricted Google/Alphabet grant to RMIT University — worth knowing as context, alongside the fact that the paper states this plainly itself and that YouTube’s own “AI Likeness” detection tool is one of the interventions discussed favorably in the paper.
What to Watch
Watch whether platforms move beyond reactive hash-matching toward provenance-based tools like C2PA and SynthID at meaningful scale over the next year, and whether more jurisdictions pass deepfake-specific legislation rather than relying on older image-abuse or harassment statutes not written with AI generation in mind.