Violation of Consent: Public Attitudes Toward AI-Generated Non-Consensual Intimate Imagery and Deepfakes

Advances in artificial intelligence have made it possible to create hyper-realistic synthetic media, known as deepfakes. Blending 'deep learning' and 'fake,' the term describes AI-generated images, audio, and video that depict real individuals. While deepfake technology offers benefits in creativity, accessibility, and entertainment, it has also been used to spread disinformation, commit fraud, and non-consensually generate intimate imagery. The last of these has drawn growing concern. Because generative AI can now produce deepfakes from text prompts alone, without a dataset of images of the target, the barrier to creating such content has dropped considerably. Understanding how the public views these uses is essential, since technical capability alone does not determine how society will accept or reject them.

The modern use of the term 'deepfake' traces to 2017, when a Reddit user posted synthetic videos of celebrities in sexual acts. A subreddit dedicated to such content quickly attracted over 90,000 users before being banned. Despite platform bans, AI-generated non-consensual intimate imagery (AIG-NCII) continues to circulate on dedicated forums. AIG-NCII is one form of image-based sexual abuse (IBSA), which encompasses the non-consensual creation, distribution, or threatened distribution of intimate images. Victim-survivors of IBSA often experience post-traumatic stress, anxiety, depression, isolation, and diminished self-esteem. Victim-blaming attitudes can further discourage help-seeking. Because IBSA is predominantly, though not exclusively, perpetrated against women, it is widely understood as a form of technology-facilitated gender-based violence. Legislation addressing these harms remains sparse in many countries, including the United States, underscoring the need to understand public attitudes that can inform policy.

Prior studies have found significant public concern about deepfakes used for disinformation, with less concern when synthetic media is used for entertainment or humor, or when it is labeled and traceable. Research on AIG-NCII has focused largely on perceptions of criminality and harm. One UK study found strong support for criminalizing both sexual and non-sexual deepfakes, and found that labeling content as fake reduced perceived harm for non-sexual deepfakes but not for AIG-NCII. Other work found that deepfakes of celebrities, especially men, were seen as less harmful, and that creating content for personal sexual gratification was viewed as less harmful than sharing it. Cross-national research found low awareness of AIG-NCII, while men reported more both perpetration and victimization. Community-level studies have also documented pro-deepfake attitudes alongside misuse concerns. Yet prior work has rarely examined acceptability as distinct from criminality, and has seldom explored qualitative reasoning behind respondents' views.

This study asks: What are people's attitudes toward the hypothetical non-consensual creation, sharing, and seeking out of deepfakes depicting them? Specifically, it addresses four research questions. First, how do attitudes differ between AIG-NCII and non-consensual content depicting non-sexual acts? Second, how do attitudes shift depending on who creates the media and for what purpose? Third, how do attitudes toward sexual consent and sexual content shape these judgments? Fourth, how does gender influence them? By focusing on acceptability rather than criminality alone, and by incorporating contextual and individual factors, the study aims to inform public, technical, legal, and policy discussions around deepfakes.

The researchers surveyed 315 U.S. adults recruited through Prolific. The survey first explained generative AI's capacity to produce realistic fake media, deliberately avoiding the word 'deepfake' to prevent priming effects. Respondents then read vignettes describing hypothetical scenarios in which generative AI creates a video of them without their knowledge. Each vignette varied three factors: the action depicted (performing a sexual act, playing a sport, or saying something), the creator (an intimate partner or a stranger), and the creator's intent (harming the respondent, entertainment, or sexual pleasure). This produced 18 vignettes in a full-factorial design, with each respondent rating three at random. Respondents rated acceptability on a five-point Likert scale and provided short written rationales. The study also measured attitudes toward sexual consent using validated subscales of the Sexual Consent Scale-Revised, and attitudes toward genuine intimate imagery and non-consensual distribution of intimate images. Analysis used cumulative link mixed models for quantitative data and a coding reliability approach for qualitative responses.

Respondents expressed strong opposition to creating and, even more strongly, to sharing non-consensually created synthetic content, particularly when it depicted sexual acts. Seeking out such content, however, appeared more acceptable to some respondents, revealing a notable gap between passive consumption and active distribution. Acceptability also varied with context. The creator's relationship to the respondent mattered, as did the creator's intent. Respondents' gender shaped judgments as well: the study grouped men separately from women, non-binary, and agender respondents, and found meaningful differences in attitudes toward creation. Attitudes toward sexual consent further influenced how acceptable respondents found these scenarios. Together, these findings suggest that public norms are not uniform but depend on who is involved, why the content is made, and how individuals understand consent.

Open-ended rationales help explain the quantitative patterns. Respondents' reasoning centered on violation of bodily autonomy and personal dignity, with some describing the creation of sexual content of themselves without consent as a deep violation. Others focused on the harm of distribution, emphasizing reputational damage and loss of control over how they are seen. Some respondents distinguished between entertainment and malicious intent, while others considered the relationship with the creator as a factor in their judgment. These rationales illustrate that people intuitively apply concepts of consent and privacy to synthetic media, even when the depicted act did not actually occur.

The findings suggest that public opposition to AIG-NCII is strong, and that sharing is viewed as more harmful than creating or seeking content. This has implications for policy: laws may need to address distribution and intent explicitly, not only creation. Platforms can use these insights to prioritize the removal and prevention of sexual deepfakes, and to design friction around sharing. On the technical side, the research highlights the value of detection tools, labeling, and mechanisms that disrupt unauthorized generation. Because attitudes vary with gender and consent beliefs, educational efforts that build norms around consent may complement legal and technical measures. Aligning technology development with public expectations can help ensure that generative AI's benefits are realized without enabling image-based sexual abuse.

As a hypothetical vignette survey of U.S. respondents, this study captures stated attitudes rather than actual behavior, and its findings may not generalize across cultures or legal contexts. The gender groupings, while necessary for statistical power, combine diverse identities. Future research should examine real-world behavior, compare attitudes across countries, and explore how interventions such as labeling or consent education shift public norms. Longitudinal work could also track how attitudes evolve as the technology and its legal treatment mature.

AI-generated non-consensual intimate imagery poses a growing threat, and public attitudes provide essential guidance for policy and technical design. This survey shows that Americans broadly oppose the creation and especially the sharing of non-consensual synthetic sexual content, while context, intent, relationship, gender, and consent attitudes all shape judgments. Addressing this emerging form of image-based sexual abuse will require coordinated action across law, platforms, technology developers, and public education.