Case Description
In September 2025, an Instagram user posted a short video of a woman in a form-fitting dress. In the video, the woman is adjusting her dress and moving her body, with her underwear visible in a few frames.
The next day, Meta’s automated system that detects and prioritizes content that may pose harm to individuals, especially content with a high likelihood of virality, identified the post. However, the report was not prioritized for human review.
A few days later, two users reported the content for pornography, but their reports were not prioritized for human review either, and the content remained on Instagram. One of the users who had reported the content appealed to Meta to take the content down, but once again the post was not prioritized for human review. The user then appealed to the Board.
When the Board brought the case to Meta’s attention, Meta’s subject matter experts reviewed the post and concluded that it did not violate the company’s Adult Nudity and Sexual Activity Community Standard, as it did not contain any “visual representations of a sexual encounter” or “language that facilitates or encourages sexual encounter.” The company therefore did not remove the content but determined it should only be visible to adults. Meta highlighted that under this policy, the company restricts visibility of “photorealistic or digital imagery of persons where crotch, buttock or female breast(s) are the focus of the image” to users aged 18 and up. It further stated that the individual featured in the video, whom Meta does not consider a public figure, draws attention to her crotch area and underwear by moving her legs and adjusting the clothing. The company added that keeping the content on the platform with age restrictions reflects their fundamental commitment to expression, while preventing younger audiences from material that may be inappropriate for their age.
The Board selected this case to assess Meta’s moderation practices in enforcing its Adult Nudity and Sexual Activity and related Community Standards, particularly in the context where a person’s image or likeness is used in a sexually explicit or suggestive manner without their consent. This case falls within the Board’s Automated Enforcement of Policies and Curation of Content and Gender strategic priorities.
The Board would appreciate public comments that address:
- Contextual information about the use and prevalence of AI-generated sexually explicit or suggestive imagery that is created without the consent of the person that is depicted.
- Views on protection of image or likeness in the context of AI-generated sexually explicit or suggestive imagery, while preserving, e.g., artistic expression.
- Best practices and recommendations for enforcing policies on AI-generated non-consensual intimate explicit or suggestive imagery, including approaches to determining non-public figures’ consent on image use.
- Views on the effectiveness of age-gating approaches to AI-generated non-consensual intimate explicit or suggestive
- Approaches to designing effective mechanisms to report AI-generated content where a person’s image or likeness is used in a sexually explicit or suggestive manner without their consent
In its decisions, the Board can issue policy recommendations to Meta. While recommendations are not binding, Meta must respond to them within 60 days. As such, the Board welcomes public comments proposing recommendations that are relevant to this case.
Public Comments
If you or your organization feel you can contribute valuable perspectives that can help with reaching a decision on the case announced today, you can submit your contributions using the button below. Please note that public comments can be provided anonymously. The public comment window is open for 14 days, closing at 23.59 Pacific Standard Time (PST) on Thursday 26 February.
What’s Next
Over the next few weeks, Board Members will be deliberating this case. Once they have reached their decision, we will post it on the Decisions page.
Comments
Prevalence and harms of image-based abuse:
Image-based is a deeply gendered and increasingly prevalent issue. My Image My Choice found there were over 270 thousand videos on the top 40 sites dedicated to AI-generated image-based abuse, a 3,000% increase from 2019. Sippy et al. (2024) surveyed 1,403 UK adults to understand exposure to and perceptions of AI-generated imagery and found that 18.8% of participants had been exposed to AI-generated intimate image abuse. In the 2023 State of Deepfakes report, 98% of AI-generated videos online were found to be sexualised and 99% of those depicted in these sexualised AI-generated videos were women.
Experiencing image-based abuse can have significant impacts on survivors, including their physical safety, emotional wellbeing, economic status, and more. This impacts marginalised communities in distinct and compounding ways. For example, Glitch has found that Black women are disproportionately likely to experience AI-generated image-based abuse, and WIRED estimated that 5% of imagery generated by Grok featured women who were stripped from or made to wear religious or cultural clothing.
The Oversight Board has in a previous case stated that deepfake intimate images have a disproportionate impact on women and girls, their rights to privacy, and protection from mental and physical harm. The findings of 2024-007-IG-UA and 2024-008-FB-UA found that "Given the severity of harms, removing the content is the only effective way to protect the people impacted."
Review process and policy enforcement:
In case 2026-021-IG-RA, multiple days passed before human review, and a timely response is vital when it comes to image-based abuse. Potential image-based abuse must be prioritised for human review, and it should not require contact with the Board for content to be reviewed. Decreased turnaround of review/takedowns must be prioritised in cases of image-based abuse, and this will be in line with the UK Government's proposed 48 hour maximum for the takedown of image-based abuse. This would be an absolute maximum, and Meta should aim for shorter timeframes than this as every moment such imagery is available on platform, it is available to be seen and saved by others. At a minimum, while this content is awaiting human review, it should be downranked to limit its further dissemination.
Once reviewed and identified as AI-generated image abuse, this content should be added to a hashing database such as StopNCII to prevent the upload or re-upload of this content across platforms. Sippy et al.’s (2024) research indicates that 87.3% of UK adults support the ban or suspension of users who distribute AI-generated imagery that may be harmful, and 82.4% believed social media platforms should make it easier to report harmful AI-generated imagery and request content removal.
The Bullying and Harassment Community Standard states that as a Tier 1 universal protection for everyone, everyone is protected from "derogatory sexualised photoshop or drawings". This means this post was by its nature due for removal, and I would also recommend that the language in this Community Standard is updated to "non-consensual synthetic intimate imagery" or similar. This would expand beyond "photoshop or drawings" to clarify that any synthetic content would be unacceptable, and it would go beyond the indeterminate "derogatory" to be clear that regardless of intention, no non-consensual image-based abuse can be allowed on the platform. This is in line with previous Board findings. I would also recommend that Meta’s position on AI-generated image-based abuse is clarified in the NCII section of the Bullying and Harassment topic as this is not mentioned currently.
The Board decision relating to 2024-007-IG-UA and 2024-008-FB-UA found that labelling is not appropriate as the harm comes from sharing and viewing, and this equally applies in this case where age-gating may limit the spread, but ultimately does not address the core harms and leaves the imagery available to a significant audience.
Further recommendations:
In Ofcom's Deepfake Defences paper, they share potential enforcement actions such as issuing warnings/strikes to users who have broken such guidelines, taking down content, suspending or removing users, and labelling content where there isn't a clear breach. Their guidance on improving women and girls' online safety sets out good practice steps including: consulting with subject-matter experts around setting policies such as this, training employees responsible for setting policies on online gender-based harms, requiring evidence of consent from those depicted in intimate content prior to upload, and adding deterrence messaging to the upload process. I would also recommend that if these images are believed to be consensual, this consent would still have to be understood to be ongoing and if at any stage someone wishes to be able to revoke consent in imagery that depicts them, they should be able to do so.
Ofcom note that "it is important for providers to recognise that user reporting relies on survivors and victims of online gender-based harms, and that reporting processes are time-intensive and risk re-traumatising survivors and victims". They state that users can be encouraged and enabled to report by making these reporting systems accessible, transparent, easy-to-use, and accounting for the dynamics of online gender-based harms. Foundational steps in this guidance and the related Codes include acknowledging receipt of complaints, providing indicative timeframes, and setting out information about how the complaint will be handled. Good practice steps to go further could be quick exit buttons, allowing report tracking, allowing users to give feedback on reporting processes, signposting to relevant supportive materials and organisations during reporting, and establishing trusted flagger programmes.
Finally, Meta is a signatory of the IBSA Principles for combatting image-based sexual abuse which highlight trauma-informed approaches and investing in resources and tooling to ensure the rapid processing of reports. For further resources on improving reporting processes, Chayn and End Cyber Abuse have created the Orbits guide which includes detailed principles for trauma-informed approaches to technology-facilitated abuse and an audit template for reviewing platforms according to these principles.
I oppose to AI generated sexually suggested imagery and expect Meta to remove these images
WITNESS Public Comment on Non-Consensual AI Sexualized Impersonation Case
Introduction
WITNESS submits this public comment as a human rights organisation with over thirty years of experience in media technology, documentation, and accountability, with particular expertise in AI governance, platform accountability, and how AI-generated content affects evidence verification and gender-based violence.
In July 2024, WITNESS submitted public comment to the Board’s Explicit AI Images of Female Public Figures case. The Board cited WITNESS twice in that decision, and our analysis shaped the Board’s consent-centred framing of non-consensual intimate imagery (NCII) and its critical separation of NCII harms from the broader “manipulated media” detection frame.
The Board made four recommendations. To date, Meta has not meaningfully implemented them.
Meta's September 2024 response to the four recommendations from the Board declined the most important (moving the prohibition into Adult Sexual Exploitation), deferred the most urgent (treating AI-generated context as a signal of non-consent), and agreed in principle only to updating "photoshop" terminology while ruling out replacing "derogatory" with "non-consensual." Meta's own Transparency Center page states that it "do[es] not expect to replace 'derogatory' with 'non-consensual'" and "do[es] not expect that this will result in moving the prohibition." These are not pending implementations. They are documented refusals. The Bullying and Harassment policy still contains the phrase "derogatory sexualized photoshop or drawings," and Meta told CBS News in February 2025 that it was "still considering" how to treat AI-generated context as a non-consent signal, seven months after the Board's recommendation. In the case now before the Board, Meta's subject matter experts assessed the content under Adult Nudity and Sexual Activity, a policy with no framework for consent, rather than the Adult Sexual Exploitation standard the Board recommended in 2024, or even the Bullying and Harassment standard where Meta itself insists the prohibition should remain.
Meanwhile, a June 2025 CBS News investigation found hundreds of nudification app advertisements running across Meta's platforms, removed only after publication.
Every vulnerability the Board identified in 2024 including over-reliance on media reports as a consent proxy, auto-closure without human review, policy routing that diverts NCII from appropriate enforcement, recurs here. But the victim is a non-public figure: the exact population the Board warned was most at risk.
Since PC-27095, WITNESS has continued to build its evidence base: our March 2025 submission to the UN Human Rights Council documented AI-driven sexual violence as a top-tier threat; our December 2025 submission to the Board's Iran-Israel case addressed likeness protection; and our analysis of the January 2026 Grok NCII cases demonstrated that harm occurs at creation, not only distribution.
This case highlights a structural failure to treat AI-generated sexualized impersonation as sexual exploitation.
Tech facilitated Gender based violence
AI generated NCII is a form of technology-facilitated gender-based violence that disproportionately affects women. A recent study by the Oxford Internet Institute analyzed nearly 35,000 deepfake models available for public download and found out that 96% of these models targeted identifiable women. Platforms must address this as image-based sexual abuse, not merely as a content moderation challenge.
Responses to the Board’s Questions
Question 1: Prevalence and Use of AI-Generated NCII
AI-generated NCII has rapidly shifted from fringe abuse forums into scalable, commercialized and mainstream distribution ecosystems. The key development is not only increased volume, but increased ease, plausibility, and reach. The issue is no longer episodic misuse. It reflects structural features of generative systems, advertising infrastructure, and enforcement design. We highlight six interrelated trends.
1. Industrialization of abuse: Consumer-facing nudification and deepfake tools have dramatically lowered the barrier to producing sexualized impersonations at scale. What distinguishes the current moment is the embedding of these tools into commercial ecosystems: nudification apps advertised on mainstream platforms, abuse services operating through subscription models, and platform advertising systems functioning as discovery infrastructure for exploitation.
2. Gendered targeting: As documented in WITNESS’ March 2025 submission to the UN Human Rights Council, AI-driven sexualized impersonation disproportionately targets women, girls, LGBTQ+ individuals, journalists, and human rights defenders. It functions as technology-facilitated gender-based violence (TFGBV), reinforcing patterns of misogyny, coercion, humiliation, and silencing. For non-public figures, consequences including reputational harm, employment loss, extortion, psychological trauma, frequently occur long before content moderation concludes.
3. Likeness exploitation and the erosion of consent signals: High-fidelity generative systems (including tools such as OpenAI's Sora) mark a structural shift in the synthetic media ecosystem, enabling increasingly precise exploitation of likeness without consent. As realism increases, harm expands beyond explicit nudity to include sexualized impersonation that may not meet strict nudity thresholds but nonetheless constitutes intimate exploitation. Platforms often rely on contextual cues of non-consent such as media reporting or public figure status, yet AI-generated impersonation frequently targets non-public figures in private networks where no public documentation exists. The absence of media coverage cannot be treated as evidence of consent.
4. Plausible deniability and verification collapse: Increasing realism enables perpetrators to evade accountability through plausible deniability ("it's fake," "it's a joke"), shifting the burden onto victims. Technical safeguards such as watermarking remain inconsistent and are often stripped across platforms. And as WITNESS documented, detection tools are unreliable in real-world conditions, particularly across multimodality, file quality, and global contexts. This is compounded by platforms' own failure to identify synthetic content on their services: an Indicator audit found that platforms fail to label AI-generated content 70% of the time, including content produced with their own tools. Enforcement frameworks that depend on proving AI generation or conclusively verifying manipulation will therefore systematically fail victims.
5. Harm at creation: The January 2026 Grok incident illustrates that harm occurs at the moment of synthetic creation when a system generates non-consensual sexualized depictions upon prompt, the violation has already occurred regardless of distribution. AI-generated NCII is not only a moderation issue; it is a product design and system safeguards issue.
6. Capture-to-exploitation pipeline: The proliferation of AI-enabled wearable devices is creating new vectors for non-consensual sexualized impersonation. Smart glasses equipped with discreet cameras, including Meta's Ray-Ban smart glasses, enable covert capture of likeness in public and semi-public spaces. In February 2026, a Russian national used such devices to secretly record encounters with women across Kenya and Ghana, monetising the footage through a multi-platform pipeline: free clips on TikTok and YouTube driving traffic to a paid Telegram channel, the same advertising-to-exploitation model documented in the nudification app ecosystem Kenya's Ministry of Gender, Culture and Children Services characterised the incident as "a serious form of technology-facilitated gender-based violence and exploitation." As global smart glasses shipments grew 210% in 2024 and major manufacturers integrate generative AI capabilities directly into wearable hardware, covertly recorded imagery can increasingly be fed into nudification or synthetic impersonation tools with minimal friction, creating a seamless pipeline from capture to abuse. This is particularly relevant to the Board because Meta is simultaneously the platform where NCII circulates and quickly becoming a manufacturer of the hardware that facilitates covert likeness capture.
Question 2: Likeness Protection and Artistic Expression
The core principle is consent. There are categorical distinctions between: transformative or satirical expression that does not sexualise an identifiable individual; sexualized impersonation that exploits a person's likeness without consent; and consensual use of synthetic media by individuals, including sex workers, to protect their identity or exercise agency over their own likeness. The distinguishing principle across all three is consent.
The recent Kenya and Ghana incident described above underscores this. As Kenyan journalist Ferdinand Omondi wrote: "Consent to sex is not consent to filming. Consent to filming is not consent to publication." Neither artistic nor commercial framing can cure non-consensual exploitation.
International human rights law protects expression, but also dignity, privacy, and bodily autonomy. Consent must be the governing standard. In practice: platforms should treat realistic AI-generated sexual content depicting identifiable individuals as presumptively non-consensual; public figure status does not nullify consent requirements; artistic exemptions should not apply where sexualisation is involved; and likeness protection should not depend on whether content is labelled as AI-generated. The violation lies in the non-consensual sexualisation, not the technical method.
Question 3: Enforcement Best Practices
a. Correct policy architecture: The prohibition on non-consensual sexualized manipulated media should be moved from Bullying and Harassment into the Adult Sexual Exploitation (ASE) standard, where enforcement expertise is aligned with exploitation harms. The requirement that content be "non-commercial or produced in a private setting" should be removed: AI-generated content is by definition not produced in a private setting, and the nudification ecosystem is inherently commercial. Retaining this language creates a loophole excluding the precise harm this case involves. AI-generated context should be treated as a signal of non-consent — consent to share an image publicly does not mean consent to its sexualised manipulation.
b. Adopt a presumption of non-consent: For realistic AI-generated sexual content depicting identifiable individuals, platforms should presume non-consent unless affirmative, verifiable consent is demonstrated. Public figure status and absence of media coverage cannot serve as proxies for consent. As Danielle Citron has argued, public revelation of a person's sexual expression without consent interferes with autonomy and self-respect. The burden should not rest on victims.
c. Mandate human review and prohibit auto-closure: Reports involving AI-generated sexual impersonation must not be auto-closed. "Intimate" content is highly contextual and cannot be reliably inferred from nudity alone — as this case demonstrates. There is currently no independent benchmark for evaluating NCII detection across diverse content types, file qualities, and cultural contexts. WITNESS sees a need to extend its TRIED AI Detection Benchmark to NCII. Until such evaluation exists, automated detection claims cannot substitute for human review.
d. Address creation-level risk through product safeguards: Platforms deploying or integrating generative systems should implement: prompt-level safeguards preventing sexualized impersonation of identifiable individuals; likeness-protection guardrails for everyone; escalation protocols when models generate prohibited impersonations; and responsible design standards for AI-enabled capture devices (including smart glasses) to prevent covert likeness capture from feeding into exploitation pipelines, particularly where, as with Meta, the hardware manufacturer and the content platform are the same company.
e. Transparency as a precondition: Without independently verifiable transparency, enforcement claims lack credibility. Building on WITNESS's 2024 submission and CDT's Model NCII Policy, platforms should publish detection rates, false positive/negative rates, removal timelines, geographic and language disparities, and appeals outcomes.
Question 4: Effectiveness of Age-Gating
Age-gating is not an effective safeguard against AI-generated NCII. Age verification mechanisms can be easily circumvented, and the core harm is tied to consent and impersonation, not viewer age. Age-gating does not prevent the creation of abusive content and does not meaningfully address the exploitation of victims, many of whom are adults.
Age-based restrictions should complement enforcement, but they cannot substitute for consent-based prohibition and rapid removal mechanisms.
Question 5: Reporting Mechanisms
Effective reporting requires: a dedicated category for AI-generated sexualized impersonation distinct from harassment or nudity; evidence-sensitive design allowing victims to report without re-uploading abusive content; removal of requirements to demonstrate public visibility or private-setting production; time-bound specialist review rather than auto-closure; and cross-platform coordination through mechanisms such as StopNCII, with due process safeguards.
3. Recommendations
WITNESS agrees with and urges the Board to reiterate its 2024 recommendations, which remain unimplemented. We consider the need to change the framing of Synthetic NCII, lack of consent in the generation of these contents and call to strengthen transparency still timely and needed in regards to many of the points highlighted in our submission.
Besides that, we would like to recommend the following:
1. Recognize AI-generated NCII as technology-facilitated gender-based violence, aligning policy language accordingly.
2. Adopt a presumption of non-consent for realistic AI-generated sexualized impersonations of identifiable individuals, regardless of the fact that such images might have been generated based on media made available by the individuals themselves - as it implies a manipulation of the purposes regarding the image uses.
3. Mandate human review for all AI sexual impersonation reports.
4. Create a dedicated reporting pathway for AI-generated sexualized impersonation.
5. Implement product-level safeguards to prevent the generation of sexualized impersonations at the system design stage.
6. Extend product accountability to AI-enabled capture devices. Meta should establish responsible design standards for its hardware products to prevent covert likeness capture from feeding into exploitation pipelines.
Any AI generated image, video, or graphic depicting someone in a sexual manner or containing any form of sexual harassment (unless purely for educational purposes) should be removed. Additionally, any AI generated image, video, graphic of any other person without their consent should be removed if it is sexual in any capacity or if used for false advertisement.
The creation of AI-generated sexually explicit or suggestive imagery without the consent of the person depicted raises significant human rights concerns that extend beyond questions of platform governance or content moderation. Such practices directly infringe upon the rights of individuals to privacy, dignity, reputation, equality, and security of person, and may hinder their ability to participate freely in social, professional, and civic life. These harms are particularly acute for women, who are disproportionately targeted by sexualised manipulation of their image or likeness.
Freedom of expression, including artistic expression, is a fundamental right. However, under well-established human rights principles, it is not absolute and must be exercised in a manner that respects the rights and freedoms of others. It is important to note that expression cannot be used as a defence for conduct that violates another person's privacy, appropriates their likeness, or exposes them to harassment, reputational damage, or sexualised abuse. In instances where expression directly conflicts with these rights, a careful balancing is required, and the absence of consent must weigh decisively against the continued distribution of the content.
This is especially true when the person portrayed is a private individual who has not voluntarily entered the public sphere. Non-public figures should not be exposed to the risks associated with public scrutiny, nor be implicitly forced to relinquish control over the use of their image. In such circumstances, consent must be regarded as non-negotiable. The fact that content may be presented as satire, art, or commentary does not override an individual's right to control the use of their likeness, notably in sexualised contexts.
It is therefore recommended that a rights-respecting approach adopts a clear default rule: whenever an identifiable person states that they did not consent to the creation or distribution of AI-generated sexually explicit or suggestive imagery depicting them, the content should be removed without delay. The onus is on the uploader, as the party initiating the dissemination, to demonstrate lawful and consensual use. This is because they are the party who could benefit from the dissemination of the content, whether in terms of attention, reputation, or other forms of value. Platforms themselves also benefit from engagement generated by such material and therefore bear responsibility to ensure robust safeguards.
In instances of non-consensual sexualised imagery, age-gating measures alone are insufficient. Restricting access to adults does not mitigate the core harm arising from the existence and circulation of the content itself, nor does it prevent reputational injury, harassment, or downstream redistribution.
In order to ensure effective policy enforcement, it is essential to establish clear and accessible reporting channels specifically designed for AI-generated non-consensual content. These channels should be supported by prompt human review by trained teams, transparent standards explaining how consent is assessed, and proactive detection measures to limit the spread of such material. Platforms should also provide clear communication to affected individuals regarding actions taken and available remedies.
In considering this case, the Board has an opportunity to clarify that protecting freedom of expression must go hand in hand with protecting individuals from the misuse of emerging technologies in ways that undermine their fundamental rights. Establishing consent as the central criterion for the treatment of AI-generated sexualised depictions of identifiable persons would align platform governance with widely recognised human rights principles and help ensure that technological innovation does not come at the expense of human dignity.