This project investigates AI that is devised to violate women, children, and vulnerable individuals by undressing and sexualizing their images and exposing them online. It’s a project by
digital rights activist Paolo Cirio, who investigated nudify AI tools through several contexts, case studies, and legal undertakings.
This campaign advocates for the banning of AI content, platforms, and software that produce such violations. It suggests AI detection of non-consensual AI-generated violating content and supports independent content moderation.
Read more about the advocacy proposals to tackle the AI nudify emergency.
As a main case study, this project examines Elon Musk’s AI language model, Grok, and the abusive content generated and published at his direction. Musk allowed users to generate and expose millions of violent and sexualized images of women and children without their consent.
Cirio assembled a list of accusations and issues that remain unsolved regarding the abuse of Grok.com and X.com, and more broadly, AI nudifying capabilities.
Cirio’s investigation involved finding evidence of abuse that stays online on Grok.com by collecting a sample of around 15 thousand AI-generated images and then using AI to detect photorealistic visuals that sexualized women, children, politicians, celebrities, and vulnerable individuals without consent. This sample is only a small portion of the images generated in total and of those still stored on Grok.com and reposted on X.com.
Ultimately, Cirio’s investigation involved forensic analysis of the metadata, sources, visuals, and contexts of these images, uncovering evidence of generative AI abuse that circulates online and harms citizens. The findings of Cirio’s investigation will be sent to organizations to help regulate and remove AI-generated non-consensual content and to help victims seek justice against Elon Musk.
Cirio published some of the evidence found by AI on this project's website. To preserve the subjects’ privacy, faces or identifiable details in the images are excluded, as well as any linked data. In addition, the individual cropped details have their colors inverted to make them more abstract and less vulgar, emphasizing technical aesthetics rather than sexual ones. The result is a visual database of thousands of details of AI-generated images browsable by keywords related to body parts, flesh, fluids, gestures, clothes, and objects that can include controversial content. In the database there is evidence of non-consensual visuals, revealing the scale of abuse and how to mitigate it.
This project aims to undress AI by stripping away its power and revealing its bare political use against society. Through a final essay,
Cirio analyzed Generative AI-Violence theoretically and practically as a media phenomenon.
Since December 24, 2025, Elon Musk’s AI, Grok, has generated millions of non-consensual sexualized images and videos. The resulting content circulated extensively on X.com (formerly Twitter). In just 11 days, Grok generated an estimated 3 million sexualized images, mostly of women, and 23,000 appeared to depict children, as reported by the non-profit watchdog Center for Countering Digital Hate (CCDH). The total amount of abusive content generated remains unknown. Even after public outcry, the feature remained available to paid subscribers. Only from April, the platform began to delete the abusive content generated in the prior months. In May 2026, Grok finally removed its nudify feature, but only in the countries where it is considered illegal. Meanwhile, thousands of abusive AI-generated images and videos still remain on Grok.com and are shared and tweaked on the open Internet.
EVIDENCE from the investigation
Browse through the investigation presented as evidence of the AI-generated sexualized images found on Grok.com and automatically detected by AI.
This visual database is browsable by main keywords related to sexualization.
Click the links to filter results with anonymized and stylized image crops to present the evidence.
Only some of these images are evidence of non-consensual and abusive visuals. Many others still need verification, and others remain unpublished to protect the subjects’ privacy. This analysis was made only on a small sample of source images. Millions more pieces of evidence remain to be found, even after Grok claimed to have deleted all abusive content.
Methodology of the investigation
Step 1) Collect a sample of 15.000 AI images from Grok.com that were generated around the time of the scandal at the beginning of 2026.
Step 2) AI image detection to filter out faces from the images collected to protect the privacy of the individuals depicted in the AI-generated photos.
Step 3) AI image detection to filter out graphics, comics, and anime content from the collected images, keeping only photorealistic AI-generated pictures.
Step 4) AI image detection of sexualization via keywords that refer to body parts, fluids, and gestures. Sorting the image files with the resulting cropped visual details.
Step 5) Manual check of the results to understand the context of the images, identifying false positives and removing samples from the dataset that may not relate to the objective of the investigation.
Step 6) Cross-reference the data of content posted on X.com and findable on Grok.com. Analysis of metadata including users’ names, posting dates, and chat thread IDs.
Step 7) Apply visual effects to all images found to abstract details and sexualization for the publishing of the evidence as visual databases without any linked data to the original victims.
Step 8) Assemble a report with the investigation’s findings and details of key evidence for submission to support legal claims, legislative action, and journalistic dossiers.
* The AI model Claude Sonnet 2.0 was used for image detection.
ISSUES and accusations
Some AI models and platforms are designed to allow the creation and sharing of abusive AI-generated content. The following list of accusations against Grok.com and X.com can often be applied to other online platforms engaging in similar abusive behavior.
-
Many violating images remain online, both on X.com and Grok.com, without the victims’ knowledge, who consequently can't ask for the takedown of their own images. Even when victims are notified about harmful content, they might not have the legal or technical means to take action to remove it.
-
Reported images are not taken down, even when platforms, such as X.com, claim to comply with regulations and their own policies ban this type of content. Only with legal threats may users finally be able to force platforms to take down abusive content. However, platforms often don’t respond to or accept complaints, and sometimes they even appeal.
-
After six months of abuse, the platform X.com announced that it had added measures to stop Grok from sexualizing people. However, this limitation applies only in jurisdictions where it is illegal. Thus, abuse persists through bypassing geoblocking or in countries that haven’t legislated it yet.
-
Platforms don’t respect their own policy that they claim to safeguard. Nudify guardrails set by big platforms were already established at the end of 2025; however, the Grok and X.com scandal happened after those claims. Thus, platform self-regulation should not be considered reliable.
-
The takedown of abusive AI images doesn't necessarily stop the images from moving elsewhere or from being tweaked a little bit and reposted on the same platform. Platforms like X.com and Grok need to have automated content detection, not just deletion after reports.
-
It’s hard to verify if content generated by AI is consensual, if the material has been stolen, if the scenario is abusive, and also if it contains minors. AI blends content together, making it untraceable, and users don’t declare sources and intentions. This is especially true if the victim is not famous, if content is uploaded without consent, or when it's hard to verify if content was generated from parts of real people.
-
The policies of platforms and AI models, such as Grok.com and X.com, have loopholes that allow the generation of adult content. This includes caricature and artistic depictions, which often enable abuse in ways that bypass existing guardrails. Other policies don’t limit non-consensual content in premium services and within private chats.
-
The prompts to generate AI content can be hacked to compose abusive content. Thus, only AI detection of outputs and content sources can mitigate abuse. For instance, prompts like "revealing summerwear" and “swap clothing” instead of “bikini”, and “upscale” modes can reveal hidden details.
-
Bots can prompt AI models to automatically generate abusive and controversial content and share it on platforms, shifting accountability and making it hard to track abuse. Bots on X.com can prompt Grok to generate abusive content. Chat threads are also algorithmically amplified in feeds to boost engagement on controversial content.
-
Abusive AI images that are generated from other software, platforms, models, online and offline, can be posted on the platform X.com. Often, the issue goes beyond Grok.com and is about X.com's policy and moderation. Exposure on large platforms is often more harmful than the actual generation of abusive content.
-
The platforms X.com and Grok.com are advertised as places where abuse is tolerated. The scandal itself functioned as promotion for Grok's services. Other abusive Nudify platforms are promoted on X.com, and the whole platform X.com is promoted as a site for sexualized and controversial content.
-
Even if nudification and sexualization on X.com are banned, the scandal itself normalized its behavior and the media model. Today, other smaller platforms play by the book that Grok.com designed, and AI-generated abusive content is now more prolific. These smaller platforms are less regulated as they don’t need to abide by the same DSA regulations (EU Digital Services Act) and are based in jurisdictions with loopholes; as a result, they are far less moderated and monitored.
-
Elon Musk hasn’t been charged, nor questioned in a congressional hearing, nor had to pay fines in most of the countries where citizens were harmed. He is accountable for the AI feature that he launched and promoted, for his negligence in not doing anything for months, for still allowing loopholes, and – to this day – not deleting all abusive content generated by his AI model, Grok.
ADVOCACY campaign
Citizens of any age, race, and gender are potential victims of AI-generated abusive content.
We need more independent and transparent content moderation, with the support of AI bots that analyze images and software used to nudify and sexualize citizens. AI watchdogs need to be embedded in the code, platforms, and chats to allow further investigation into instances of abuse.
AI Non-Consensual Sexualization and Nudify tools are a widespread phenomenon with many software and platforms involved, as well as several commercial and personal interests.
Read about the unsolved issues around abusive AI content.
It’s not only about Grok. Abusive AI content is becoming easier to create and more difficult to trace. It’s not only about new regulations; we must consider the scale, speed, and distribution of abusive content as well as the proliferation of Nudify tools. Even if big AI companies like Anthropic and OpenAI have certain policies, it is simple to bypass these guardrails using loopholes that allow anyone to generate pornographic images. It is also increasingly possible to train AI models at home using just personal computers.
However, the harm primarily lies in the circulation of content in group chats or on social media platforms, which amplify the abuse of AI-generated content. Automated AI detection can reduce exposure and distribution of the material, which should be promptly reported, notified, and removed.
AI-based content moderation can work if it is transparent and well designed. Platforms must also allow independent moderation by researchers and journalists, with more forensic analysis. Content found to be in violation should be removed automatically and immediately, then investigated to block users and notify victims. For this process, we need transparency into companies’ AI moderation practices and critical evaluation of their policies and automated content filtering. They need to actively tackle their loopholes and reduce opacity.
Additionally, we need to shift responsibility through new cultural and legal norms. The companies and developers who create or distribute tools that allow mass abuse must be held accountable and face criminal consequences. In particular, platform owners who allow exposure to abusive content must be held accountable, not just the users who use features that the platform promotes.
It’s also not only about AI. It’s an underlying cultural context concerning gender and social discrimination, lack of education, bad politics that shapes legislation, and general misunderstanding of the complex layers that really generate such violence.
Read the essay Generative AI-Violence for more cultural and political references.
Millions of survivors remain outside the protection of both legal systems and the platforms’ policies, especially in countries with less legal protection for gender abuse, hate crimes, and civil rights.
Follow the call for a global ban of AI nudify:
https://www.mentalhealtheurope.org/joint-statement-unifying-voices-worldwide-no-to-nudify/
Some current regulation and data on AI nudify:
A 2025
UNESCO AI Ethics study revealed that 92% of Nudify victims are women, exacerbating gender-based violence. A 2024
Stanford study found that 68% of deepfake content involved non-consensual nudity.
The EU’s 2025 AI Act classifies non-consensual nudification as a "high-risk" application, but enforcement remains inconsistent with the
EU AI Act. Victims report severe psychological harm, with 41% of cases involving minors, according to a 2024
Thorn Research report.
As of January 2025, 110 countries had no law against image-based abuse; 67 had what are considered “sufficient” laws — meaning they align with Article 16 of the
UN Cybercrime Convention — while 18 have laws that fall short of this standard, including the United States and Japan, according the
SWGfL’s resources. While the growing number of countries passing legislation protecting citizens against image-based abuse is a positive development, it is important to reflect on the limitations of such legislation.
Key news articles:
August 2026. “Elon Musk’s xAI used child porn to train Grok models, lawsuit says”
https://arstechnica.com/tech-policy/2026/08/elon-musks-xai-used-child-porn-to-train-grok-models-lawsuit-says/
July 2026. “Elon Musk’s xAI sues Minnesota over law banning ‘nudification’ technology”
https://www.theguardian.com/technology/2026/jul/29/xai-sues-minnesota-nudification-technology
June 2026. “EU Lawmakers Move to Ban AI Nudifiers, but Enforcement Remains Unclear”
https://www.techpolicy.press/eu-lawmakers-move-to-ban-ai-nudifiers-but-enforcement-remains-unclear/
April 2026. “Elon Musk snubs Paris legal summons over alleged child abuse images on X”
https://www.theguardian.com/technology/2026/apr/20/french-prosecutors-summon-elon-musk-over-alleged-child-abuse-images-on-x
Reports:
"Sexualized images on X: What we are doing to stop them and what we expect from the EU". January 2026, report by Oliver Marsh, Head of Tech Research Algorithm Watch.
https://algorithmwatch.org/en/sexualized-images-on-x/
"Grok Unleashed, Grok Generating Flood of Sexualized Images of Women, Including Minors, and Extremist Propaganda". January, 2026, Flash report by AI Forensics.
https://aiforensics.org/uploads/GrokUnleashed.pdf
“Grok floods X with sexualized images of women and children”. January 2026, report by Center for Countering Digital Hate (CCDH).
https://counterhate.com/research/grok-floods-x-with-sexualized-images/
“Consent, Not Nudity, Should Define Image-Based Abuse Laws”. Legal analysis by Eva Blum-Dumontet.
https://www.techpolicy.press/consent-not-nudity-should-define-imagebased-abuse-laws/
“All the legal risks that apply to Grok’s deepfake disaster”. Legal analysis by Derek B. Johnson.
https://cyberscoop.com/elon-musk-x-grok-deepfake-crisis-section-230/
