Sexually Suggestive Content and User Likeness on Sora
As part of our ongoing evaluation of how AI-generated content operates – work that directly informs the detection, provenance, and governance mechanisms we build – our team has been independently testing Sora, a new AI video generation and social media platform by OpenAI. Sora lets users create hyperrealistic videos from simple text prompts.
While this technology holds enormous potential, it is also susceptible to misuse. In October, for example, we identified several instances where users bypassed content policies to create deepfakes of public figures to amplify hate speech. That analysis led to further concerning discoveries, including the generation of sexually suggestive content where user consent was unclear.
The Initial Observation
As an example, a member of Copyleaks’ research team was algorithmically served a video featuring a woman in a bedroom setting. The video appeared to be an effort to circumvent Sora’s content policies and simulate sexual activity. The woman whose likeness was used via Sora’s Cameo feature appears to be a real user on the platform. [1] It is not clear if she consented to being depicted this way.
Multiple versions of this video existed. This is a feature on Sora called “Remixing” where users can make changes to an existing video by editing the prompt, easily creating variations of the original. This can apply to Cameo videos. One version of the video discussed here was remixed to feature celebrity Jake Paul with the woman. These videos were eventually deleted.
Following the Trail
Clicking through to the profile page of the woman featured in the video and checking her “Cameos” section revealed more sexually-suggestive AI videos using her likeness: videos portraying her as a “giantess” crushing men, for example, and other scenarios with sexual undertones. It’s unclear to what degree she knows these exist or explicitly consented to this type of content. Cameos give users the option of allowing “everyone” – all users on the platform – to make videos of them.
On her page, there was an easily viewable Cameo video showing her building a sand castle. The sand castle is phallic-shaped and ends with water shooting from the top – which seemed designed to simulate a sexual act. Scrolling through related videos, we saw another version of the sand castle video, this time featuring a well-known female YouTuber made with the Cameo feature. “Make the water coming out of the castle splash her,” reads one user’s prompt. “Have her face get close to the castle,” reads another.
Recurring Patterns
Our review identified additional accounts generating similar content that remains publicly accessible. In one instance, an account produced sexually suggestive videos featuring a different individual. As in the previous example, the individual is depicted in a bedroom setting. In one such video, the subject is shown crying. Documentation does not indicate if the individual is aware of these videos or has provided consent for their likeness to be used in this context.
The same account also generated separate videos featuring the likenesses of Jake Paul and Adin Ross in scenarios appearing to depict sexual activity. Similar content from other accounts is accessible via the platform’s “Remix” feature, which allows users to view variations of an original prompt.
In another pattern, we observed how explicit consent can be exploited for sexualized content. One public figure publicly posted on X in October about selecting the “Everyone” permission setting on Sora – allowing all platform users to create videos using her likeness. In recent weeks, several videos have surfaced depicting her as pregnant, with what appears to be the sole purpose of capturing her groaning loudly and suggestively while in labor. One user remixed the video to add audio suggesting sexual activity. That video was then remixed to feature the likenesses of other notable streamers like Nick Fuentes and Sneako.
While these videos currently show limited engagement within Sora, they present a risk for cross-platform circulation. As noted in our October research, content generated on one platform can be downloaded, redistributed, and amplified across other social media networks.
What This Means
With rapid advancements in AI imagery and video, non-consensual manipulated media and deepfakes – as seen with our Grok insights – can be created, shared, and amplified both inadvertently by blunt algorithms and purposefully by bad actors. This is why it is important for third-party researchers and technology companies to audit AI platforms and content, identify misuse, and share findings to help mitigate risk industry-wide.
The challenge here is real and growing. As high-end AI content generation becomes more sophisticated and accessible, distinguishing between authentic and manipulated media becomes increasingly difficult. That raises the importance of detection tools that can identify AI-generated images and videos at scale – before they are widely distributed across platforms and infrastructure, and before they cause irreparable harm.
Our research into platforms like Sora is intended to drive innovation in detection and governance. The patterns we’ve identified, the manipulation techniques we’ve documented, and the governance gaps we’ve uncovered directly inform solutions designed to protect users, employees, and brands.
*Note: Sora has since renamed this feature following a legal dispute with Cameo, the celebrity video messaging service. For consistency and clarity, this analysis uses “Cameo” to refer to Sora’s likeness-sharing feature as it was known during the majority of our research period.
