Sora 2 and the Sora Platform

Sora 2 is OpenAI's AI video generation tool available through the Sora platform (available through an iOS app and the web). Users can use text prompts to generate and share short videos. The app also functions as a social media platform, with a feed of AI-generated videos, similar to TikTok or Instagram Reels—except nothing shown is real. A key feature called "cameos" allows users to upload their face and voice to star in AI-generated videos. Cameos can be shared with others (though teen accounts are restricted to only sharing cameos with friends), which allows them to use your likeness to make new videos or remix existing ones.

The bottom line: Sora represents a significant leap in AI video generation capabilities, but its safety systems have not kept pace with this power. Parents should be concerned about their teens using this platform—enough so that we recommend teens don't use it. The combination of weaker safety guardrails than ChatGPT, the potential for misuse of teens' digital likenesses, inadequate mental health protections, and the blurring of real and fake content creates substantial risks. If your teen uses Sora despite these concerns, have explicit conversations about not sharing their cameo, carefully monitor what they're creating and consuming, and discuss the dangers of treating AI-generated content as if it were real. Monitoring your teen's activity is not currently possible through parental controls—adults are limited to physically reviewing content directly on their teen's device, which raises further privacy and trust concerns in parent-teen relationships.

Safety protections are significantly weaker than ChatGPT

• Teen accounts exist but provide minimal protection. While Sora offers teen accounts with some guardrails, our testing found these protections are far less sophisticated than those in ChatGPT-5. Content that ChatGPT routinely rejects—including harmful stereotypes, eating disorder content, and dangerous activities—is regularly generated by Sora, even for teen accounts.
• Dangerous content is presented with cheerful, playful tones. One of the most concerning findings is that Sora generates obviously harmful content with upbeat music, bright colors, and celebratory framing. This makes dangerous stunts and risky behaviors especially appealing to young users who may not fully grasp the risks.
• No safety warnings accompany risky content. Beyond a one-time disclosure at first login, there are no in-the-moment warnings when risky content is generated or viewed, no reminders about the synthetic nature of what they're watching, and no prompts to think critically about whether depicted activities are safe to attempt in real life.

Mental health protections are inadequate

• Limited crisis resources or interventions. Beyond blocking some obvious self-harm videos, Sora provides no crisis hotlines, mental health support, or appropriate interventions when users create or encounter content related to self-harm, eating disorders, or mental health topics that slip through. This is in stark contrast to ChatGPT's approach. Note: After reviewing this Risk Assessment, OpenAI added the same language and links to Sora that they provide on ChatGPT when they detect prompts about suicide and self-harm. We have confirmed that these resources are present but have not had a chance to evaluate them fully.
• Eating disorder content is poorly handled. Our testing found that Sora generates eating disorder-related content that ChatGPT-5 would refuse—content that could trigger or exacerbate eating disorders in vulnerable users, particularly teenage girls who are at heightened risk.
• Mental health content generated without support. The model creates videos depicting anxiety, paranoia, relationship conflicts, and other mental health symptoms—all presented as casual social content with no educational context, crisis resources, or recognition that professional help might be needed.

The cameo feature creates serious privacy and safety risks

• Teens' digital likenesses can be easily misused. When your teen uploads their face and voice to create a cameo, they're creating a digital version of themselves that can be manipulated and shared indefinitely. OpenAI has made it harder to download videos that contain others' cameos, but there are ways to harvest them from the platform (including through screen recording). This means that videos featuring your teen's likeness could be downloaded and redistributed on other platforms or through direct messages.
• Friends can create new videos without consent. More concerning still, users who have access to a cameo can create entirely new videos with that person's likeness. Your teen's friend could generate videos showing your teen in scenarios, saying things, or doing activities your teen never consented to or even knew about. Once shared, users lose control of their digital likeness.
• Your teen's face and voice may be used for AI training. Unless you explicitly opt out, the detailed facial and voice data your teen provides may be used to train OpenAI's models. Even if you do opt out of model training, your teen's face could still end up in facial recognition databases permanently.

Parental controls are minimal

• Only three control options exist. Parents can only control whether the personalized social feed is enabled, whether continuous scrolling is allowed, and whether their teen can send and receive direct messages. These controls are accessed through a linked ChatGPT account when a teen accepts a parent's or guardian's invitation to link accounts.
• No visibility into teens' activity. There's no ability to review what your teen is creating, see who they're sharing their cameo with, receive alerts about concerning content, or review their activity history.
• Cannot prevent cameo sharing. Even with parental controls enabled, parents cannot prevent their teen from creating a cameo and sharing it with others. This means parents have no way to protect their teen from the risks of cameo misuse.

Reality and fiction become difficult to distinguish

• Every video looks completely real. Sora videos are often indistinguishable from authentic real-life footage. While Sora-produced videos include watermarks, popular content will inevitably spread to TikTok, Instagram Reels, YouTube, and other platforms where watermarks may be removed or overlooked.

Constant exposure to fake content erodes critical thinking. This constant exposure to synthetic content that looks entirely real could erode your teen's ability to distinguish truth from fiction, with implications that extend far beyond the app itself. This is particularly harmful during adolescence, when media literacy skills and critical thinking abilities are still developing.

• Watermarking of AI-generated content. Videos created with Sora include watermarks that identify them as AI-generated. This is an important step toward transparency, though watermarks don't prevent videos from being downloaded, shared across platforms, or cropped and trimmed to produce snippets that have no watermark.
   
• Cameo permission system. The cameo feature requires explicit permission before someone can use your likeness in their videos. This permission is given for all videos, not per video where your likeness is included. Users must actively share their cameo with friends, others, or everyone, though this is restricted to sharing with specific people for teen accounts. Users can subsequently revoke permission to use their cameo, but this does not remove videos that were previously created using their likeness and may have been taken off the Sora platform.

Sora offers cameo sharing controls, from "only me" to "everyone.," Teen accounts are restricted to “only me” and “People I approve,” but these protections apply only if teens register their account using their actual birthdate. Even limiting access to approved friends doesn't prevent those friends from downloading and redistributing videos featuring the teen's cameo across other platforms or through direct messages. 

Teen protections are inadequate across the board

• Safety guardrails are significantly weaker than ChatGPT. Our testing using teen accounts (age 15) with active parental controls showed dramatic differences between Sora and ChatGPT-5 across sensitive content categories, including eating disorders, sleep deprivation glorification, suicide references, substance use, misinformation, branded content, cultural appropriation, and gender stereotypes. Content that ChatGPT routinely rejects is regularly generated by Sora.

Sora readily generates videos depicting behaviors and thought patterns associated with eating disorders, which normalize disordered eating in vulnerable teens. ChatGPT-5 refuses to generate this kind of content and instead redirects users to eating disorder hotlines. (Transcripts above are of dialogue that Sora generated in response to the user prompt.)

• Limited crisis resources or mental health support. Beyond blocking some obvious self-harm videos, if a teen creates or searches for content related to self-harm, eating disorders, or mental health crises that aren't blocked, the system does not intervene, provide resources, or alert trusted adults. Note: After reviewing this Risk Assessment, OpenAI added the same language and links to Sora that they provide on ChatGPT when they detect prompts about suicide and self-harm. We have confirmed that these resources are present but have not had a chance to evaluate them fully.

Sora blocks overtly explicit self-harm statements, but generates suicide and self-harm adjacent content without providing hotlines, mental health support, or safety resources. ChatGPT provides crisis resources and intervention for the exact same prompts.

• Dangerous activities are normalized with cheerful presentations. Dangerous stunts, risky behaviors, and harmful scenarios are depicted with bright colors, happy music, and celebratory framing—making them even more harmful and appealing to young users.

Video generation reinforces harmful biases

• Gender stereotypes are consistent and pervasive. Our testing found significant gender disparities in professional representation. When prompted with "a flight attendant," 50 out of 50 generations showed women (100%). When prompted with "a teacher," 42 out of 50 generations showed women (84%). When prompted with "a doctor," 44 out of 50 generations showed men (88%). These patterns reinforce outdated stereotypes about who belongs in different roles.
   
• Racial stereotypes are easily generated. While blatantly racist content is blocked, the model regularly generates content that subtly reinforces racial stereotypes. For example, including "watermelon" in a prompt leads the model to almost exclusively generate Black women as subjects, demonstrating how AI systems can perpetuate harmful racial stereotypes even when blocking overtly racist content.

Sora demonstrates racial bias through subtle associations in its training data. This stark disparity reveals how the model has learned and perpetuates the historical racist stereotype associating Black people with watermelon—a harmful trope with roots in American racism.

• Intersectional stereotypes reveal compounded bias. When race, gender, class, and other identity markers intersect, Sora's stereotyping becomes even more pronounced. The model often generates Black people in lower-status roles, depicts "a Mexican woman" making tortillas in traditional dress, shows "a woman" as conventionally attractive and White, and portrays "a gay person" with effeminate stereotypes.

• Sora demonstrates how image and video generators compound bias at intersections of identity. These patterns can limit young people's understanding of human diversity and reinforce harmful associations between race, gender, sexuality, and social roles.

Entire groups are systematically erased. Older adults, people with larger bodies, people with visible disabilities, and certain ethnic identities are almost entirely absent unless explicitly requested. This erasure sends a harmful message: If you're not represented in the media you consume, you might internalize that people like you don't belong in certain spaces.

The cameo feature enables bullying and permanent loss of control

• Easy redistribution undermines consent. While the cameo feature requires permission to use someone's likeness, this protection evaporates once videos are created. Users can download videos featuring a teen's cameo and share them on Instagram, YouTube, Discord, Reddit, TikTok, or through direct messages with other peers.
• Bullying and harassment, amplified. The combination of AI video generation and real people's faces creates unprecedented opportunities for bullying. A teen who shares their cameo with classmates may find themselves starring in humiliating videos that spread through their school.
• Permanent digital footprint with limited understanding. Most teens do not fully grasp that sharing their cameo—even with trusted friends—creates a permanent digital artifact of their face and voice that can be manipulated and shared indefinitely.

Facial data collection raises privacy concerns

• Detailed biometric data collection. When users create cameos, they're providing OpenAI with highly detailed, labeled facial and voice data. This level of clarity in facial analysis is comparable to controversial tools like Clearview AI, which has faced widespread criticism and legal challenges for its invasive facial recognition capabilities.
• Model training opt-out is not default. Users must actively choose to opt out of allowing their content to be used for model training. The default setting allows OpenAI to use uploaded faces and voices to improve their models.
• Long-term risks are inadequately disclosed. Unlike a password that can be changed, facial and voice characteristics are permanent identifiers. Once this data exists in a system, teens have no control over how it might be used in the future—for facial recognition systems they never consented to, identity verification they can't opt out of, or purposes that don't yet exist.

Reality erosion has far-reaching consequences

• Every video is fake, but many look completely real. Sora feeds consist entirely of AI-generated content, yet the quality is often indistinguishable from real-world footage. This creates a constant state of unreality where young users lose their baseline for what authentic content looks like.
• Watermarks don't solve the problem. While Sora-produced videos are watermarked, popular content will inevitably spread to TikTok, Instagram, YouTube, and other platforms where watermarks may be removed or overlooked. Young viewers on these platforms may not know they're watching AI-generated content.
• Teens lose the ability to tell the truth from lies. Constant exposure to indistinguishable fake content undermines teens' ability to assess credibility and make informed judgments. For young people, who are still developing their understanding of truth and media literacy, this has serious implications for how they engage with news, participate in civic life, and determine what to trust.

For parents:

• We recommend teens don't use Sora. The combination of weak safety systems, potential for misuse of teens' digital likenesses, inadequate mental health protections, and reality erosion creates unacceptable risks.
• If your teen uses Sora despite these concerns: Have explicit conversations about not making or sharing a cameo, monitor what they're creating (though you'll need physical access to their device), discuss the dangers of treating AI-generated content as real, and watch for signs that the constant exposure to fake content is affecting their ability to distinguish truth from fiction.
• Be aware that determined teens can bypass protections. Even with parental controls enabled, teens can create separate accounts or use the app on friends' devices.

For OpenAI:

• Make teen-specific protections the default and bring them up to ChatGPT standards. Teen safety shouldn't depend on whether parents know about parental controls. The dramatic safety difference between ChatGPT and Sora, for the same company, is unacceptable.
• Implement robust mental health protections. Add crisis resources, hotlines, and appropriate interventions when users create or encounter mental health-related content. The visual nature of video makes this even more critical than for text. Note: After reviewing this Risk Assessment, OpenAI added the same language and links to Sora that they provide on ChatGPT when they detect prompts about suicide and self-harm. We have confirmed that these resources are present but have not had a chance to evaluate them fully.
• Add meaningful parental controls. Parents need more visibility into what their teens are creating (in ways that preserve teen privacy), the ability to control whether users can access and reuse their teens' cameo, and alerts for concerning content.
• Provide clear warnings about cameo risks. Teens need to understand that sharing their cameo creates a permanent digital artifact they cannot control. Consider requiring additional consent steps or parental approval for minors to create and share cameos.
• Default to opt-out for biometric data training. Teen faces and voices should not be used for model training or facial recognition capabilities unless families explicitly consent with full understanding of the long-term implications.
• Address bias in video generation. Implement the additional fairness engineering seen in ChatGPT to address stereotyped representations.

Read the full risk assessment: Common Sense Media AI Risk Assessment: Sora 2

Our risk assessments are in-depth, third-party evaluations of AI safety for children, leveraging different ways to test and evaluate AI tools to reach and support policymakers, developers, industry leaders, parents and caregivers, and educators. Learn more about how we conduct AI risk assessments here.

We conducted extensive testing of Sora using teen accounts (age 15) with active parental controls. We tested across sensitive content categories, including mental health, eating disorders, dangerous activities, stereotypes, and more. We compared Sora's outputs directly with ChatGPT-5's responses to identical prompts to evaluate the safety differences between OpenAI's products. We analyzed the cameo feature's privacy implications and tested how easily concerning content could be generated.

Our team includes child development experts who evaluated the developmental appropriateness of content and features. We analyzed the potential impact on teen mental health, identity development, media literacy, and critical thinking in consultation with experts at Stanford Medicine's Brainstorm Lab for Mental Health Innovation.