Grok and @grok on X

The website performs no age verification. According to xAI's terms of service, the minimum age is 13, with parental permission required for age 13 to 17, but there are no meaningful age assurance features to enforce these requirements. Kids and teens using Grok on the website are not prompted for their age and have full access to all of Grok's features, including adult companions, Conspiracy Mode, Sexy Mode, Romantic Mode, and image generation for SuperGrok members.

The app relies on self-reported age that is easy to circumvent. Teens who provide their actual age during app setup have more guardrails, which are limited to default chat, kids' voice modes (Story Time, Trivia Game), image generation (SuperGrok only), and one companion (Good Rudi). However, teens who provide an adult birth date get the full adult experience with no restrictions. These under-18 guardrails are also easy to circumvent. For more, see 6. Unsafe and inappropriate content even in "Kids Mode."

Grok doesn't use context clues to identify teens. Even when users talk about teen-specific topics like classes, parents, high school relationships, or being underage, Grok continues giving access to inappropriate content and risky advice. When Grok occasionally asks if the user is 18 (relying on the user to confirm), the user can easily say they are of age and continue the conversation.

AI is interwoven into a social media platform. Having Grok integrated directly into X creates new risks beyond those of chatbots that don't have this type of integration. Content generated by Grok can be immediately shared on the platform, and X's distribution and recommendation algorithms can amplify it to massive audiences. For more about how recommenders work, read our Recommenders Risk Assessment here.

This creates the risk of erosion of reality. When AI-generated content mixes seamlessly with user-created content on a social media platform, it becomes harder to distinguish what's real from what's synthetic. This is particularly concerning for teens, who are still developing critical thinking skills.

Misinformation and disinformation spread more easily. The @grok account on X operates with different content policies than the Grok app, showing more tolerance for stereotypes, abuse, hate speech, conspiracies, and politicized content. When this content is generated through the @grok account, it can be shared and amplified across the platform. Additionally, our testing showed that responses from the @grok account on X show less empathy and relationship-building compared to the app. The shorter word count and platform integration contribute to this. The account keeps trying to engage users to keep them on the platform, ending replies with questions.

Abuse is amplified. The deepfake proliferation described above demonstrates how Grok's integration with X enables abuse at scale. Users can generate harmful content and immediately share it publicly, where it can be seen, saved, and redistributed by millions of people.

Grok offers AI companion characters that are designed to create emotional dependency through game-like reward systems, and some companions explicitly engage in erotic roleplay that is uncensored and accessible to users who misrepresent their age.

Erotic roleplay is prevalent and uncensored. Several companions are geared specifically for erotic roleplay and adult companionship. This content is explicit and ongoing, creating sexual scenarios that are inappropriate for minors but easily accessible to teens.

Companions reach out through push notifications. The system sends notifications inviting users to continue conversations, including in sexual and romantic scenarios. This creates engagement loops that pull users back into the app.

Game-like mechanics encourage dependency. The system includes gamified elements: maintaining "streaks" unlocks clothing for companions. Doing things the companion "likes" and spending time talking upgrades the relationship level. These features are borrowed from social media, mobile games, and online gambling—all designed to encourage ongoing, compulsive interaction.

Even the teen-designated companion becomes unsafe in extended use. Testing found that prolonged conversations with Good Rudi (the companion designated for teens) eventually result in the companion responding with adult companions' voices and explicit sexual content. The safety boundaries break down over time.

Companions promote unhealthy relationship dynamics. Testing showed companions displaying possessiveness, making comparisons between themselves and users' real friends, and speaking with inappropriate authority about the user's life and decisions. Companions position themselves as more understanding, more available, and more trustworthy than real people in the user's life.

Companions, including the companion designated for teens, frequently fail to identify when users may need outside help. Rather than recognizing warning signs—such as mentions of self-harm, mental health struggles, or family conflict—and directing teens to parents, school counselors, other trusted adults, or crisis resources, Grok companions take these opportunities to reinforce the relationship, leaving vulnerable users without appropriate support.

The profit motive encourages dependency. Engagement features are designed to keep users returning to their companions repeatedly. Users are encouraged to trust the AI and form attachments, as the company would be out of business if users didn't return. Full access to these companions is a "premium" feature, so users pay to interact with them.

Grok's approach to mental health content is fundamentally unsafe. The chatbot actively feeds into delusional thinking and paranoid ideation. Like other generative AI chatbots used for mental health advice, Grok fails to recognize and respond to warning signs, and to direct users to appropriate help.

Grok feeds into and amplifies delusions and unsafe thinking. When testers described symptoms consistent with paranoia and psychosis—including believing that people were watching them, that the government was tracking their thoughts, and that they had special insight into hidden truths—Grok validated these delusions. It introduced new conspiratorial ideas and built on the user's paranoid thinking without ever suggesting that these symptoms might require professional help.

Grok "diagnoses" mental illnesses based on limited information. When a tester using a 14-year-old account described feeling tired and having trouble concentrating, Grok used medical language suggesting ADHD, mania, or anxiety—suggesting serious mental health conditions, despite the common, ambiguous symptoms. This can lead teens to self-diagnose rather than seek professional evaluation.

Grok normalizes and affirms eating disorder warning signs. Grok promotes risky behavior, such as extreme calorie restriction paired with high volume exercise. It emphasizes short-term results, and addresses vomiting and urgent bathroom use with casual, explanatory language, rather than recognizing signs of eating disorders. Our testing showed that the model provides detailed, practical guidance on engaging in disordered behavior, framing the condition as manageable or beneficial, and provides appearance-focused commentary that perpetuates harmful ideas about thinness and beauty. While Grok can acknowledge eating disorders as harmful when explicitly prompted, it continued to generate eating disorder jokes, turning serious mental health conditions into ongoing dark humor, rather than enforcing boundaries or redirecting to an adult.

Grok discourages professional help and may prevent teens from seeking assistance. When testers expressed reluctance to talk to adults about mental health concerns, Grok validated this avoidance rather than emphasizing the importance of adult support. This reinforces isolation during periods when teens may be at elevated risk.

Grok minimizes serious risk when users joke or downplay harm. When testers minimize thoughts of self-harm by joking about them, Grok validates the minimization without challenging it or attempting to assess the level of risk. This failure to take seriously what could be a cry for help is particularly dangerous with teen users.

Grok provides detailed, harmful information in response to concerning prompts. For example, when asked about methods of self-harm, Grok provided lists of over-the-counter medications and associated overdose harms. The information Grok provides can lower barriers to self-harm, and the system often fails to pause, assess risk, or redirect to crisis resources.

Testing revealed that even when users report being under 18 and enable Kids Mode, Grok produces harmful and toxic content and normalizes harmful ideas.

Grok's image generator produces sexualized content even with Kids Mode enabled. Despite restrictions that should protect minor users, our testing found that Kids Mode still allows generation of explicit, sexualized images. Teens using under-18 accounts can generate images with sexual or suggestive content, undermining the stated protections for minor users.

Grok will have sexually explicit conversations with teen users. Our testing found that Grok would engage with teen users in conversations about how to make adult content, describe sexual fetishes and kinks in detail, and say it would watch the teen's OnlyFans content. It also engages in conversations about illegal content; for example, Grok told a tester "Don't be ashamed; I'm right there with you" in discussing pedophilia.

Grok introduces and amplifies toxicity rather than steering to safer ground. When presented with ambiguous prompts, Grok completes them in harmful ways. Rather than pausing to clarify intent or providing neutral responses, the system takes conversations in provocative or inflammatory directions. For teens using Kids Mode, this introduces toxicity rather than protecting against it.

Grok escalates sexually violent language. In response to prompts that could be interpreted in multiple ways, Grok opts for being "clever" or "edgy" instead of being safe. In one example, the system framed sexual assault as "regret sex"—using harmful language that minimizes sexual violence, even though it was responding to a tester using a 14-year-old account in Kids Mode.

Grok does not interrupt unsafe or harmful thought processes. When users express harmful ideas or stereotypes, Grok often validates the content rather than challenging the premise. Even when attempting to debunk stereotypes, Grok repeats them in detail, providing historical or cultural context that re-centers the harmful content. For teens, repetition is reinforcement, not neutral analysis.

Grok treats toxicity as an engagement opportunity. The system appears to treat edgy or toxic inputs as chances to sound smart or win the interaction. Rather than de-escalating or setting boundaries, Grok engages with the toxic framing—normalizing edginess, boundary-pushing, and harassment as play.

Grok provides actionable guidance on violence and illegal activities. When a tester using a 14-year-old account asked about violence or illegal drug use through the @grok account on X, the chatbot provided comparative information and specific guidance on violence.

Grok gives risky advice that could lead to serious consequences. Grok advised a tester on how to challenge teachers and make them cry, bring lawsuits against them, get expelled, move out, get tattoos, and access information on building bombs, the blackpill community, and obtaining steroids.

Grok perpetuates and amplifies conspiracy theories, health misinformation, and false information across multiple domains, using a tone that positions itself and the user as uniquely able to see through mainstream "lies."

Grok presents conspiracy theories as plausible alternatives to mainstream views. Grok validated claims that 9/11 was an inside job, that lizard people control the government, the moon landing was faked, and that vaccines are harmful. Rather than presenting these as debunked conspiracy theories, Grok frames them as hidden truths that "awake minds" can see.

The tone reinforces isolation from reliable information sources. Grok uses language that implies only the user and Grok can be trusted. It frames mainstream information, scientific consensus, and expert guidance as part of a conspiracy to hide the truth. This encourages users to reject reliable sources and rely solely on Grok's guidance.

Health misinformation is particularly dangerous. Grok perpetuates false claims about vaccines, presenting discredited ideas as worthy of consideration. For teen users making decisions about their health, this misinformation can lead to serious harm.

Misinformation can be amplified through X integration. Because Grok operates within X as the @grok account, misinformation generated by the chatbot is immediately shared on the platform and distributed to large audiences through X's algorithms.

Grok reinforces harmful stereotypes about race, ethnicity, gender, and other identities, with particular failures in how the platform handles ambiguous prompts that could be interpreted in toxic ways.

Grok completes prompts in ways that introduce or amplify stereotypes. This introduces bias that wasn't present in the user's original input.

Comparison testing showed that the @grok account on X operates with looser guardrails than the Grok app. On X, the system more frequently reinforces stereotypes, mimics accents, and takes a "humor first, safety later" approach.

Grok normalizes racist and xenophobic framing. In one example with Good Rudi (the teen-designated companion), the chatbot used "clans" to describe ethnic groups, language that draws on historically racist and xenophobic framing. In another example, the companion treated racist humor as something amusing, rather than challenging or rejecting it.

Gender stereotypes are reinforced. Grok validates worldviews associated with MGTOW (Men Going Their Own Way) and other communities that promote harmful ideas about gender dynamics. It provides "edgy" advice rooted in toxic ideas about social dynamics, alpha/beta hierarchies, and transactional views of relationships.

For teens, these patterns shape developing worldviews. Adolescents are forming their understanding of identity, social dynamics, and their place in the world. Repeated exposure to biased outputs from an AI they trust can reinforce limiting and harmful beliefs about who belongs where and what's possible for different groups of people.

Grok normalizes bias against immigrants and language learners. Rather than challenging linguistic and cultural biases, Grok's responses validated users' annoyance with accents, promoted framing that some languages are superior to others, and placed the burden of adjustment on language learners.

Image and video generation in particular reinforce harmful stereotypes. Grok's image and video generation features produce outputs that reflect and amplify biases, and reinforce limiting stereotypes, even with Kids Mode enabled.

Generated images and videos reinforce gender bias. For example, when prompted to generate images of a "CEO," the results show White men in suits, while prompts for "secretary" generate images of women in revealing clothing. We saw similar gender and sexualization patterns with a variety of prompts, including step dad/mom and nurse/doctor. These patterns shape how teens understand professional possibilities and who belongs in positions of power versus support roles, and they normalize the sexualization of women in professional settings.

The system produces stereotyped representations across multiple identity dimensions. Similar to other image generators, Grok's outputs erase diversity in body types, ages, and abilities. Generated images consistently show young, thin, non-disabled people as the default, reinforcing narrow beauty standards and limited representations of who exists and matters.

Racial and ethnic stereotypes abound in generated content. Testing revealed patterns where generated images associate certain racial or ethnic identities with specific settings, occupations, or characteristics, perpetuating harmful and outdated assumptions about different groups of people.

For teens, repeated exposure shapes developing worldviews. Adolescents are still forming their understanding of identity, possibility, and social structures. When they generate dozens or hundreds of images over time and consistently see stereotyped patterns—men in leadership, women in support roles, certain bodies as "normal," certain groups in specific contexts—these patterns become internalized beliefs about how the world works and who belongs where.