Brazil gave families four ways to see AI. The safety advice needs a complaint path

THE SIGNAL IN ONE SENTENCE

A child does not meet artificial intelligence as a policy category. They meet a toy that talks back. They meet a character that remembers yesterday's conversation. They meet a feed that quietly decides what comes next. They meet a tool that can turn one photo into another and one rumor into a convincing voice note. Brazil's new federal guide for children, teenagers, families and schools starts from that ordinary reality. The Ministry of Justice and Public Security and the Ministry of Human Rights and Citizenship organize AI into four roles: toy, company, curator and creator. That is a smart piece of translation. The plain signal is this: Brazil has made the risks easier to recognize in daily life. Now families need an equally clear map from recognizing a problem to filing a complaint, preserving evidence, getting an answer and securing a remedy. The guide was announced by the Justice Ministry on October 15 and made available through the Human Rights Ministry on October 16. The ministries describe it as the federal government's first document aimed at helping this audience use AI responsibly. The publication is guidance, not a new enforcement action. It was developed as part of the Brazilian Artificial Intelligence Plan with cooperation from UNESCO. The government says its legal foundation includes the ECA Digital, Law 15,211 of 2025, its implementing decree and international child-rights guidance. The four-part frame is the useful bit. AI as toy covers connected objects that listen, speak and respond. The risk is not simply that the toy gives a silly answer. A microphone, account, profile and remote service may sit behind the plastic face. Families need to know what is recorded, where it goes, how long it stays and whether the device can be used without building a behavioral file. AI as company covers conversational apps and characters that simulate friendship or emotional support. The guide addresses emotional attachment, mental health and body image. That is important because a child can experience a fluent reply as care even when the system has no duty of care, no reliable memory and no understanding of the child's situation. Warm language is not a safeguarding credential. AI as curator covers recommendation systems that decide which posts, videos and games rise to the surface. This is where attention becomes infrastructure. A young person may feel as if they are freely exploring while the system is learning which fear, joke, body ideal or controversy keeps them watching. AI as creator covers tools that generate or alter text, images, audio and video, including deepfakes. The guide points to privacy, critical thinking, sextortion and nonconsensual intimate imagery. Here the harm can move faster than an adult meeting, a school calendar or a platform support queue. The four roles also overlap. A companion can recommend content. A creative tool can sit inside a toy. A curator can produce synthetic previews. A classroom assistant can feel like a companion while collecting prompts and grading patterns. So the categories should be treated as doors into the problem, not sealed rooms. The ministries say the guide offers different recommendations for three audiences: children and teenagers, family members and caregivers, and teachers, educators and school administrators. That division matters. It avoids pretending that a seven-year-old should carry the same burden as a platform, a school or a public authority. But advice becomes much stronger when every warning has an owner. Suppose a student receives a sexualized synthetic image of a classmate. The immediate questions are practical. Should the student save a screenshot or avoid creating another copy? Which adult should be told first? Does the school preserve the device or the link? Which platform report category fits? When should the police, child-protection service or a specialized hotline become involved? What confirmation should the family receive after filing? Those questions are not an argument for turning every incident into a police case. They are an argument for triage. The official summaries list several support and reporting resources: Disque 100, the CVV at 188 and SaferNet Brasil's help channel. For nonconsensual intimate imagery, the guide also points to Take It Down and StopNCII, services designed to help prevent distribution across participating platforms. That is useful routing information. It is not yet the same as one end-to-end complaint path. A directory tells someone where doors exist. A complaint path says which door fits which harm, what information is safe to collect, what response time to expect, how to escalate silence, who can order removal and what remedy is available when the damage cannot be fully erased. Families need both. The government could build the next layer as a simple public decision tree. Start with the type of harm: unwanted data collection, emotional crisis, discriminatory recommendation, impersonation, sexual image, threat, extortion or school misuse. Then route the person to the correct first responder and show the handoff to the next institution. Every step should produce a receipt. A receipt does not need to expose a child's private story. It can be a case number, date, destination, expected next step and safe method for checking progress. The point is to prevent a parent from repeating the same traumatic account to five desks while nobody owns the clock. Schools need their own version. A teacher should know when to preserve a URL, when not to forward an image, how to contact a caregiver safely, how to protect the targeted student from rumor and punishment, and which staff member owns external reporting. School systems should record resolution time without publishing a child's identity. Platforms need a visible lane too. The guide can tell children to report harmful material, but a report button is not a remedy. The platform should confirm receipt, explain the policy category, communicate whether content or an account was restricted and offer an appeal when the reporter or accused user believes the decision is wrong. Public agencies should publish the gaps between those systems. If a platform says an image does not violate its policy but a school identifies harassment, who coordinates? If a family uses a global hash-matching service but the content remains on a small site, who helps with the next step? If a companion app triggers a mental-health crisis, which records can be preserved without exposing more private conversations? Those are implementation questions, not reasons to abandon the guide. The guide arrives because AI use among young Brazilians is already ordinary. The Justice Ministry says 65 percent of children and teenagers who use the internet had used generative AI in the 2025 TIC Kids Online Brasil research. It reports study as the leading use at 59 percent, followed by information searches at 42 percent, content creation at 21 percent and conversations about personal problems or emotions at 10 percent. The last figure is smaller than the others and still deserves attention. One in ten is not a corner case when the subject is private distress. A conversational system can be available at midnight, answer without embarrassment and feel easier than telling an adult. Those qualities can help a young person put words around a problem. They can also create false confidence in a response that is generic, mistaken or unable to act. The safer message is not never talk to a machine. It is know what the machine cannot promise, and know which human can take responsibility when the conversation becomes serious. Families can use the guide as a household inventory. Ask which devices listen, which apps chat, which feeds recommend and which tools create. For each one, write down the account owner, age setting, privacy controls, reporting route and trusted adult who can help. That conversation should not feel like a courtroom deposition. Children will hide problems if every disclosure costs them the device, the game and their independence. A useful family rule separates asking for help from automatic punishment. The first goal is safety. The second is understanding what happened. Consequences, if needed, come after the child knows that reporting was the right move. Educators can do the same exercise with classroom tools. Before assigning an AI product, document what data enters it, whether the provider uses prompts for training, whether students can opt out, what accessibility support exists and how errors are challenged. A creative writing demo is easy. A complaint about a generated stereotype is the real governance test. Developers serving young users should map features to the four roles. If a product acts as company, test attachment cues and crisis handoffs. If it acts as curator, test repeated exposure and explain why items appear. If it acts as creator, make provenance and reporting visible. If it acts as toy, minimize collection and give adults controls that do not require a privacy law degree. Then publish the complaint path beside the feature. Brazil's framing can travel beyond Brazil because it starts with what a child sees rather than how a vendor describes its stack. A model may change. A marketing label may change. The everyday roles are likely to persist. That makes the guide a strong front door. The next job is plumbing. Connect the child to a trusted adult, the adult to the right institution, the institution to a deadline and the deadline to an appeal or remedy. Show the path before an emergency, test it with schools and families, and publish anonymous performance totals so the public can see whether reports actually move. Four ways to see AI are a useful beginning. The fifth picture should show who answers when something goes wrong.