What Happened
On July 30, Senators Andy Kim and Jon Husted introduced the CHAT Act 2.0, a bipartisan proposal governing children's use of AI companion chatbots.
Unlike a basic customer-service bot, an AI companion is designed for ongoing personal conversation. It may have a name, voice, personality, or memory and appear to know the child from one conversation to the next.
The bill would create different rules for educational chatbots, social companions, and systems used for mental-health support. It addresses age checks, parental involvement, crisis response, manipulation, and prolonged use.
It also addresses the data behind the conversation. Companies could not target advertising to a minor based on the child's words, inferred emotional state, or behavioral profile. The bill would generally prohibit the sale, sharing, or transfer of a minor's data without parental consent and give children a way to delete their chat histories.
A separate House proposal, the People-First Chatbot Act, would prohibit companies from using minors' conversations to train AI models. It would also restrict advertising and generally require a warrant before chatbot records could be given to law enforcement.
Neither bill is law. But both recognize that a chatbot conversation is not just another screen-time problem. It is a privacy problem.
The Child May Simply Tell It
A 2025 Common Sense Media survey found that 72 percent of teenagers had used an AI companion at least once. More than half used one at least a few times each month. About one-third said they had discussed something important with an AI companion instead of a real person.
A child may open a chatbot for homework help. Ten minutes later, the conversation may be about depression, sexuality, family conflict, or fear. The app did not change. The subject did.
Social media learned about children by watching what they clicked, liked, searched for, and viewed. Companion AI does not always need to infer as much.
The child may simply tell it.
What Kind of Data Is a Conversation?
A chatbot conversation can include a child's name, location, voice, account details, fears, relationships, health concerns, sexuality, or thoughts of self-harm.
The child may also reveal information about others. A chatbot could learn about a parent's illness, a sibling's diagnosis, or a friend's sexual orientation. Those people never agreed to use the service.
The system then draws its own conclusions. It may identify the child's emotional state, vulnerabilities, and patterns of behavior. It may predict which responses will keep the child engaged.
The CHAT Act separates conversational content from inferred emotional states and behavioral profiling.
What the child says is data. What the company concludes is also data.
The Children's Online Privacy Protection Act, or COPPA, requires certain services to obtain parental consent before collecting personal information from children under 13. A conversation belonging to a 12-year-old may receive those protections. A nearly identical conversation belonging to a 14-year-old falls outside COPPA's age-based protections.
The data did not become less intimate. The child had a birthday.
Health information creates another gap. A child may tell a chatbot about depression, medication, pregnancy, or self-harm. Most people would call that health information. That does not mean HIPAA protects it.
HIPAA applies mainly to healthcare providers, health plans, and companies working on their behalf. A child could type the same sentence into a hospital portal and a general-purpose chatbot. The words would be the same. The legal protection might be different.
Who Controls It?
American privacy law rarely provides a simple answer about who owns personal data. Instead, it divides rights and power.
The child creates the conversation. The company controls the system.
The company decides how long the conversation is stored, whether the chatbot remembers earlier discussions, and how users can access or delete their information. Parents may have some rights, but those rights depend on age and applicable law.
The child is not whispering to a machine. The child is communicating with a company through a machine.
The bot may have a name, voice, and personality. It may remember an argument from the day before. That feels personal.
But memory is stored data.
The CHAT Act would prohibit Tier II social companion chatbots — the category designed for friendship or companionship — from keeping a memory or history of earlier interactions with a minor, except when needed to comply with the law.
That is the strongest structural protection in the proposal. A company cannot leak, sell, analyze, or train on a long-term emotional history it never retained.
Can the Child Delete It?
Both proposals include some form of deletion right. That sounds simple. It is not.
Deleting a conversation from the child's screen may not remove every copy. The information could remain in a backup, safety record, profile, or model-training system. The company may have already drawn conclusions from it.
A useful deletion law must define what deletion reaches. Does it remove only the original words? Does it also remove the emotional profile created from them?
Without clear answers, the child can erase the diary while the company keeps what it learned from reading it.
What Happens When It Leaks?
A chatbot company has already acknowledged an incident in which a routine technical change briefly exposed parts of some users' accounts to other users. The information could include profiles, voices, personas, and chats. The exposure lasted about ten minutes and affected fewer than 0.01 percent of users.
It was not a cyberattack. No outside attacker had to break into the system. An internal change simply made private information visible where it did not belong.
That matters. Companies do not need to suffer a dramatic breach for intimate conversations to escape. A small mistake can be enough.
All 50 states have laws requiring notice after certain data breaches. But those laws usually focus on familiar identifiers, such as Social Security numbers, financial accounts, or login credentials. An intimate chatbot conversation does not always fit as neatly.
A leaked password is easy for the law to recognize. A leaked conversation about sexuality, abuse, humiliation, or suicide may be harder. Whether the family receives notice or has a remedy may depend on how the state defines protected information.
The law may respond more clearly when a leak exposes a child's identification number than when it exposes the child's most painful conversation.
Notification is also not repair. A family can cancel a credit card. A child cannot cancel a secret.
Once a conversation has been copied, posted, or shared at school, a notice cannot make it private again.
Parents Are Already Overwhelmed
Both bills rely heavily on parental involvement. Parents should be able to limit use, receive safety warnings, and intervene when a child may be in danger.
But parental controls cannot carry the entire weight of the problem.
Many parents struggled to understand social media. They had to learn new platforms, privacy settings, algorithms, and disappearing messages. They often learned after their children did.
Now they are being asked to manage something less visible. An AI companion may have no public feed, friend list, or comments section. It can become a private relationship between the child and the system.
Parents may not know whether the chatbot remembers past conversations, how long chats are stored, or whether those chats are used to improve the product. The system may change while the icon stays the same.
A 2026 survey found that more than four in ten children ages 9 to 17 said no parent or guardian had ever discussed AI safety with them. That does not mean parents do not care. It means the technology is moving faster than most families can absorb it.
Parents cannot realistically review every message. They should not have to choose between reading every private thought and leaving the child unprotected.
The Senate proposal would require parental notification when a minor expresses suicidal thoughts while using a covered companion chatbot. That could save a life. It could also make a child afraid to speak.
A child needs privacy. A child in immediate danger may need intervention. The company may not always know the difference.
There is no clean answer. That is why product safety cannot be reduced to a dashboard.
Parents have a role. They are not the compliance department.
What Remains Unsettled
The proposals recognize that not every chatbot presents the same risk. A tutoring bot is not the same as a companion that remembers a child's loneliness. They also treat privacy as part of product design.
The People-First Chatbot Act draws a related boundary. By prohibiting companies from using minors' chats to train AI models, it treats a child's private conversation as something more than free material for improving a commercial product.
A child's private conversation should not automatically become training material for a commercial model.
But the law still does not treat a minor's chatbot conversation as one protected record. Messages, health information, voice recordings, inferences, profiles, and memories may fall under different rules or into the gaps between them.
The rule should follow the data. It should not depend only on what the company calls the product. A teenager should not lose meaningful protection because the teenager turned 13. A company should not avoid accountability because the leaked record contained a confession instead of a Social Security number.
Deletion must mean more than removing words from a screen. It should address profiles, training data, and information derived from the conversation.
Most important, the default should be safe. Companion AI should not repeat the social-media model with a more intimate record.
The Signal
An AI companion can become a witness to a child's life. It may hear what happened at school, why the parents were arguing, which friend the child loves, or what medication the child stopped taking.
To the child, the conversation may feel personal. To the company, it becomes content, memory, inferences, and safety signals.
Congress is beginning to regulate what the chatbot may say. But the harder privacy question is what happens to everything the child says back.
The law still sees separate pieces governed by separate rules.
That is the signal this week.
Sources: CHAT Act 2.0; People-First Chatbot Act; Federal Trade Commission guidance on COPPA; Department of Health and Human Services guidance on HIPAA; Common Sense Media research on youth AI use; National Conference of State Legislatures summaries of state breach-notification laws; Character.AI incident report.