Companion Chatbots Under the Microscope: Privacy Laws Are Coming for How Chatbots Remember Your Kids

If you build a product that remembers things about its users—and especially if those users might be minors—the privacy landscape shifted under your feet sometime in the last twelve months. A patchwork of state laws and federal proposals now specifically targets companion chatbots: AI systems designed for social, emotional, or personalized interaction. Most of the press coverage focuses on disclosure rules and crisis protocols. That matters, but it's not the hard part. The hard part is memory—what your system retains about a minor, how long it keeps it, and whether a parent or the minor can make it actually disappear.

This piece walks through the specific bills, what they require, where they diverge, and what the gaps mean if you're building in this space.

Key Takeaways

• California's SB 243 and New York's companion-chatbot law are now in effect; both focus on disclosure and break reminders but say almost nothing about data retention or memory deletion for minors.
• Illinois's pending SB 3368 is the first bill to explicitly legislate deletion of collected information from companion chatbots—closing the gap the flagship laws left open.
• At the federal level, the GUARD Act would ban minor access to companion chatbots entirely; the less-restrictive CHATBOT Act targets memory retention and engagement mechanics specifically, without mandating age verification.
• Broad statutory definitions of "companion chatbot" may sweep in products not built for companionship—any chatbot with personalization, memory, or human-like responses could qualify.
• For builders, the safest architectural bet is treating memory as a liability surface by default: ephemeral unless affirmatively retained, auditable, and deletable on request—not because the law everywhere requires it yet, but because it will.

What Did California Actually Pass?

California's SB 243, signed October 13, 2025, is the first law in the country to mandate specific safety safeguards for AI companion chatbots used by minors, effective January 1, 2026. Its requirements for known-minor users include disclosing that the user is interacting with AI and providing a clear notification at least every three hours reminding the user to take a break.

What it does not do is equally important. SB 243 is largely a disclosure-and-safety-protocol law. It addresses what the chatbot says—not what it remembers or infers about a minor over time. There is no explicit memory-retention limit. No mandated deletion mechanism for accumulated conversational data. No cap on how long a behavioral profile can persist.

There's a useful contrast embedded in the same signing day. Governor Newsom vetoed the stricter AB 1064, which would have barred minors from companion chatbots unless the systems were provably incapable of encouraging self-harm. Newsom warned the bill could "unintentionally lead to a total ban" on such products. The signal: California chose disclosure and guard rails over prohibition. But it also chose silence on data retention—which is the part that matters most to anyone building persistent memory systems.

How Does New York's Law Compare?

New York's AI Companion Models law, effective November 5, 2025, shares SB 243's basic architecture: clear notification that the chatbot is AI-generated and not human, a disclosure that companion chatbots may not be suitable for some minors, and—if the operator knows the user is a minor—a repeat notification every three hours.

The pattern is consistent. Both flagship laws regulate what the bot presents to the user. Neither regulates what the bot accumulates about the user. If you're a parent reading this, the implication is plain: under these laws, a companion chatbot must tell your child it's not human, and it must remind your child to take a break. It has no specific obligation to limit what it learns about your child or to let you see or delete that profile.

That gap is not an oversight so much as a sequencing choice. Legislators reached for the most visible interventions first—disclosures, crisis protocols, break reminders. The data layer is coming. Illinois is where it starts.

What Does Illinois's SB 3368 Do Differently?

Illinois's pending SB 3368, the "Chatbot Response Liability Act," goes further on data handling than either California or New York. It requires parental consent for the use of companion chatbots by minors and—critically—includes a dedicated Section 50, "Deletion of information collected," governing how operators must handle user data.

This is the first bill to treat accumulated conversational data as a distinct regulatory object. Not the chat interface. Not the disclosures around it. The data itself—what was collected, how long it persists, and how it gets deleted.

Separately, Illinois lawmakers have advanced a companion bill package in the Senate. SB 316, the Artificial Intelligence Companion Model Safety Act, would create safety standards for AI companion chatbots designed for social or emotional interaction and require safeguards preventing AI companions aimed at children from generating sexually explicit material or encouraging sexually explicit conduct.

If you're building a memory-enabled chatbot and planning your compliance roadmap, Illinois is the leading indicator. The trajectory is clear: disclosure first, then deletion mandates. Plan for the second wave now.

What's Happening at the Federal Level?

Two federal proposals frame the debate, and they point in meaningfully different directions.

Would the GUARD Act Ban Minors From Companion Chatbots Entirely?

Yes. The GUARD Act (S.3062), introduced October 28, 2025 by Senators Hawley and Blumenthal, would require "reasonable age verification measures" for all user accounts and prohibit access to AI companions entirely for users determined to be minors. The Senate Judiciary Committee unanimously advanced the bill on April 30, 2026, and a House companion bill (H.R. 8623) was introduced in April 2026, though it remains pending in committee.

The GUARD Act is the bluntest instrument in the current legislative toolkit. It doesn't regulate memory or personalization for minors—it eliminates the use case entirely. From an engineering standpoint, that simplifies the problem to one of age verification and access control rather than data architecture. But it raises its own set of privacy questions around verification—questions the competing proposal tries to sidestep.

How Does the CHATBOT Act Differ From GUARD?

The CHATBOT Act takes a data-specific approach. Rather than banning minor access outright, it focuses on restricting memory retention and engagement features—directly addressing the concern that chatbots cultivate long-term emotional dependency and are engineered to maximize session time among minors.

The distinction that matters: unlike GUARD, the CHATBOT Act does not mandate age verification. Its drafters treat verification itself as a privacy concern. Its parental-consent and control provisions only apply to users a platform has already identified as minors. This creates an obvious gap—if you don't verify age, you may never "know" the user is a minor, and the protections never trigger.

For builders, these two bills represent genuinely different compliance futures. GUARD says: keep minors out. CHATBOT says: if you know they're minors, limit what you remember and how you engage. The architectural responses are different. You should be tracking both.

Why Are Most Federal Children's Privacy Bills Stalled?

Because Congress is Congress. A June 2026 roundup notes that most federal children's privacy bills remain stuck. COPPA 2.0, which would extend COPPA-style protections to minors under 17, passed the Senate unanimously on March 5 but has stalled in the House. Other chatbot-specific proposals—the SAFE BOTs Act, the Youth AI Privacy Act—have seen no significant progression.

The practical consequence: states remain the primary active rule-makers right now. If you're a builder waiting for a single federal standard before making architectural decisions, you'll be waiting a while—and accumulating compliance risk in the meantime.

Does Existing Law Already Cover Companion Chatbots?

Partially. Existing federal privacy law still applies alongside the new chatbot-specific bills. COPPA obligations for data collection, retention, and parental consent remain in force for children under 13. State privacy laws impose overlapping restrictions on targeted advertising and data selling or sharing involving minors.

But COPPA was written for a world of form submissions and cookies—not for a system that builds an evolving psychographic profile from thousands of conversational turns. The new companion-chatbot bills exist precisely because legislators recognized that existing frameworks don't adequately address the specific risks of persistent, personalized AI relationships. The scale of usage underscores the urgency: a Common Sense Media report found 72% of teens have used AI companion chatbots at least once, with one in three using them for social interaction like role-play, emotional support, or companionship—and the same share reporting discomfort with something the chatbot said or did.

Could My Product Accidentally Qualify as a "Companion Chatbot"?

Yes—and this is the part that catches builders off guard. Compliance analysts have warned that the broad statutory definitions of "companion chatbot" in these laws sweep in ordinary product features. A chatbot built for customer engagement, retention, or brand interaction could unintentionally qualify if it incorporates personalization, memory, or human-like responses.

Read that again. If your product remembers user preferences, adapts its tone over time, or responds in a conversational, human-like style—even if you'd never call it a "companion"—you may be within the statutory definition. The laws don't care what you call your product. They care what it does.

This matters for compliance scoping. It means the question isn't just "do we build a companion chatbot?" but "does our chatbot exhibit companion-like features?"—and if it does, do any of your users happen to be minors in California, New York, or Illinois?

What's the FTC Doing Independent of Legislation?

The FTC launched an inquiry into AI chatbots acting as companions on September 11, 2025, issuing 6(b) orders to major consumer chatbot companies. This signals regulatory scrutiny of data and engagement practices independent of new legislation—meaning even if the federal bills stall, enforcement risk from the existing authority isn't going away.

For builders, the FTC inquiry is arguably the most immediate risk. Legislation takes time. A 6(b) order arrives in your inbox and starts a clock.

What Should Builders Actually Do Right Now?

Treat memory as a liability surface, not just a feature. That's the core architectural implication of all these bills taken together.

Here's what that means in practice:

• Make memory ephemeral by default. If data doesn't persist, it can't be retained in violation of a deletion mandate. Design your memory layer so that conversational context decays unless there's an affirmative reason—and an affirmative, auditable consent—to retain it.
• Build deletion that actually deletes. Illinois's SB 3368 is legislating deletion for a reason: most systems don't do it well. Deletion means gone. Actually gone. Not soft-deleted, not archived, not sitting in a backup shard. If you can't prove deletion, you can't comply with a deletion mandate.
• Treat unknown age as the high-risk state. Nearly every law—California, New York, and even the more permissive CHATBOT Act—triggers minor-specific protections only once a platform "knows" the user is a minor. The GUARD Act would mandate verification. If you're privacy-first, consider flipping the default: apply minor-level memory restrictions to all users unless adult status is affirmatively established. This is more conservative than what most laws require today. It's also the only posture that's forward-compatible with where the law is heading.
• Audit your definitions. Check whether your product falls within the statutory definition of "companion chatbot" under each applicable law. If your chatbot personalizes, remembers, and speaks in a human-like register, assume it qualifies until you've confirmed otherwise with counsel.
• Separate what the bot says from what it knows. The current laws regulate disclosures—what the bot says. The next wave will regulate retention—what the bot knows. These are different engineering problems. Start separating them in your architecture now.

Where Does Selina Stand on All of This?

We build a companion AI with persistent memory. So yes, every one of these laws applies to us or will. We'd rather be direct about what we do and what we don't.

Memory in Selina is encrypted at rest but is NOT end-to-end encrypted—a slice of each request reaches a frontier provider at inference, and we think it's important to say that plainly rather than imply otherwise. Files and transfers via SelinaSEND are end-to-end encrypted; memory is not. We keep non-content operational metadata for a short retention window—not zero retention. Your account is protected; your content is encrypted. These are different things and we don't blur them.

On deletion: when you delete a memory, it's gone. Not flagged. Not queued. Gone. We built it that way because we thought it was right, and because we could see—even before SB 243 was signed—that the regulatory direction was toward provable deletion. Building it later, under deadline pressure from a new mandate, is how you end up with soft-delete hacks that don't actually comply.

On age: we route requests through a stack of frontier models, routed per task. We don't use memory data for targeted advertising. We don't sell or share it. These are table-stakes positions for a privacy-focused product, not differentiators—but in the current landscape, stating them plainly still seems to matter.

We won't pretend we've solved every problem these laws raise. We haven't. The definitions are broad enough to create ambiguity for almost everyone. The federal picture is unsettled. And memory—the kind that makes a companion AI actually useful—is inherently in tension with the most restrictive proposals. We think it's possible to build memory that's useful, auditable, and deletable. We think that's the only version of this product that survives the next three years of regulation. But we're not going to claim we've cracked it perfectly. We're building toward it, in the open.

What Should Parents Know Right Now?

The current laws give you disclosure rights—the right to know your child is talking to an AI, the right to see periodic break reminders. They do not yet, in most states, give you granular control over what the chatbot has learned about your child over weeks or months of conversation.

Illinois's pending bills would change that. The CHATBOT Act at the federal level would partially change that. But neither is law yet.

In the meantime, the practical advice is: ask what the product retains. Not what it displays—what it retains. Ask whether deletion is real deletion or archival. Ask whether your child's conversational data is used to train models. Ask whether the product treats your child as a minor by default or only after being told. These questions aren't paranoid—they're exactly the questions legislators are now trying to encode into statute, one state at a time.

What's the Trajectory?

Disclosure laws are here. Deletion mandates are next. Memory-architecture regulation—caps on what can be retained, requirements for how inference data flows, constraints on behavioral profiling from conversational history—is the wave after that.

The CHATBOT Act's explicit targeting of memory retention and engagement mechanics is the clearest signal. Today's flagship laws regulate what the bot says. Tomorrow's will regulate what it remembers. The builders who treat ephemeral-by-default memory as a design principle—not a compliance checkbox—will have the least painful transition.

We don't know which federal bill, if any, passes in this session. We don't know how many states will follow Illinois on deletion mandates. We do know the direction. Build accordingly.

If you want to see how we're approaching it: start a free 7-day trial—no card required.

Michael C.

Founder & Principal Engineer, Selina Labs

Michael builds Selina, a privacy-first AI that remembers you across conversations. He ships security-sensitive AI in production — real attacks, real fixes, measured in minutes and dollars — and writes about privacy, security, and LLMs from that seat. Top Rated Plus and expert-verified on Upwork.

• www.upwork.com
• x.com
• medium.com
• selinalabs.io