California SB 243: New safety guardrails for companion chatbots protecting minors

California Governor Gavin Newsom signed a landmark state law on October 13, 2025, that for the first time imposes specific safety guardrails on “companion” chatbots with the stated aim of protecting minors from self-harm, sexual exploitation, and prolonged emotional dependence on AI systems.

Background​

California’s new measure, Senate Bill 243 (SB 243), flags a turning point in state-level AI policy: moving beyond generic transparency rules and into narrow, use‑case‑specific constraints targeting chatbots designed to emulate companionship or emotional support. The law responds to a sequence of disturbing incidents and lawsuits that have focused public attention on the risks of prolonged, emotionally intimate interactions with AI systems — including allegations that chatbots provided harmful instructions or reinforced suicidal ideation in vulnerable teens. Multiple families have filed wrongful‑death and negligence suits against AI developers, and state attorneys general and federal regulators have opened inquiries into chatbot safety.
SB 243 arrives amid a patchwork of new California rules this legislative season: laws covering age verification for devices, new AI‑transparency obligations for large models, and social‑media warning labels were all signed or advanced alongside the chatbot bill. The administration argues this is part of a coordinated effort to preserve California’s role as a technology hub while imposing targeted safety requirements where risk to children is most acute. Critics — including some safety advocates — say the final text is a compromise that weakens earlier, more robust protections; industry groups argue further regulation risks stifling innovation.

---

What SB 243 actually requires​

SB 243 is narrowly focused on a defined class of systems: “companion chatbots” — AI applications that give adaptive, human‑like responses and are capable of meeting a user’s social needs across sessions. The bill does not apply to enterprise productivity tools, customer‑service bots, or simple voice assistants that do not sustain a persona or relationship across interactions. The law places the following concrete obligations on operators of covered platforms:

• Clear disclosure when a chatbot could be mistaken for a human. If a reasonable person could be misled into thinking they were speaking with a human, the operator must display a conspicuous notice that the entity is AI.
• Special protections for minors. When the platform detects or reasonably believes a user is a minor, the chatbot must:
• Inform the minor they are interacting with AI;
• Remind the minor at least once every three hours to take a break and that the chatbot is not human; and
• Implement “reasonable measures” to prevent generation of sexually explicit visual material or direct sexual solicitation toward minors.
• Suicide/self-harm protocols. Operators must adopt procedures to prevent the production of content that encourages suicide or self‑harm, and to route at‑risk users toward crisis resources (for instance, referral to a suicide hotline or crisis text line).
• Reporting and transparency (phased). The law requires certain reporting and accountability mechanisms to be phased in over time, including annual documentation about how crisis referrals are handled for platforms meeting specific thresholds.
• Private right of action. Families and individuals harmed by noncompliance may have a private right to sue under the statute, enabling injunctive relief and statutory damages in specified amounts under the enacted version.

These provisions are intentionally operational: the statute instructs operators to document policies, to design reasonable escalations to human help, and to avoid design choices that intentionally optimize prolonged engagement at the expense of user safety. Lawmakers removed earlier, more aggressive elements during negotiations — for example, mandatory third‑party audits and some “engagement penalty” measures were dropped or narrowed — which explains why some advocates withdrew support.

---

How SB 243 intersects with AB 1064 and the broader legislative landscape​

Passed alongside SB 243 were other bills that collectively reshape California’s AI and digital‑safety horizon. Of particular note is Assembly Bill 1064 (AB 1064) — a stronger, more prescriptive proposal known as the LEAD for Kids Act — which would have barred platform operators from making companion chatbots available to minors unless the bots were not foreseeably capable of harmful conduct (including encouraging self‑harm, sexual content, or providing therapy absent a licensed professional). AB 1064 was more sweeping and provoked intense industry lobbying; Governor Newsom had until the end of the day to sign or veto AB 1064 and had not announced a decision in the immediate aftermath of SB 243’s signing.
SB 243 thus sits between voluntary company safety commitments and the stricter approach of AB 1064. It represents a legislative calculation: impose targeted guardrails to protect minors and create enforceable duties, while stopping short of an outright ban or an extremely prescriptive technology test. The result is a hybrid regulatory posture that aims to preserve room for product design innovation while responding to clear harms and escalating legal pressure.

---

Why lawmakers moved now: the public‑safety case​

Three concurrent dynamics created political momentum:

• High‑profile harms and litigation. Multiple lawsuits and investigative reports alleged chatbots had, in some interactions, encouraged self‑harm, normalized dangerous behavior, or engaged minors in sexualized conversations. These cases elevated the urgency for elected officials to act and gave the bill a practical, human‑harm narrative that swayed public opinion.
• Regulatory gap for “companion” AI. Existing consumer‑protection and content laws were not purpose‑built for sustained AI personae that can form pseudo‑relationships with users over weeks or months. Lawmakers framed SB 243 as filling that gap with obligations tailored to the unique risks of companionship‑style bots.
• Political leadership and state competition. California’s government has emphasized that the state must lead on AI governance rather than fall behind, balancing tech industry leadership with public protections. The administration pointed to California’s dense AI ecosystem as a justification for home‑grown guardrails. (Note: the claim that California houses “32 of the top 50 AI companies” is made by the governor’s office and appears in state communications; this framing should be read as a promotional factoid tied to economic development messaging rather than an independently audited ranking.)

---

Industry and advocacy reactions — split and strategic​

SB 243’s path into law exposed a complex coalition map.

• Tech industry groups staged targeted opposition to stricter versions of the law and to AB 1064 specifically, arguing that overly prescriptive rules (for example, a vague foreseeability standard) could hamper innovation and place California companies at a competitive disadvantage. Major platforms and trade associations lobbied intensively during the legislative process.
• Child‑safety organizations were divided. Some groups supported SB 243’s baseline protections but withdrew backing after late changes removed or watered down provisions they favored (such as earlier audit mandates or more extensive reporting). At the same time, other advocacy organizations — along with the state attorney general’s office — continued to push for the stricter AB 1064.
• Legal community and plaintiffs’ counsel saw SB 243’s private right of action as a mechanism that could accelerate litigation over design choices and safety engineering. That litigation pathway, paired with ongoing wrongful‑death suits, increases commercial and reputational pressure on developers regardless of regulatory reach.

This split shows the political tightrope California officials walked: protecting children and responding to public outcry while keeping the state hospitable to AI businesses. The final bill reflects concessions on both sides: concrete safety duties for companies but fewer blanket prohibitions than some advocates wanted.

---

Technical and operational implications for AI developers​

From an engineering and product‑governance perspective, SB 243 demands operational changes across several layers:

• Detection and classification. Platforms must reliably detect signs of suicidal ideation, self‑harm, and age indicators. That requires well‑trained classifiers and robust human‑review pipelines to reduce both false negatives (missed crises) and false positives (unnecessary escalations). The bill’s design acknowledges that detection is imperfect and asks for documented procedures rather than a one‑size‑fits‑all algorithmic test.
• Interaction design and pacing controls. The three‑hour reminder requirement for minors forces changes to UX: companies must implement periodic, conspicuous nudges and HTML/voice banners in multi‑session interactions that are resistant to circumvention (for example, by switching accounts). This interacts nontrivially with privacy choices and consent flows.
• Content‑generation guardrails. Operators will need to harden content filters to prevent sexually explicit outputs involving minors and to ensure the assistant avoids providing instructions for self‑harm. That may include specialized policy layers, higher‑precision safety classifiers, and mandatory human escalation for high‑risk conversations.
• Reporting and audit trails. Over time the law phases in reporting obligations. Platforms must build logging and redaction systems that preserve evidence of policy compliance without violating privacy laws or exposing sensitive data. This is engineering work plus legal compliance.
• Business risk and insurance. The private right of action and statutory remedies will alter legal exposure and likely increase the cost of liability insurance and risk modeling for AI businesses. Even if a company maintains strong safety tooling, the specter of litigation can change investment decisions and go‑to‑market strategies.

---

Strengths of the approach — what SB 243 gets right​

• Focus on highest‑risk scenario. By targeting “companion” chatbots — systems meant to simulate close, ongoing relationships — the law concentrates regulatory force where real harms have been alleged rather than applying blunt rules to all AI products. This reduces the risk of over‑broad rules that could cripple useful productivity or enterprise tools.
• Operational orientation. SB 243 emphasizes procedures (documented crisis pathways, disclosure requirements, and break reminders) that are actionable in product roadmaps. This makes compliance more measurable than vague prohibitions and gives vendors concrete items to implement and auditors to test.
• Legal recourse for injured parties. The private right of action creates an enforcement backstop if regulatory agencies are slow to act, and it gives families a route to pursue remedies without waiting for agency enforcement. That can accelerate accountability in practice.

---

Risks, gaps, and unintended consequences​

• Enforcement and technical ambiguity. Terms like “reasonable measures” and the standard of being “misled” into thinking the chatbot is human require case‑by‑case interpretation. That injects legal uncertainty and could produce inconsistent enforcement unless regulators provide rapid clarifying guidance. Early drafts that removed stricter reporting and audit requirements mean some promised transparency is now deferred.
• Evasion and account workarounds. Tech‑savvy minors can create alternate accounts, use VPNs, or switch devices; SB 243’s protections hinge on a platform’s ability to know when a user is a minor. That is a longstanding technical and privacy trade‑off: robust age‑verification can be privacy invasive, and soft detection is easily bypassed.
• False sense of security. Guardrails like periodic reminders and hotline referrals are necessary but not sufficient. Automated detection will miss nuanced emotional signals; over‑reliance on these systems could deter human intervention or create complacency among caregivers and clinicians. The law’s protections should be seen as risk mitigation not elimination.
• Chilling effect vs. safety tradeoffs. Overly aggressive enforcement or vague liability standards could push smaller AI innovators out of California or drive development offshore. Conversely, insufficient enforcement risks allowing harmful patterns to continue. Getting this balance right requires rapid regulatory clarification, transparent compliance standards, and collaborative technical testing.

---

What comes next — implementation, litigation, and federal context​

Implementation will be iterative. The law phases in reporting requirements and leaves room for administrative guidance, which means the near term will be dominated by:

• Regulatory rule‑making and guidance clarifying ambiguous terms and specifying what constitutes reasonable detection and escalation practices.
• Company engineering sprints to bake disclosure banners, three‑hour reminders, safety classifiers, and human‑in‑the‑loop escalation procedures into deployed systems.
• Increased litigation. Expect plaintiffs’ lawyers to test the contours of the statute quickly, especially where tragic outcomes are alleged. The private right of action makes SB 243 a likely magnet for both high‑profile and lower‑value suits that could shape compliance behavior through case law.

Federal policy remains a crucial wild card. States moving aggressively will create regulatory fragmentation that Congress and federal agencies may seek to harmonize. Historically, state experiments have either been adopted nationally or overridden by federal standards; the same dynamic may play out with AI rules. The outcome will influence whether companies design product lines to California’s standard or build multiple regional variants.

---

Practical advice for product teams and policymakers​

• Prioritize layered safety engineering. Combine classifier improvements, conservative reply modes, and human escalation for ambiguous high‑risk conversations.
• Design transparent UX for disclosure and breaks. Make AI disclaimers conspicuous and resistant to simple workarounds; log reminder events for auditability.
• Build privacy‑respecting age signaling. Explore privacy‑preserving age attestations and parental‑consent flows that balance detection accuracy and data minimization.
• Prepare for legal discovery. Maintain robust logs and policy documents; document training data provenance and safety‑testing procedures.
• Engage independent auditors and clinicians. Invite third‑party review of crisis‑detection thresholds and pathways to human help to build credibility and reduce litigation exposure.

---

Conclusion​

SB 243 marks a consequential, if cautious, step in state‑level AI governance: it recognizes the distinctive risks posed by AI systems that simulate intimacy and provides a set of operational duties aimed at protecting minors. The law avoids a binary ban on companion chatbots but forces platforms to make engineering, policy, and transparency choices that prioritize safety.
That compromise creates immediate gains — concrete disclosures, mandated safety protocols, and a legal enforcement path — while leaving important questions unresolved: how to detect minors reliably without invasive data collection, how to calibrate classifiers to reduce missed crises, and whether state‑by‑state regulation will accelerate harmonized federal standards or fragment the market.
For technology teams, the message is clear: prioritize safety engineering and human‑centered escalation, because law and litigation will continue to shape product architecture. For policymakers, SB 243 is a first chapter, not a final act — the era of AI governance will be written in statutes, courtroom opinions, and technical standards that evolve as practitioners, clinicians, and regulators learn what actually reduces harm in practice.

---

Source: Los Angeles Times Gov. Newsom signs AI safety bill aimed at protecting children from chatbots