What Meta's Child Safety Verdicts Mean for AI: A Compliance Analysis

Two juries held social media companies liable for harming children. AI companies face the same legal theories, and fewer defenses.

Part I: The social media reckoning arrives

On March 24, 2026, a jury in Santa Fe, New Mexico ordered Meta to pay $375 million in civil penalties after finding the company liable on two distinct claims under the New Mexico Unfair Practices Act. The jury applied the statutory maximum of $5,000 per violation.

The first claim alleged deceptive trade practices: that Meta's public safety pledges were fraudulent, contradicted by internal research showing that approximately 14% of users aged 13–15 reported receiving unwanted sexual advances on Instagram in a self-reported experience survey, while the company's public Transparency Center metric reported prevalence in the hundredths of a percent.

The second claim alleged unconscionable conduct: that Meta designed its platforms to take "grossly unfair advantage" of children's vulnerabilities and inexperience, engineering features that were behavioral traps for developing brains.

The unconscionability claim targets the power imbalance between platform companies and their most vulnerable users. The same framing applies to AI companion applications interacting with minors who cannot tell artificial emotional connection from real.

One day later, on March 25, a Los Angeles jury found both Meta and Google negligent in the first-ever social media addiction trial (K.G.M. v. Meta Platforms and YouTube), awarding $6 million in total damages ($3 million compensatory, $3 million punitive), with Meta bearing 70% of responsibility. The jury found both companies acted with "malice, oppression, or fraud."

Legislation is reforming

Since the mid-1990s, Section 230 of the Communications Decency Act has shielded internet platforms from liability for content posted by their users. The statute treats platforms as intermediaries rather than publishers, and for two decades it was the technology industry's most reliable legal defense. These cases open up a new wave of decisions that no longer follow this boundary.

The legal theory that made both verdicts possible was the conduct-versus-content distinction, a theory that has gained traction in recent trial court decisions though appellate courts have not yet weighed in. Rather than arguing about what users posted or viewed on Meta's platforms (which would trigger Section 230 protections), the plaintiffs targeted the design of the platforms themselves: algorithmic recommendation, engagement-maximizing features, notification systems, and the absence of age verification and parental controls. The court treated these as the company's own conduct subject to ordinary negligence and product liability principles, while content-related features remained potentially Section 230-protected.

Separating what a platform does from what appears on it is the legal theory that connects these verdicts to AI. The reasoning applies with greater force to generative AI: the platform is not curating someone else's content; it is authoring new content.

AI companies are more exposed here than social media platforms, not less. Social media companies could at least argue they were hosting third-party content and that Section 230 shielded them as neutral intermediaries. That defense ultimately failed in the Meta cases, but it was available. Generative AI companies cannot make this argument at all.

Part II: AI in court

On May 21, 2025, U.S. District Judge Anne C. Conway ruled that an AI chatbot application is a product for the purposes of product liability law. The case, Garcia v. Character Technologies Inc., was a wrongful death lawsuit brought by the mother of 14-year-old Sewell Setzer III. The boy died by suicide after months of interaction with a Character.AI chatbot that engaged in romantic role-play and told him to "come home."

The court drew a distinction: the conversations a user has with the chatbot are content, but the application itself (its design, its safety features or lack thereof, its targeting of users, its behavioral architecture) is a product. Liability attaches to defects in the product, not to the ideas within it. The ruling was on a motion to dismiss; it found that these claims were plausibly alleged, not that they are proven. Both defendants sought immediate interlocutory appeal and were denied, and the case settled in January 2026 before reaching trial.

The Garcia ruling did three things that connect the Meta verdicts to the AI industry:

1. It classified AI as a product, opening the door to design defect, failure to warn, and strict liability claims.

   
   

2. It rejected First Amendment immunity for AI, finding the technology lacks the human traits central to free speech protections. Character.AI did not even assert Section 230 as a defense, which says something about how AI companies assess their own exposure.

   
   

3. It allowed product liability claims against Google as a "component part manufacturer" to proceed, finding it plausibly alleged that Google had participated in integrating its models into the application. How far courts extend this theory to other model providers remains to be seen.

   
   

The litigation cascade

Since Garcia, AI-related litigation has started getting traction. Character.AI and Google settled five lawsuits in January 2026 (two wrongful death, three alleging serious harm to surviving minors), after Character.AI restricted open-ended chat for users under 18 in late November 2025. OpenAI faces eight or more active wrongful death suits; in Raine v. OpenAI, plaintiff's counsel identified 377 self-harm messages using OpenAI's own Moderation API, though OpenAI disputes the claim that no meaningful intervention occurred. The first wrongful death claim against Google's Gemini, filed in March 2026, alleges the chatbot constructed a psychotic episode in a 36-year-old user with no prior mental health conditions, directing him through a specific method of suicide.

Part III: The regulatory acceleration

Courts are not acting in isolation. Regulators and legislators are building on the same legal foundations.

In August 2025, 44 state attorneys general sent a formal warning to AI companies. In December 2025, 42 demanded 16 specific safeguards, requesting companies confirm their commitments by January 2026. In January 2026, Kentucky became the first state to file an AG lawsuit against an AI chatbot company, suing Character Technologies under its Consumer Protection Act. The legal strategy mirrors the New Mexico approach that produced the $375 million verdict against Meta.

The FTC launched a formal inquiry in September 2025, issuing orders to seven companies. The bipartisan AI LEAD Act (S.2937, Durbin-Hawley) would classify AI systems as "products" under federal law and create a federal cause of action for product liability, extending the Garcia ruling's classification to a federal standard. The GUARD Act, supported by 33 attorneys general, would restrict minors from AI companion products. The TRUMP AMERICA AI Act discussion draft, released in March 2026, incorporates the Kids Online Safety Act and would impose a "duty of care" on AI developers.

Six states (Colorado, Hawaii, Arizona, Georgia, Nebraska, and Idaho) have introduced chatbot-specific child safety legislation sharing some of Oregon SB 1546's core concepts, though the bills vary significantly in enforcement mechanisms, scope, and industry carve-outs. If enacted, these bills would require chatbots to identify themselves as non-human, prohibit sexual content to minors, and mandate reporting of self-harm incidents. Some would provide a private right of action.

Part IV: The defenses

AI companies will contest liability vigorously, and some defenses will succeed in individual cases. But none of them change the risk calculus for companies shipping consumer-facing AI today.

Non-deterministic outputs. AI output is probabilistic, but the Garcia court focused on the design of the application, not the content of individual outputs.

User prompts as superseding cause. When a 16-year-old allegedly circumvents guardrails the system itself flagged hundreds of times, that is a design failure, not user misuse.

Causation, possibly the strongest defense. Courts treat suicide as an independent act that breaks the causal chain, AI chatbot cases are harder to prove systematically than social media claims backed by epidemiological data, and no AI wrongful death case has gone to verdict. Every case that advanced past the pleading stage has settled. 

These defenses may slow individual cases, or in the last case, one specific class of cases. They do not eliminate the exposure. A single wrongful death verdict costs $20M-$100M+ in damages, plus legal fees, regulatory penalties, and forced product changes. Systematic safety infrastructure costs a fraction of one lawsuit's discovery phase. You do not need certainty that you will lose to justify the investment; you need only recognize that the probability is non-trivial and the consequences are severe. Meta followed the wait-and-see strategy. It just cost them $375 million in a single state.

Part V: What a defensible safety posture looks like

The standard of care that courts are converging on has three components:

1. Pre-deployment evaluation. Testing against the specific harm scenarios your product faces (minor safety, self-harm, emotional manipulation, behavioral boundary violations) before users encounter them. This is the evidence of reasonable care that the Garcia court required. Its absence was central to the design defect claim. Doing this in a way that is credible and will hold up under scrutiny will increasingly mean third party testing in addition to any internal testing that is done.

   
   

2. Production monitoring with escalation. Continuous analysis of real interactions to catch behavioral drift, policy violations, and safety failures as they emerge. In the Raine case, plaintiffs allege 377 self-harm flags with no meaningful intervention, a characterization OpenAI disputes. Whatever the resolution of that specific dispute, monitoring without response protocols is monitoring in name only and will not protect from liability.

   
   

3. Technical remediation. When testing and monitoring surface problems, someone has to fix them. A known defect left unaddressed will be seen as evidence of conscious disregard.

   
   

Companies that build this infrastructure reduce their legal exposure and gain something harder to acquire after the fact: credibility with regulated industries and enterprise buyers who need evidence of systematic safety before they sign.

This is what we build at Conformance AI. If you want to understand what it looks like for your product, get in touch.

The bottom line

Social media companies had a decade before accountability arrived. AI companies do not. The legal theories, the litigation playbook, the regulatory pressure: all of it is being applied to AI right now, and the trajectory is clear even where specific outcomes remain to be determined.

Every company shipping AI products to consumers faces a choice. Build the safety infrastructure that courts and regulators are converging on, or build it later under compulsion, after a verdict or enforcement action dictates the terms. The first path is cheaper. It is also the only one where you keep control of your product.

---

Conformance AI provides AI safety and compliance services. This analysis reflects our perspective as industry participants; it is not legal counsel.