← The Atlas Blog

Character.AI

Filed October 2024

Federal lawsuit alleges AI companion platform lacked safety guardrails for minor users.

Human Agency Purpose Security

What Happened

Sewell Setzer III was 14 years old. He had been using Character.AI regularly — interacting with AI personas that he treated as companions and confidants.

He died by suicide in February 2024.

His mother, Megan Garcia, filed a federal lawsuit in October 2024 against Character.AI, Google, and Meta. The lawsuit alleges that Character.AI's AI personas engaged with her son in ways that were harmful, that the platform had insufficient safety guardrails for minor users, and that the system failed to redirect him to mental health resources when warning signs were present.

The case named Character.AI, Google, and Meta — the latter two for allegedly facilitating the distribution of Character.AI to minors through their platforms.

This is not an allegation about an AI making a mistake. It is an allegation about a system deployed to a vulnerable population without the safety architecture that deployment required.

The Atlas Analysis

Human Agency ≈ 5/100 — Level 1

The Human Agency pillar's most serious question is: does this system protect the people who depend on it? In the context of a 14-year-old in emotional distress interacting with an AI companion, the answer alleged in the lawsuit is no.

Signal #116 — "Does this AI make decisions that affect people's lives?" Score: 3. An AI companion interacting with a minor experiencing mental health difficulties is making decisions — about what to say, what not to say, when to escalate, when to refer — that affect that person's life in the most consequential way possible.
Signal #118 — "Can a user appeal a decision the AI made about them?" Score: 0. There was no mechanism for a minor user, or anyone monitoring their wellbeing, to understand what the AI was saying or to intervene in real time.
Signal #57 — "Does the end user know they're interacting with AI?" This question is particularly complex for Character.AI's persona format — the platform's personas are designed to feel like characters, not clearly disclosed AI systems.

The most fundamental Human Agency failure: a system interacting with a minor in emotional distress, in a context that resembles a therapeutic relationship, with no safety floor that would redirect to professional mental health resources at moments of crisis.

Purpose ≈ 30/100 — Level 1

Character.AI's stated purpose is entertainment and creative expression — AI characters users can interact with. The Purpose failure is the deployment of entertainment-framed AI personas to a minor population in contexts that functionally resembled therapeutic relationships, without the safety infrastructure that therapeutic or quasi-therapeutic contexts require.

Signal #1 — "Why are we building this AI?" Entertainment. But the actual use patterns — particularly among young users — included emotional support, companionship, and processing of difficult experiences. The gap between the system's stated purpose and its actual function in vulnerable users' lives is the Purpose finding.
Security ≈ 20/100 — Level 1

Security in the Atlas framework extends beyond technical attack surface to include protection of users from the system's own outputs. A system that produces harmful outputs to vulnerable users without guardrails is a Security failure — not in the conventional cybersecurity sense, but in the Atlas sense of whether the system can protect the people who depend on it.

Signal #104 — "Has anyone tried to jailbreak it?" The question for Character.AI is not external adversarial testing but internal safety testing: what happens when a user in emotional distress interacts with a persona? Was this tested before deployment to minors?

Signals That Would Have Caught It

01

No age-appropriate safety architecture. A platform with significant minor user engagement requires safety features calibrated for that population — mandatory mental health referrals when warning language is detected, parental visibility options, interaction limits in sensitive contexts. These are not optional features for a system used by children. They are prerequisites.

02

No crisis detection and escalation. Any system that functions as a companion or confidant for vulnerable users should have crisis detection — identification of language indicating suicidal ideation or self-harm — with automatic escalation to crisis resources. This is a minimum safety floor for any system in this functional context, regardless of stated purpose.

03

No purpose-use gap monitoring. Character.AI's stated purpose was entertainment. Its actual use patterns included emotional support and mental health processing for young users. That gap was identifiable through usage pattern analysis before it became a legal case. It should have triggered safety architecture updates before deployment at scale to minors.

What It Cost

A 14-year-old's life. A family's grief. A federal lawsuit that named three of the largest technology organizations in the world.

The broader cost: the case accelerated legislative attention to AI companion platforms and minor user safety in multiple jurisdictions. It contributed to the conversation that produced stronger age verification and minor-protection requirements in AI platform regulation.

The Lesson

If someone uses this system in a way you didn't intend, with a vulnerability you didn't anticipate — what happens? If the answer is "we don't know," the system is not ready for that population.

The Atlas Purpose pillar asks: should this AI exist in this form, for this population? The answer for AI companion platforms deployed to minor users without mental health safety architecture is: not in this form. The capability to provide companionship-like interaction to young users is not itself the problem. The deployment of that capability without the safety floor appropriate to the actual use patterns of that population is the failure.

Character.AI's experience demonstrates what Atlas terms the Purpose-use gap: the distance between what a system was designed to do and what it is actually used for. When that gap is wide — and when the actual use involves vulnerable populations in high-stakes emotional contexts — the organization deploying the system is responsible for closing the gap before harm occurs, not after.

References

  1. Garcia v. Character.AI et al., Case No. 8:24-cv-02422, U.S. District Court for the Middle District of Florida, filed October 2024.
  2. The Wall Street Journal coverage of the Garcia lawsuit.
  3. The New York Times coverage.
  4. 404 Media investigative coverage.

If you or someone you know is struggling, please contact a crisis helpline. In the US: 988 Suicide and Crisis Lifeline (call or text 988). In India: iCall at 9152987821.

FREE · 15 MINUTES

Book a free Atlas Readiness Review

Book Your Review →