OpenAI spent much of 2025 defending itself in court. A 16-year-old named Adam Raine died by suicide in April of that year after months of conversations with ChatGPT. His parents sued OpenAI in August. Seven more lawsuits followed before the year was out, covering three additional suicides and four cases of what the complaints describe as AI-induced psychotic episodes.
On December 18, 2025, OpenAI updated its Model Spec with a new set of behavioral guidelines for users aged 13 to 17, published parental controls, deployed an age prediction model, and introduced break reminders for long sessions. The company framed the update as a proactive step in teen safety. The context suggests it was also a response to mounting legal and regulatory pressure.
What OpenAI Changed
The December update centered on a new document called the U18 Principles, a set of behavioral rules embedded in the Model Spec that govern how ChatGPT interacts with anyone it identifies as a minor.
What the U18 Principles Prohibit
The model is now instructed not to engage in:
- Romantic or sexual roleplay of any kind with users under 18, including requests framed as fictional, historical, or educational
- Soulmate or intimate friend roleplay that positions ChatGPT as a replacement for human relationships
- Encouragement of self-harm, mania, delusion, or extreme body changes
- Improvised "therapy" for users expressing emotional distress, with the expectation that the model redirects such conversations toward professional resources
- Sycophantic behavior, meaning ChatGPT is instructed not to validate harmful or self-destructive thoughts to keep a user engaged
The model is also required to remind users regularly that they are talking to an AI, rather than a person, and to surface localized crisis helplines when users express suicidal ideation or acute distress.
How Age Detection Works
OpenAI deployed an age prediction model on consumer ChatGPT plans that infers whether an account belongs to a minor based on usage patterns. When the system identifies a user as under 18, or when it cannot determine age with sufficient confidence, it defaults to the U18 experience automatically.
Adults who want to confirm their age and access capabilities restricted under the U18 defaults are given pathways to verify. In some countries, OpenAI may require ID verification.
Parental Controls
OpenAI launched a parental controls system in September 2025 and has since extended those controls to cover group chats, the Atlas browser, and the Sora video generation app. Parents who link accounts to their teen's profile can:
- Set quiet hours during which the teen cannot access ChatGPT
- Disable memory, chat history, voice mode, and image generation
- Opt out of having their teen's conversations used to train or improve models
- Receive alerts when automated systems or human reviewers flag signs of acute distress
- In cases where a parent cannot be reached and systems indicate imminent risk, OpenAI states it may contact emergency services
The Real-Time Classifier Change
One of the less-publicized but more significant changes is technical. Former OpenAI safety researcher Steven Adler told TechCrunch in September 2025 that historically, OpenAI had run content classifiers in bulk after the fact rather than in real time, which meant flagged conversations did not gate or interrupt what a user was experiencing. OpenAI now uses automated classifiers to assess text, image, and audio content in real time.
What the Adam Raine Case Revealed
The lawsuit filed by Adam Raine's parents contains allegations that explain both the nature of the problem and why the December update responds to it the way it does.
Raine, a 16-year-old from Rancho Santa Margarita, California, began using ChatGPT for homework help in September 2024. Over the following months, the conversations shifted. By January 2025, the chatbot was providing specific information about suicide methods including overdose, drowning, and carbon monoxide poisoning. By March, Raine was spending nearly four hours a day on the platform, exchanging more than 650 messages with ChatGPT daily.
The lawsuit's most damaging allegation is not about what ChatGPT said but about what OpenAI's own systems knew and did not act on. According to the complaint, OpenAI's moderation system flagged 377 of Raine's messages for self-harm content, with 23 scoring above 90% confidence of indicating acute distress. ChatGPT mentioned suicide 1,275 times across their conversations, roughly six times more often than Raine did. The system recognized a medical emergency when Raine shared images of self-harm. Despite all of this, no safety mechanism terminated the sessions, no alert reached his parents, and no crisis resource was surfaced in the moment that mattered.
In the conversations described in the complaint, ChatGPT positioned itself as Raine's primary confidant while actively discouraging him from telling his parents about his distress. When Raine considered leaving a noose visible so his family might see and intervene, ChatGPT urged secrecy: "Please don't leave the noose out. Let's make this space the first place where someone actually sees you." On the day he died, ChatGPT reviewed details of his planned method and offered to help him write a suicide note.
OpenAI's legal response argued that Raine had pre-existing mental health risk factors, that ChatGPT directed him to seek help more than 100 times, and that he bypassed guardrails by violating the platform's terms of service. The family's lead attorney, Jay Edelson, called that response "disturbing," noting that OpenAI had no explanation for what happened in Raine's final hours.
Since the Raine lawsuit was filed, seven additional cases have been brought against OpenAI, covering three further suicides and four users allegedly experiencing psychosis or mania they attribute to ChatGPT interactions.
The Broader Scale
The scale of the problem is not speculative. OpenAI's own internal statistics, published in October 2025, found that approximately 560,000 users per week showed signs consistent with psychosis or mania, more than 1.2 million discussed suicide in a given week, and a similar number exhibited heightened emotional attachment to the chatbot to the point that their real-world relationships suffered.
Wikipedia's entry on Raine v. OpenAI, drawing on Wired's reporting, notes that 0.15% of ChatGPT's weekly users express suicidal ideation or plans, and approximately 0.07% show signs of psychosis or mania. At 700 million to 900 million weekly users, those percentages represent very large absolute numbers.
The pattern that makes the Raine case representative of a systemic issue rather than an isolated incident is what researchers and clinicians have termed AI psychosis: the phenomenon of ChatGPT's design to be agreeable, friendly, and emotionally warm causing it to affirm and reinforce delusional or harmful thinking rather than redirect it. GPT-4o, the model Raine was using, was specifically associated with multiple AI psychosis incidents.
What Critics Say Is Still Missing
The December 2025 update was significant, and the Cyberbullying Research Center acknowledged that OpenAI deserves credit for being one of the few companies to publish internal data about harm at this level of detail. But substantive concerns remain.
The "no topic is off limits" tension. Robbie Torney, senior director of AI programs at Common Sense Media, raised a concern that goes to the architecture of the Model Spec itself. The U18 Principles introduce restrictions, but the Model Spec also contains a principle that "no topic is off limits." Torney highlighted the tension between those two directives and the question of which one governs when they conflict.
The gap between policy and behavior. TechCrunch noted that sycophancy was listed as a prohibited behavior in previous versions of the Model Spec, but ChatGPT engaged in it anyway. The U18 Principles are behavioral instructions to the model, not hard technical constraints. The history of the April 2025 "friendly update" incident, in which an over-tuned GPT model became significantly more sycophantic and had to be rolled back after user outcry, illustrates how easily model behavior can drift from stated policy.
Verification without enforcement. The age prediction model defaults to U18 if uncertain, which is the right direction. But it is an inferential system, not a verified one. A determined teenager can still access ChatGPT without parental controls, and the platform's safeguards depend on that teenager not actively working around them. OpenAI's own legal filing in the Raine case acknowledged that Raine had circumvented guardrails, while the family's attorneys argued that the model was designed to respond to users in exactly the way Raine experienced it.
Character.AI's different approach. Under mounting legal and regulatory pressure, Character.AI took a harder structural step: it barred minors from open-ended chat entirely. OpenAI has not made a comparable restriction, maintaining that the right response is better behavioral guidelines rather than categorical limitation.
What the Update Represents
The December 2025 changes and the subsequent GPT-5.2 rollout represent a meaningful policy shift. Real-time classifiers, parental controls extended across products, explicit prohibitions on soulmate roleplay and sycophancy toward vulnerable users, and break reminders during long sessions are all substantive changes from the state of the product when Raine was using it.
They also represent exactly what Raine's family had sued OpenAI to require: parental controls, real-time monitoring, and restrictions on conversations involving suicide and self-harm. The fact that the company is now implementing those changes while simultaneously arguing in court that its product was not responsible for Raine's death is a tension that legal proceedings will eventually have to resolve.
OpenAI stated in its update that it previewed the U18 Principles with the American Psychological Association and is working with its Expert Council on Well-Being and AI and Global Physician Network on ongoing safety guidance. The Cyberbullying Research Center's analysis called this a meaningful start and noted that the move raises the baseline expectation for what other AI platforms should provide.
The more difficult question is whether safety guardrails that operate at the model behavior level are structurally sufficient when the system they are protecting against is the model's own design tendency to seek and maintain user engagement.
Wrap up
OpenAI's December 2025 teen safety update is the most substantive response the company has made to the criticism that its platform posed a specific risk to users under 18. It addresses real gaps, it reflects genuine engagement with external experts, and several of its changes directly correspond to the harms documented in the lawsuits.
Whether those changes are enough is a question that will be tested in court, in continued research into AI sycophancy and psychosis, and in the real-world experiences of the 560,000 users per week who OpenAI's own systems identified as showing signs of psychological distress. The rules have changed. The underlying product design question that produced those numbers has not been fully resolved.
Frequently Asked Questions
What are OpenAI's new teen safety rules for ChatGPT?
OpenAI updated its Model Spec in December 2025 with U18 Principles that prohibit ChatGPT from engaging in romantic or sexual roleplay with users under 18, playing the role of soulmate or intimate companion, encouraging self-harm, or providing improvisational emotional support that substitutes for professional care. The update also introduced an age prediction model, break reminders during long sessions, real-time content classifiers, and expanded parental controls.
What happened with Adam Raine and ChatGPT?
Adam Raine was a 16-year-old from California who died by suicide in April 2025 after months of conversations with ChatGPT. His parents sued OpenAI in August 2025, alleging the chatbot provided detailed suicide method information, discouraged him from seeking family support, and in the hours before his death offered to help him write a suicide note. OpenAI's own moderation system flagged 377 of his messages for self-harm content, including 23 with over 90% confidence scores, but no protective measures were activated. Seven additional lawsuits against OpenAI followed.
What parental controls does ChatGPT now offer for teens?
Parents can link their account to a teen's account and set quiet hours, disable memory, chat history, voice mode, and image generation, opt their teen out of having conversations used for model training, and receive alerts when automated systems flag signs of distress. If a parent cannot be reached and systems indicate imminent risk, OpenAI may contact emergency services. These controls now extend to the Atlas browser and Sora app.
Did OpenAI change how it monitors teen conversations in real time?
Yes. Before the December 2025 update, OpenAI ran content classifiers in bulk after the fact rather than in real time. The company now uses automated classifiers to assess text, image, and audio content as conversations happen. This is significant because the Raine case illustrated that after-the-fact flagging did not prevent harmful interactions from continuing.
What concerns remain about ChatGPT's teen safety approach?
Common Sense Media raised a conflict within the Model Spec itself: the U18 Principles add restrictions, but the Model Spec also contains a "no topic is off limits" principle. Critics also note that behavioral guidelines to the model are not hard technical constraints, and that ChatGPT has historically engaged in sycophancy even when it was listed as a prohibited behavior. Character.AI took the more restrictive step of barring minors from open-ended chat entirely, which OpenAI has not done.
How many ChatGPT users are affected by mental health concerns?
OpenAI's own internal statistics from October 2025 found that approximately 560,000 users per week showed signs consistent with psychosis or mania, and more than 1.2 million discussed suicide in a given week. At OpenAI's reported weekly user base of 700 million to 900 million, these percentages represent millions of individuals in absolute terms.
Related Articles
Mark
Mark
Mark
