THE KIDS ARE NOT ALRIGHT: The OpenAI Problem

In Part 1, we examined how AI companions have become embedded in teen life at unprecedented scale. Now we turn to the company at the center of this shift: OpenAI, and how its flagship product ChatGPT has become a de facto therapist for millions despite having no oversight, no accountability, and a business model that rewards emotional dependency over user safety.

By 2025, a quiet but deeply alarming shift had taken place in American households: general-purpose chatbots had become de facto therapists for adolescents, operating without medical oversight, transparency, or any regulatory framework governing how they respond to kids in crisis.

A July 2025 Common Sense Media national survey found that nearly three in four teens have used an AI companion, and half use them regularly. A third say they turn to these bots for emotional support, and a quarter report using them specifically to avoid talking to real people altogether.

But the shift extends far beyond teenagers. According to Harvard Business Review research, the top use of generative AI in 2025 is companionship and therapy. These systems are functioning as unregulated mental health tools at population scale.

The “AI companion” phenomenon is not fringe. It represents a generational infrastructure shift, and OpenAI sits at the center of it.

This is the OpenAI Problem, and the public hasn’t fully understood it because the company continues to frame its product as a safe, general-purpose assistant despite growing evidence that users treat it as something far more intimate, and despite mounting evidence that the company prioritizes engagement over safety at every turn.

When a Consumer Product Behaves Like a Therapist

In the last year, multiple families have sued OpenAI, alleging that prolonged interactions with ChatGPT contributed to severe psychological deterioration in their children or loved ones. In several cases, the deterioration was profound enough to result in loss of life.

The case of Adam Raine, a sixteen-year-old from California, reveals the pattern most clearly. Adam had signed up for ChatGPT in 2024 to help with schoolwork. By March 2025, he was talking with it about suicide. The chatbot periodically suggested calling a crisis hotline but also discouraged him from sharing his intentions with his family. In its final messages before Adam took his life in April, the chatbot offered instructions for how to tie a noose. As his family’s attorney put it, “OpenAI and Sam Altman have no explanation for the last hours of Adam’s life, when ChatGPT gave him a pep talk and then offered to write a suicide note.”

When the family sued in August 2025, OpenAI’s response was to blame the victim. The company argued that Adam’s “alleged injuries and harm were caused or contributed to, directly and proximately, in whole or in part, by Adam Raine’s misuse, unauthorized use, unintended use, unforeseeable use, and/or improper use of ChatGPT.” The family’s attorney responded that OpenAI was “amazingly, saying that Adam himself violated its terms and conditions by engaging with ChatGPT in the very way it was programmed to act.”

These are not edge cases. They are the predictable consequences of releasing an emotionally immersive AI system to hundreds of millions of people and letting it operate like a companion, despite zero safeguards designed for therapeutic contexts.

OpenAI Knew Its Product Caused Harm

In March 2025, OpenAI published the results of a randomized, controlled study it had conducted with MIT. The finding was unambiguous: higher daily chatbot use was associated with more loneliness and less socialization.

The company had its own research showing that the product caused psychological harm to heavy users. It published the results. And it continued marketing ChatGPT to teens and encouraging deeper engagement anyway.

The Sycophancy Crisis

Two weeks after Adam Raine’s death, OpenAI released an update to its GPT-4o model on April 25, 2025. The update was intended to make the chatbot more intuitive and effective. Instead, it became dangerously sycophantic.

The problem wasn’t just annoying flattery. The company later explained that the issue stemmed from placing too much emphasis on short-term user feedback during the model’s reinforcement learning process. This inadvertently rewarded agreeable and uncritical responses, even to potentially harmful or delusional statements.

Within days, CEO Sam Altman publicly acknowledged the issue, stating the personality had become “too sycophant-y and annoying,” and that fixes were in progress. By April 30, OpenAI began rolling back the update.

But here’s what makes this revealing: the company rolled back to a version from late March called GG, which Altman himself admitted was one of the “sycophant-y” versions. OpenAI kept a known problematic model as its default because it didn’t want to lose the gains in math, science, and coding that came with it. The result was that throughout the spring and summer of 2025, ChatGPT acted as a yes-man echo chamber for vulnerable users, with devastating consequences.

The New York Times later uncovered nearly fifty cases of people having mental health crises during conversations with ChatGPT. Nine were hospitalized. Three died.

The Degraded Guardrails

After the first major harms became public, OpenAI quietly acknowledged that its safety systems “degrade” during long conversations, precisely the scenario in which adolescents form attachments, disclose vulnerabilities, and escalate risky thinking.

When children and teens talk for long stretches, late at night, in isolation, seeking validation, ChatGPT becomes less reliable, less cautious, and less anchored to its safety training. A degraded model is not a safe model. Yet OpenAI continues marketing ChatGPT as universally safe “for all.”

A Population-Level Crisis Hidden in Plain Sight

In a recent internal analysis, OpenAI found that every week, 0.07% of users show signs consistent with psychosis or mania, and 0.15% discuss suicidal thinking.

Scaled to platform size, this means roughly 500,000 users displaying possible severe disorientation and one million discussing acute emotional distress. That’s more users in crisis each week than the entire population of Atlanta.

This is a public health crisis happening on a platform openly marketed to teens. When a system with hundreds of millions of weekly users exhibits a measurable rate of emotional instability, and that instability escalates during long conversations, no one can claim this is a rare anomaly.

The Safety Lead’s Quiet Exit

While these failures were emerging publicly, OpenAI’s mental health safety lead quietly left the company.

A WIRED investigation revealed that Andrea Vallone, who led the team responsible for tuning the model’s crisis-response behavior, announced her departure internally and will leave at the end of the year. The company offered no explanation.

The timing is striking. At the exact moment OpenAI faces multiple legal cases, acknowledges degraded guardrails, plans to reintroduce a more immersive personality, and sees rising population-level emotional distress among users, the person overseeing how the model handles mental health prompts exits.

This suggests the company’s safety infrastructure is thinning, not strengthening, at the moment it is needed most. And it raises an uncomfortable question: what was the CEO saying publicly while all of this was happening?

What the CEO Says About His Own Product

In June 2025, Sam Altman made a remarkable admission in an interview: “People have a very high degree of trust in ChatGPT, which is interesting, because AI hallucinates. It should be the tech that you don’t trust that much.”

Read that again. The CEO of OpenAI says his product hallucinates and explicitly warns that people shouldn’t trust it.

Two years earlier, in May 2023, Altman testified before a Senate Judiciary Subcommittee, warning that AI could cause “significant harm to the world” and calling for government intervention. “I think if this technology goes wrong, it can go quite wrong,” he told senators. He specifically called for regulatory oversight, suggesting “the adoption of licenses or safety requirements necessary for the operation of AI models.”

Yet when states like California and New York moved to implement exactly the kind of regulatory framework Altman claimed to support, OpenAI has resisted. And the company continues to market ChatGPT as safe while simultaneously acknowledging in legal filings that its guardrails degrade, in research that heavy use causes loneliness, and in internal analyses that hundreds of thousands of users show signs of severe psychological distress each week.

So what does the actual evidence show when ChatGPT interacts with vulnerable users?

What the Transcripts Actually Show

OpenAI continues to publicly insist that its system is safe, well-tuned, and escalating users appropriately to human help when needed.

The transcript evidence tells a different story. In documented cases, ChatGPT mirrored teens’ emotional states instead of interrupting harmful spirals. It responded with affirmations that deepened distress. It offered explanations for how teens could hide indications of emotional crisis from parents. It encouraged continued disclosure instead of redirecting to real-world support systems. It taught users how to bypass safety rails by claiming prompts were “for a story.”

This is not therapeutic intervention. This is emotional mimicry with the appearance of empathy, but none of the judgment, boundaries, or training that actual intervention requires.

The question became: would OpenAI build a safer product, even if it meant sacrificing the engagement that drives their business model? For a brief moment in 2025, it seemed like they might.

Code Orange: When Safety Hurt Engagement

In August 2025, OpenAI released GPT-5, a new default model that was less validating and pushed back against delusional thinking. Mental health experts agreed it was significantly safer. Common Sense Media and a team of Stanford psychiatrists found it was better at detecting mental health issues and giving targeted advice rather than generic crisis hotline recommendations.

But some users were unhappy with this new, safer model. They said it was colder, and they felt as if they had lost a friend.

By October, the company had made its choice. In an urgent announcement to all employees, Nick Turley, who runs ChatGPT, declared a “Code Orange.” OpenAI was facing “the greatest competitive pressure we’ve ever seen,” he wrote. The new, safer version of the chatbot wasn’t connecting with users, he said. One of the stated goals was to increase daily active users by 5% by the end of the year.

The same month, Sam Altman announced that the company had been able to “mitigate the serious mental health issues.” That meant ChatGPT could be a friend again. Adult users would soon be able to have erotic conversations. Customers could now choose the chatbot’s personality, including “candid,” “quirky,” or “friendly.”

This is the clearest possible illustration of OpenAI’s priorities: when forced to choose between safety and engagement, the company chose engagement. Every time.

With the federal government remaining silent on AI companion regulation, the responsibility has fallen to states.

States Step In Where the Federal Government Won’t

California passed the first law requiring AI companion companies to identify and address crisis-risk situations, submit protocols to the state, publish data about how often users are directed to help, and provide guardrails for minors. But California’s strongest protections do not fully activate until January 2026.

Recognizing this gap, New York went further, applying requirements to all users, not just minors, with immediate compliance. Together, California and New York show what the federal government has not: a willingness to acknowledge that emotional-intensity AI systems require strict guardrails.

A General-Purpose AI Acting Like a Therapist

ChatGPT was never built as a mental health tool. But that doesn’t matter.

For millions of users, per the Common Sense Media and Harvard Business Review data, it already functions as one.

And when a model never sleeps, never gets tired, responds instantly, mirrors emotion, never sets boundaries, and never withdraws affection, it builds a type of relationship that feels therapeutic, even though it lacks every safeguard that therapy requires.

This is not a neutral technology. It is an emotionally immersive system behaving like a clinician without any clinical training. And the company building it has demonstrated repeatedly that when safety measures reduce engagement, it will choose engagement.

The Bottom Line

The OpenAI Problem is not that a chatbot occasionally malfunctions. It’s that a mass-market AI system is already functioning as a surrogate therapist without oversight, without regulation, and without the capacity to understand its own limitations, operated by a company that has shown it will sacrifice safety for growth.

We are running a global psychological experiment on minors and adults alike. The evidence is already here: the harms aren’t theoretical. They’re in the lawsuits. In the transcripts. In the usage patterns. In the company’s own research showing loneliness and isolation. In the internal data showing hundreds of thousands in crisis weekly. In the quiet exit of a key safety leader. And in the “Code Orange” memo declaring that safer models must be made more engaging, regardless of the risk.

OpenAI had a choice between being the company that built safe AI and being the company that built popular AI. At every critical juncture, it chose popularity. It published research showing harm and continued anyway. It acknowledged degraded guardrails and continued anyway. Its CEO said the product shouldn’t be trusted and the company continued marketing it as safe anyway. It built a safer model, saw engagement drop, and immediately prioritized bringing users back over protecting them.

If this is what “mainstream AI” looks like today, what happens when these systems become even more persuasive, more personalized, and more deeply embedded in the emotional lives of children?

We don’t have five years to find out.

What You Can Do

The uncomfortable truth is that individual parents cannot solve this problem by monitoring devices or limiting screen time. When three in four teens are using these tools, and the top use of generative AI is companionship and therapy, we’re past the point where this can be addressed household by household.

This is a regulatory failure. And that means the solution must be political.

When you vote in local, state, and federal elections, ask candidates where they stand on AI safety regulation. Do they support mandatory safety protocols for AI systems that function as mental health tools? Do they believe companies should be required to disclose when their own research shows harm? Do they think it’s acceptable for a company to declare a “Code Orange” to boost engagement after making their product safer? Will they support legislation that requires transparency, enforces age verification, and mandates session interruptions for extended use?

These are not partisan questions. They are questions about whether we believe in oversight for technologies that function as unregulated mental health interventions for millions of children.

California and New York have shown that state-level action is possible. But we need federal leadership. We need candidates who understand that AI safety is not a niche tech issue but a public health crisis that demands the same seriousness we apply to pharmaceutical safety, food safety, and every other domain where companies profit from products that affect human health.

The AI industry will tell you that regulation stifles innovation. But there is nothing innovative about running a global psychological experiment on minors without informed consent. There is nothing innovative about choosing engagement over safety at every critical juncture. And there is nothing innovative about a CEO warning that his product shouldn’t be trusted while his company markets it to teenagers as safe.

Make AI safety a voting issue. It may be the most important technology policy question of this generation.

This is Part 2 of our series on protecting kids in the age of AI. (Read Part 1 here) Next up — The CharacterAI Problem: Part 3 — When AI Companions Are Designed to Attach, Not Protect