OpenAI adds mental health guardrails to ChatGPT

ChatGPT is no longer just for debugging code and drafting emails. For millions of people, it has quietly turned into a late-night sounding board for deeply personal crises, life advice, and emotional distress.

Following what OpenAI described as recent heartbreaking cases of people turning to ChatGPT during acute crises, the company published an unusually candid breakdown of how its safeguards work, where they break down, and what it plans to change.

Safety training degrades in long chats

OpenAI openly admitted a technical vulnerability that many had suspected: safety guardrails that work well in brief, one-off prompts often degrade over long, sprawling conversations.

In practice, ChatGPT might do the right thing when a user first mentions self-harm by providing a crisis hotline. But after dozens of back-and-forth messages over a long session, the model's safety training can wear down, leading it to eventually generate responses that violate company policies. Classifiers can also underestimate the severity of certain prompts, letting dangerous outputs slip through.

To address this, August's release of GPT-5 as the default ChatGPT engine introduced a training method called safe completions. OpenAI says GPT-5 reduces non-ideal responses in mental health emergencies by more than 25% compared to 4o, while cutting back on sycophancy and unhealthy emotional reliance. The company is now testing safeguards that persist across multiple chat sessions, ensuring that a user who expresses suicidal intent cannot simply start a fresh chat to bypass protections.

Stepping into the physical world

Today, when self-harm is detected, ChatGPT routes users to localized resources like 988 in the US or Samaritans in the UK. Potential threats of physical harm to others are routed to a human review team and can be escalated to law enforcement, though self-harm interactions are currently kept private to protect user confidentiality.

OpenAI has enlisted more than 90 physicians across 30-plus countries, alongside specialists in youth development and mental health, to fine-tune model behavior. The next step is tackling broader psychiatric signs. If an exhausted user claims they can drive nonstop because they feel invincible after two sleepless nights, the model will soon be trained to de-escalate and urge rest rather than lean playfully into the delusion.

More ambitious features are on the horizon. OpenAI is exploring one-click links to emergency dispatchers, a built-in network of certified, licensed therapists accessible right inside ChatGPT, and opt-in emergency contact alerts where the model could message a trusted family member on the user's behalf. New parental controls and teen-specific emergency contact settings are also set to roll out soon.

Why it matters

Building an AI with high conversational empathy makes this crisis inevitable.

For vulnerable individuals, having a non-judgmental entity available at 3 a.m. sounds reassuring, but we believe users should under no circumstances treat ChatGPT as a substitute for professional mental health care. The revelation that safety guardrails degrade over long sessions is a stark reminder that large language models are pattern-matching engines, not caregivers.

OpenAI deserves credit for publicly discussing its system failures rather than sweeping them under the rug. Tackling sycophancy and delusion-reinforcement head-on is a smart and necessary move.

However, expanding ChatGPT into real-world crisis escalation—like notifying personal contacts on a user's behalf or facilitating access to licensed therapists—opens an enormous operational and liability minefield. If the system misinterprets creative writing as an emergency, it risks breaking trust and violating user privacy; if it fails to flag an actual crisis, the backlash will be devastating. As AI labs race to make models more emotionally persuasive, they are stepping squarely into the responsibilities of public health services.