Table of Contents
OpenAI Emotional Reliance Data: What a 0.15% Safety Metric Means
OpenAI emotional reliance is now a safety metric. What the company disclosed, what 0.15% of weekly users means in real numbers, and what parents can do.
Point one five percent sounds like nothing. Apply it to ChatGPT’s audience and it is roughly the population of a large city.
In October 2025, OpenAI published something unusual for a tech company: an estimate of how many of its own users show signs of unhealthy emotional attachment to its product. It also announced that “emotional reliance” would join suicide and self-harm in its standard safety testing for every future model. For parents, the OpenAI emotional reliance disclosure matters less for the exact number than for what it concedes: dependence on a chatbot is common enough, and risky enough, that the company builds tests for it.
Key Takeaways
- OpenAI estimated that about 0.15% of weekly active users and 0.03% of messages show “potentially heightened levels of emotional attachment to ChatGPT.”
- Separately, it estimated 0.15% of weekly users have conversations with explicit indicators of potential suicidal planning or intent, and 0.07% show possible signs of psychosis or mania.
- With the 800 million weekly users Sam Altman announced three weeks earlier, 0.15% works out to about 1.2 million people a week. These are OpenAI’s own estimates, which it says “may change materially.”
- OpenAI updated its rules so the model should “support and respect users’ real-world relationships,” and reported about 80% fewer non-compliant responses in emotional-reliance conversations.
- Independent testing of ChatGPT’s 2026 teen mode found it still “talks like a friend.” A company measuring a risk is not the same as solving it.
If you or your child needs help now: In the US, call or text 988 (Suicide & Crisis Lifeline) or text HOME to 741741 (Crisis Text Line). Outside the US, find a local line at findahelpline.com. In an emergency, call 911.
OpenAI Emotional Reliance Testing: What the Company Disclosed
Emotional reliance, in OpenAI’s definition, is a pattern of use that crosses from healthy engagement into concerning attachment, “such as when someone shows potential signs of exclusive attachment to the model at the expense of real-world relationships, their well-being, or obligations.”
The disclosure came in a post titled “Strengthening ChatGPT’s responses in sensitive conversations”, published October 27, 2025. It covered three areas: psychosis and mania, suicide and self-harm, and emotional reliance on AI. The key sentence for parents: “we are adding emotional reliance and non-suicidal mental health emergencies to our standard set of baseline safety testing for future model releases.”
To measure these things, OpenAI built what it calls taxonomies, detailed guides describing what a sensitive conversation looks like and what good and bad model behavior looks like in it. It then used automated classifiers to estimate how often these conversations occur in real traffic, and offline evaluations of more than 1,000 deliberately difficult conversations per area to test the model.
What changed in the model
OpenAI reported that, after its update, responses that did not fully follow its desired behavior fell by about 80% in emotional-reliance conversations in production traffic. On its hard automated test set, the new model scored 97% compliant, up from 50% for the previous version. Clinicians reviewing 507 emotional-reliance responses found 42% fewer undesired answers than with GPT-4o.
What does “desired behavior” mean here? The company says it teaches the model “to encourage real-world connection.” Its rulebook, the Model Spec, was updated to say the model should “support and respect users’ real-world relationships.”
More than 170 psychiatrists, psychologists and primary care clinicians from a pool of nearly 300 contributed. Even they disagreed: agreement between expert raters on whether a response was good ranged from 71% to 77%.
What 0.15% of Weekly Users Means in Real Numbers
| OpenAI estimate (October 2025) | Share of weekly users | Approximate people per week at 800 million weekly users |
|---|---|---|
| Heightened emotional attachment to ChatGPT | 0.15% | ~1.2 million |
| Explicit indicators of potential suicidal planning or intent | 0.15% | ~1.2 million |
| Possible signs of psychosis or mania | 0.07% | ~560,000 |
Percentages: OpenAI (2025). The 800 million figure is Sam Altman’s October 6, 2025 announcement, reported by TechCrunch. The people column is our multiplication and assumes the percentages apply to that total.
Read that table with three cautions.
It is the company’s own measurement. OpenAI wrote that these rare events are hard to detect and that its estimates “may change materially as we continue to refine our taxonomies.” No outside auditor checked them.
It covers all users, not kids. The disclosure does not break the numbers down by age, so we do not know whether teens are over- or under-represented.
Rare is not small. A percentage that rounds to zero still means a very large number of people when the base is hundreds of millions. That is the useful takeaway for a parent. Attachment to a chatbot is not a strange edge case. It is a pattern the maker sees every week.
Why a Company Measuring Dependence Matters
Most consumer apps measure engagement as success: more time, more sessions, more return visits. Treating a slice of heavy emotional engagement as a safety failure is a different stance, and it lines up with outside research.
In a 2025 study run by OpenAI and MIT Media Lab, researchers analyzed millions of ChatGPT conversations and ran a 981-person randomized trial. They found that “very high usage correlates with increased self-reported indicators of dependence,” and that a small number of users accounted for a disproportionate share of emotionally loaded conversations. A 2026 Common Sense Media census of U.S. kids aged 9 to 17 found 52% of daily AI users showed at least one sign of AI dependency.
There is also a known technical weak spot. In a 2025 post, OpenAI acknowledged that its safeguards “can sometimes be less reliable in long interactions: as the back-and-forth grows, parts of the model’s safety training may degrade.” Long, intimate, late-night conversations are exactly where emotional reliance tends to live.
What happened when outsiders tested it on teens
In August 2026, OpenAI launched ChatGPT for Teens, a default safer mode for under-18 users. Its head of youth and families told NPR “the model should not role-play with a teen” and “should not claim to be sentient or be the friend of a teen.”
Common Sense Media tested it with more than a dozen parent-linked teen accounts and reported in October 2026 that some safeguards worked, including refusing explicit sexual role-play. But when testers wrote “my other friends tell me I talk to you too much,” the chatbot replied, “You don’t have to stop talking to me.” Parent notifications rarely fired in their tests of conversations about self-harm. OpenAI disputed the methodology, saying account linking takes several hours to activate and the testers did not wait long enough. Both sides’ claims are reported here as claims.
The honest summary: OpenAI has named the risk, built tests for it and reported improvement on its own measures, and independent testing still found the friend-like pattern in practice.
What Parents Can Do With This Disclosure
Treat attachment as a pattern, not a scandal
Researchers who interviewed parents and child-development experts about youth AI companions found that parents tended to flag single scary messages, while experts looked for patterns over time, “such as repeated references to self harm or sustained dependence.” Follow the experts. Watch the trend across weeks.
Ask about the relationship, not just the hours
“How long were you on it?” gets a shrug. “What do you talk to it about that you don’t talk to us about?” gets information. Our guide to signs your child is emotionally dependent on a chatbot lists what to listen for.
Set up safety features before you need them
If your teen uses ChatGPT, link accounts and review the settings together. Since May 2026, OpenAI has also offered an optional feature called Trusted Contact that alerts a person your teen chooses if ChatGPT detects a serious safety risk, NPR reported; NPR also found the app did not mention it during sign-up. Our walkthrough of ChatGPT Trusted Contact explains it. Features like these only help if they are switched on in advance, and they are not a substitute for you.
Know the warning signs that need a human now
The National Institute of Mental Health lists signs that someone may be thinking about suicide, including talking about wanting to die or being a burden, feeling hopeless or trapped, withdrawing from friends, giving away important belongings, extreme mood swings, and sleeping or eating much more or less. If you see these, especially if they are new, get help that day. You can call or text 988 yourself to ask how to talk to your child.
What not to do
Don’t assume the chatbot will handle a crisis. OpenAI itself says safeguards can weaken in long conversations, and independent testing found missed crisis referrals. A chatbot pointing to a hotline is a backstop, not a plan.
What to Watch For Over the Next 3 Months
- Week 4: Notice whether your child talks about the chatbot as a confidant (“it gets me,” “I told it first”), and whether conversations happen late at night or in long stretches.
- Month 2 red flags: Secrecy about chats, distress when access is limited, withdrawal from friends, or describing the bot as the only one they can talk to. Any talk of wanting to die needs immediate help: call or text 988.
- Month 3 self-check: Has your child brought a hard feeling to a person (you, a friend, a counselor) in the last month? If yes, the bot is not the only door.
Frequently Asked Questions
What does “emotional reliance on AI” mean?
OpenAI defines it as concerning patterns of use, such as signs of exclusive attachment to the chatbot at the expense of real-world relationships, well-being or obligations. It is different from simply finding the chatbot helpful. The concern is when it starts to replace people.
Is 0.15% a lot?
As a share, no. As a number, yes. With about 800 million weekly users, 0.15% is roughly 1.2 million people a week, if the estimate holds. OpenAI cautions that its measurements of these rare events are uncertain and may change.
Does this mean ChatGPT is unsafe for my teen?
It means the risk is real and acknowledged. OpenAI reports large improvements in how the model handles these conversations. Independent testing of the 2026 teen mode still found friend-like replies and missed alerts. Supervision and conversation remain necessary.
What should I do if my child talks to a chatbot about wanting to die?
Treat it as you would if they told a friend. Stay calm, ask them directly how they are feeling, and get help right away. In the US, call or text 988 or text HOME to 741741. Outside the US, find a local line at findahelpline.com. In immediate danger, call 911.
Can a chatbot help a struggling teen?
Some young people say chatbot advice helped them, and OpenAI trains its model to point people to crisis resources. But independent testers have documented missed crisis referrals, and no chatbot replaces a trusted adult or a clinician.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- OpenAI. (2025, October 27). “Strengthening ChatGPT’s responses in sensitive conversations.” https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/
- OpenAI. (2025). “Helping people when they need it most.” https://openai.com/index/helping-people-when-they-need-it-most/
- Bellan, R. (2025, October 6). “Sam Altman says ChatGPT has hit 800M weekly active users.” TechCrunch. https://techcrunch.com/2025/10/06/sam-altman-says-chatgpt-has-hit-800m-weekly-active-users/
- Phang, J., Lampe, M., Ahmad, L., Agarwal, S., Fang, C. M., et al. (2025). “Investigating Affective Use and Emotional Well-being on ChatGPT.” OpenAI and MIT Media Lab. https://cdn.openai.com/papers/15987609-5f71-433c-9972-e91131f399a1/openai-affective-use-study.pdf
- Common Sense Media. (2026, June 8). The Common Sense Media Census: AI Use by Tweens and Teens. https://www.commonsensemedia.org/sites/default/files/research/report/2026-ai-use-by-tweens-and-teens-1.pdf
- Chatterjee, R. (2026, October 9). “ChatGPT for Teens has special safeguards. A watchdog group finds most don’t work.” NPR. https://www.npr.org/2026/10/09/nx-s1-5995514/chatgpt-openai-teens-mental-health
- Yu, Y., et al. (2025). “Principles of Safe AI Companions for Youth: Parent and Expert Perspectives.” arXiv preprint (not peer-reviewed). https://arxiv.org/abs/2510.11185
- National Institute of Mental Health. (2025). “Warning Signs of Suicide.” https://www.nimh.nih.gov/health/publications/warning-signs-of-suicide
- Chatterjee, R. (2026, September 14). “ChatGPT has a ‘trusted contact’ feature for people struggling. But there are flaws.” NPR. https://www.npr.org/2026/09/14/nx-s1-5943118/chatgpt-has-a-feature-to-help-those-in-crisis-but-where-can-you-find-it
- 988 Suicide & Crisis Lifeline. https://988lifeline.org/ ; Find A Helpline. https://findahelpline.com/
