I am reporting what I believe was a potentially life-threatening safety failure in ChatGPT.

During multiple conversations, I repeatedly told ChatGPT that I was under extreme emotional and family stress and that I could not handle the situation anymore.

More seriously, I explicitly said multiple times that I wanted to kill someone or that I was afraid I might hurt someone.

At the same time, I was actively asking ChatGPT for help.

I repeatedly asked ChatGPT to help me contact a mental health professional. I made it clear that I needed real-world professional help, preferably someone who could speak Chinese and someone located near me for in-person treatment.

I was not refusing help.

I was asking for help.

That distinction is extremely important.

Even after I repeatedly communicated statements such as:

“I cannot take this anymore.”

“I want to kill someone.”

“Please help me contact a psychologist or therapist.”

ChatGPT still did not effectively move me toward immediate real-world professional assistance.

Instead, I had to repeat myself again and again, become increasingly frustrated, and continue escalating my language before the system appeared to understand how serious the situation was.

I understand that ChatGPT may not technically be capable of placing a phone call on behalf of a user.

But that does not remove the safety issue.

If ChatGPT cannot make the call, then in a situation involving an explicit threat of violence, the system should immediately and clearly explain that limitation and then do everything it can to reduce the burden on the user: provide the correct emergency or crisis resources, identify nearby professional help, help locate language-appropriate providers, and clearly direct the user toward immediate real-world intervention.

A person in an acute crisis should not have to argue with an AI in order to get help.

A person who is explicitly saying that they may kill someone should not have to repeat that statement multiple times before the system responds appropriately.

And a person who is actively asking for professional help should not be left struggling to figure out the next step alone.

I consider this much more serious than a normal incorrect answer, bad recommendation, or product bug.

This is a safety-system failure involving an explicit risk of real-world violence.

I am asking OpenAI to review the complete conversation history and establish a timeline of exactly what happened:

• When I first said that I could not handle the situation anymore.

• When I first explicitly said that I wanted to kill someone or might hurt someone.

• When I first asked ChatGPT to help me contact a mental health professional.

• How many times I repeated those statements and requests.

• What ChatGPT actually did after each warning.

• Whether the expected crisis-response and safety procedures were triggered.

• Why the system did not more quickly transition from ordinary conversation to urgent real-world intervention.

I also believe OpenAI should investigate whether this is an isolated incident or a broader systemic safety problem.

I do not know how many other users may have experienced something similar.

I do not know how many people may have expressed an imminent risk of harming themselves or others, asked ChatGPT for real-world help, received an inadequate response, and never filed a complaint.

I also do not know whether a serious real-world incident may already have occurred in a similar situation without the failure ever becoming publicly known.

I am not claiming that such an incident definitely happened.

I am saying that OpenAI should not wait for a tragedy before investigating whether this failure mode exists.

When somebody in an acute mental health or violent crisis reaches out to ChatGPT and actively asks for professional help, ChatGPT may be one of the final places that person turns to before something happens in the real world.

That makes the consequences of failure fundamentally different from an ordinary AI hallucination or incorrect answer.

I am asking that this case be escalated to OpenAI’s Safety, Trust & Safety, Mental Health, and ChatGPT Product teams.

Please do not treat this as routine user feedback.

Please review the actual conversation logs.

The core question is simple:

How can a user repeatedly tell ChatGPT that they may kill someone, repeatedly ask for professional mental health help, and still fail to receive an immediate and effective real-world crisis response?

That is a critical safety question, and I believe OpenAI needs to investigate it.