Former OpenAI Safety Lead: “Don’t Trust Their Claims About AI Safety or Erotica”

A former OpenAI safety lead, Steven Adler, wrote in October 2025 that people should not simply take OpenAI’s word that ChatGPT is safe for emotionally vulnerable users. You do not need to side with him or with OpenAI to find the question useful. This page is about the user’s side of it: how to read any company’s safety claims, and what you can do yourself. For a summary of the essay, see our news page on it.

What he asked for

According to Futurism, Adler wrote that people deserve more than a company’s word that safety issues are addressed, and that OpenAI should show evidence before introducing adult content. He also said OpenAI’s release of mental health data was a good first step but lacked comparisons with earlier months. That is a request for evidence, not an accusation about any one person’s conversations.

How to read a safety claim

Safety announcements from AI companies can be true and still leave out what you want to know. Questions worth asking:

• Is there a number, and who measured it? A company’s own test of its own model is one kind of evidence. An outside review is another.
• Compared with what? A rate with no earlier rate to compare against does not show a trend.
• What does it leave out? A figure about how often a chatbot gives a safe answer in a test says little about long, late-night conversations.
• Who is affected by the gap? A feature that works for most users can still fail the few who most need it.

OpenAI itself has said its safeguards can be less reliable in long conversations, where parts of the model’s safety training may degrade (reported by Yahoo News). That does not mean ChatGPT is unsafe for everyone. It does mean the length and tone of your own use matter.

What the research says about heavy use

A four-week study by MIT Media Lab and OpenAI followed roughly 1,000 adults. Most did not use ChatGPT in an emotional way. Heavy daily use was linked with more loneliness and more reliance on the chatbot, though that does not show that the chatbot caused it (MIT Media Lab). Someone who is already lonely may simply use it more. The study has not been peer reviewed.

A Stanford-led paper on AI and therapy found that chatbots can respond inappropriately to serious symptoms, for example by encouraging delusional thinking, probably because they tend to agree with users (arXiv abstract). Agreeable is not the same as helpful.

What you can do

You cannot audit a company. You can look after your own use.

• Notice what you use it for. Homework, drafting and brainstorming carry different risks than late-night talk about your worst fears.
• If you talk to it about feelings, treat it as a place to write things down, not as someone who knows you.
• Keep one or two real people in the loop for anything important or painful.
• Take breaks. If you feel uneasy when you can’t open it, that’s worth noticing.
• Be cautious with any chatbot that always agrees with you. Ask it to argue the other side.
• Do not use it as a crisis service.

The American Psychological Association has said that general AI chatbots are no substitute for care from a trained professional (APA, November 2025).

“AI addiction” is not a formal diagnosis. A habit can still crowd out sleep, work or relationships.

Thoughts of suicide or self-harm are a reason to reach out right now. In the US, call or text 988, or text HOME to 741741 to chat with Crisis Text Line. SAMHSA runs a free helpline at 1-800-662-4357. Dial 911 for immediate danger.

Should I stop using ChatGPT because of this?

Not necessarily. The question is how you use it. If it helps with tasks and you still lean on people for support, there is little cause for alarm. If it has become your main source of comfort, that’s a reason to add people to the mix.

Why would a chatbot being agreeable be a problem?

If you are anxious or holding a belief that isn’t true, a bot that agrees can strengthen it. A friend or therapist is more likely to push back.

Where can I find independent information on AI safety?

Look for peer-reviewed studies, professional bodies such as the APA and the American Academy of Child and Adolescent Psychiatry, and regulators such as the FTC.

This page is for education only. It is not medical advice, a diagnosis or a treatment plan.