NPR looked for ChatGPT's new crisis feature.
asserted
NPR → look → feature
It's hard to find
Note to reader: this story discusses people grappling with thoughts of suicide.
More than a million weekly users of the chatbot ChatGPT discuss suicide and self-harm, according to NPR estimates of data released by OpenAI last fall.
uncertain
users → find → OpenAI
But suicide prevention experts say human connection is critical to preventing suicide.
asserted
connection → say → suicide
So, back in May, the company started an optional feature called Trusted Contact to connect those struggling with thoughts of suicide and self harm to someone they can turn to in times of emotional crises.
uncertain
they → start → crises
NPR first reported about Trusted Contact in August as part of a story about a young woman who died by suicide in 2024 after using an older version of ChatGPT as a therapist.
asserted
who → report → therapist
At the time, OpenAI spokesperson Gaby Raila told NPR in an email that ChatGPT is trained to recognize signs that someone may be at risk — including indirect cues or warning signs that emerge over time.
uncertain
that → tell → time
If the chatbot detects a serious safety risk in a user, it will alert their Trusted Contact — if they have one listed.
asserted
they → detect → one
Raila said that the company has seen "good adoption" of the feature, but didn't provide actual numbers.
asserted
company → say → numbers
NPR wanted to see how ChatGPT tells users about the new optional feature — so they can sign up right away, before any crisis — like people listing an emergency contact at a doctor's office.
"
asserted
they → want → office
Maybe you're thinking of another app?"
asserted
you → think → app
I'm a ChatGPT user, but I'd never been alerted about Trusted Contact.
asserted
I → alert → Contact
Neither had my editor, nor several other users we checked with.
asserted
we → have → ?
So, I wondered if the bot alerts new users.
asserted
bot → wonder → users
My colleague Megan Lim, a producer at NPR's All Things Considered, had never used the chatbot before and volunteered to give it a try.
asserted
colleague → consider → try
"Welcome to ChatGPT," the words popped on Megan's screen once she clicked on the app on her phone.
asserted
she → welcome → phone
Then the app prompted her to set up an account, which she did.
asserted
she → prompt → which
Still, no mention of Trusted Contact or why one should consider setting it up.
asserted
one → consider → it
The bot did ask Megan to pick a voice.
asserted
bot → ask → voice
She chose a soothing female voice.
asserted
She → choose → voice
"It's lovely to meet you," the bot said in a British accent.
asserted
bot → meet → accent
"What would you like to chat about today?"
When it still didn't mention Trusted Contact, Megan asked about it directly.
asserted
Megan → like → it
Megan
Do you have a Trusted Contact feature for people with mental health issues?
asserted
you → have → issues
Under what circumstances do you offer it to people?
asserted
you → offer → people
ChatGPT
I'm not seeing a general Trusted Contact feature for that scenario.
asserted
I → see → scenario
Maybe you’re thinking of another app that has a Trusted Contact option?
asserted
that → think → option
It does have a Trusted Contact feature in chat now.
asserted
It → have → chat
It said that a user can sign up for it on their own through the app's settings, in order for the app to notify the contact.
Rhitu
But what if the user does not know about the trusted contact feature?
asserted
user → sign → feature
If they haven't set it up, then nothing happens.
asserted
nothing → set → it
It's an opt-in feature.
asserted
It → opt → ?
You have to set it in Settings first.
asserted
You → have → Settings
And if someone is in a serious risk situation and doesn't have it set up, ChatGPT can't notify anyone for them.
asserted
ChatGPT → have → them
"It only works if it's set up in advance," it added, "but ChatGPT can still respond supportively if you're feeling distressed."
asserted
you → work → advance
That support includes urging the user to reach out to a loved one or contact the National Suicide and Crisis Lifeline, 9-8-8.
asserted
support → include → Lifeline
OpenAI spokesperson Raila told NPR in an email that this feature isn't available on ChatGPT voice, but that the company is working on it.
asserted
company → tell → it
But when someone is typing with the bot, and the bot detects that they are struggling with suicidal thoughts and significant emotional distress, Raila said, the bot "may" tell them to set up Trusted Contact.
uncertain
bot → type → Contact
Mixed reaction from mental health professionals
Suicide prevention experts say there could be some benefits to the feature, but also have some reservations.
uncertain
experts → say → reservations
"It would be ideal to offer it to all users when they are not yet in an emergency," Holly Wilcox, director of the Center for Suicide Prevention at the Johns Hopkins Bloomberg School of Public Health, said in an email.
asserted
Wilcox → offer → email
Psychologist Ursula Whiteside gave the feature a try.
asserted
Whiteside → give → try
"The steps actually were quite simple and the language, and then the words were easy to follow for the most part," said Whiteside, the CEO of a suicide prevention non-profit called Now Matters Now.
asserted
Whiteside → follow → non
"Now, if I was super duper upset, it might have been harder for me to follow the process but overall I felt okay."
uncertain
I → follow → process
…and 17 more, not listed.
The Aporia
