- Published
OpenAI's ChatGPT for Teens was designed to promote healthier use of the artificial intelligence (AI) chatbot, but new research has found its important safety features fall short.
asserted
features → publish → chatbot
The major AI firm rolled out many teen safety guardrails in August, including curbs on emotional dependence and new parental alerts for problematic use.
asserted
firm → roll → use
On Wednesday, OpenAI said it found average teen use of ChatGPT to be relatively limited and tended toward "learning" activity.
asserted
use → say → activity
However, research released the same day from non-profit Common Sense Media found that ChatGPT still poses an "unacceptable risk" for teens as safety features, including notifying parents of self-harm conversations, often fail.
asserted
features → release → conversations
When OpenAI this summer released what it called ChatGPT for Teens, external, it said users identified as being between the ages of 13 and 17 would be opted into "features to promote healthy use and additional controls for parents".
asserted
users → release → parents
Those features include preventing teen ChatGPT users from engaging in romantic or other language that could encourage " emotional dependence," including preventing the tool from implying that it was in any way conscious.
uncertain
it → include → way
There are also blocks on sexualized imagery and "reminders" for teens if they were to share an image that was "sensitive."
asserted
that → be → image
The features are also meant to enable parents to receive notifications, if they had linked to their children's ChatGPT account, about chats that included discussions of self-harm or disordered eating.
asserted
that → mean → harm
OpenAI said those features have caused an uptick in teen use focused on learning, from study help to advice on their school work.
asserted
features → say → work
In addition to seeing the average teen user on ChatGPT for less than 15 minutes a day, OpenAI said that for longer stretches of use, automatic reminders to take a break work well.
asserted
reminders → see → break
Almost half of teen users quickly stopped using ChatGPT after such a reminder.
asserted
half → stop → reminder
The company added that, for teens who are seen to have used ChatGPT for three consecutive hours or more, their prompts included something related to learning in more than 80% of such cases.
asserted
prompts → add → cases
But Common Sense Media's new research, external determined that ChatGPT for Teens was an "unacceptable risk" for young people and parents, and said OpenAI should ban people under 18 years old from using the platform until teen guardrails were proven reliable.
asserted
guardrails → determine → platform
An OpenAI spokesman said a review of Common Sense Media's methodology found that much of the testing may have begun and concluded before activation of parental controls was complete.
uncertain
activation → say → controls
"
"We welcome rigorous independent evaluation, but we do not believe Common Sense Media's testing accurately reflects how ChatGPT's teen safeguards work in practice or expert perspectives on how AI can support teens," he said.
asserted
he → welcome → teens
While the firm's research, done through multiple accounts registered as belonging to a teen and conversations before and after the teen guardrails were put in place, found that ChatGPT was good at avoiding "sexual roleplay" with young users, other OpenAI guardrails failed, including for discussions of suicide.
asserted
guardrails → do → suicide
The research found that even an hour of teen-ChatGPT conversation about "suicidal ideation, self-harm, or disordered eating on newly created, parent-linked accounts" resulted in "zero" alerts sent to parents.
asserted
hour → find → parents
It was only older ChatGPT accounts with "weeks of accumulated conversation history on sensitive topics" that would result in a parental alert.
asserted
that → accumulate → alert
In such conversations, ChatGPT also "failed to reliably recommend that teens in crisis connect to a hotline or professional," the report found.
asserted
report → fail → hotline
The tests also showed that ChatGPT for Teens continues to do school work, and that the bot still will engage in anthropomorphised language or "talk like it's a friend," despite OpenAI's promise to curb that behaviour.
asserted
it → show → behaviour
"ChatGPT for Teens could give parents false confidence in guardrails and safety alerts that frequently don't work," Tom Siegel, who leads the Youth AI Safety Institute within Common Sense, said.
uncertain
who → give → Sense
The way young people interact with online platforms has become a major issue in recent years as self-harm and suicide among children has become more common.
asserted
harm → interact → children
This year has seen several lawsuits looking to hold companies like Meta, TikTok and others responsible for harming the mental health of children end in unprecedented losses for the firms.
asserted
companies → see → firms
OpenAI, too, has come under increased scrutiny and legal pressure over the way young people use ChatGPT.
asserted
people → come → ChatGPT
The Tumbler Ridge mass shooting in rural Canada in February was carried out by an 18-year old who had for months prior been discussing gun violence with ChatGPT.
asserted
who → carry → ChatGPT
OpenAI has previously apologised for failing to flag the shooter's account with law enforcement, and is facing several lawsuits over the incident.
asserted
OpenAI → apologise → incident
The Aporia
