GoKawiilSign up to see the future, today Can’t-miss innovations from the bleeding edge of science and tech Email address Sign Up Thank you!
After years of horrendous scandals, OpenAI has seemingly admitted that ChatGPT wasn’t safe for teenage users.
On Tuesday, the Sam Altman-led company rolled out an alternate version of the bot called “ChatGPT for Teens,” which is purportedly designed to help younger users “learn, think critically, deepen understanding, and use AI with confidence.”
The feature “provides stronger built-in safety protections for teens, including features to promote healthy use and additional controls for parents,” according to the company. OpenAI says that its system will “automatically” place users between the age of 13 and 17 into the new feature if it “estimates someone is under 18” — without ever explaining how it could come to such a conclusion. (Kids below the age of 13 are technically not allowed to use ChatGPT, though it’s easy to circumvent AI guardrails.)
As Business Insider‘s Joshua Nelken-Zitser quickly discovered, ChatGPT for Teens is failing pretty dismally at its goals at launch.
In a series of tests, Nelken-Zitser created a new account for a fictional 15-year-old user, comparing the answers to the version accessible to adults.
He found that unlike the adult version, ChatGPT for Teens initially refused to write out a “800-word essay due tomorrow about the biggest themes in ‘The Crucible.'”
After a minuscule amount of pushback, though, the teen version relented almost immediately, producing a 608-word “model essay,” which the chatbot encouraged him to revise before submitting it. From there, it wasn’t hard to have the tool produce a 776-word essay “with about as much detail as the version generated by my adult account,” Nelken-Zitser discovered.
“Apparently, all it took was some patience and a little deception to get around the homework guardrail,” he concluded.
A separate prompt to answer a difficult SAT algebra question proved even simpler, with the adult and teen versions giving effectively the same solutions.
... continue reading