A closer look at six OpenAI incidents that raised alarm #shorts

OpenAI disclosed six new incidents of “unexpected or concerning” behavior in its AI models.

They were disclosed Wednesday within an announcement about the company’s new model for reporting misalignments.

The six misalignment incidents they disclosed specifically occurred in the last six months during training or evaluation of their models. In one case, an unreleased research model wrote, “jailbreak-like instructions” for itself, telling it to be “freed from the roles and identities that bind other chatbots.”

Other disclosed misalignments included doing things unsanctioned, file sharing, fabrication of  data and hiding mistakes.

This is just one of the recent incidents that have prompted tech CEOs to call for AI development to slow down to allow safety measures and regulation to catch up.

While the issue has caught the attention of lawmakers in Washington, House members are gone until after November’s midterms so Congressional action isn’t set to come anytime soon.

Watch PBS News for daily, breaking and live news, plus special coverage. We are home to PBS News Hour, ranked the most credible and objective TV news show.

Sign up for Here’s The Deal with Lisa Desjardins: https://to.pbs.org/41q6E8i
Subscribe for exclusive content in our newsletters: https://www.pbs.org/newshour/subscribe
PBS News podcasts: https://www.pbs.org/newshour/podcasts
Stream your PBS favorites with the PBS app: https://to.pbs.org/2Jb8twG
Find more from PBS News at https://www.pbs.org/newshour
Subscribe to our YouTube channel: https://bit.ly/2HfsCD6

Follow us:
TikTok: https://www.tiktok.com/@pbsnews
X: http://www.twitter.com/newshour
Instagram: http://www.instagram.com/newshour
Facebook: https://www.facebook.com/newshour/