#ai-hallucination

#ai-hallucination

"Um ... there we go ... uh-oh," said Meta CEO Mark Zuckerberg on stage as he attempted to answer a video call through a combination of movements between a wristband and a pair of glasses. "Well, I ... let's see what happened there ... that's too bad," he continued, shortly before cutting short the live demo. The video call went unanswered.

"What we had noticed was there was an underlying problem with our data," Ahuja said. When her team investigated what had happened, they found that Salesforce had published contradictory "knowledge articles" on its website."It wasn't actually the agent. It was the agent that helped us identify a problem that always existed," Ahuja said. "We turned it into an auditor agent that actually checked our content across our public site for anomalies. Once we'd cleaned up our underlying data, we pointed it back out, and it's been functional."

A few months ago, I asked ChatGPT to recommend books by and about Hermann Joseph Muller, the Nobel Prize-winning geneticist who showed how X-rays can cause mutations. It dutifully gave me three titles. None existed. I asked again. Three more. Still wrong. By the third attempt, I had an epiphany: the system wasn't just mistaken, it was making things up.

Popular right wing influencer Charlie Kirk was killed in a shooting in Utah yesterday, rocking the nation and spurring debate over the role of divisive rhetoric in political violence. As is often the case in breaking news about public massacres, misinformation spread quickly. And fanning the flames this time was Elon Musk's Grok AI chatbot, which is now deeply integrated into X-formerly-Twitter as a fact-checking tool - giving it a position of authority from which it made a series of ludicrously false claims in the wake of the slaying.

"Language models are optimized to be good test-takers, and guessing when uncertain improves test performance," the authors write in the paper. The current evaluation paradigm essentially uses a simple, binary grading metric, rewarding them for accurate responses and penalizing them for inaccurate ones. According to this method, admitting ignorance is judged as an inaccurate response, which pushes models toward generating what OpenAI describes as "overconfident, plausible falsehoods" -- hallucination, in other words.