Meta blames hallucinations after its AI claims Trump rally shooting didn’t happen

Meta’s AI assistant mistakenly concluded that the recent assassination attempt on former President Donald Trump did not take place, a mistake one of the company’s executives attributes to the technology that powers its chatbot and other solutions.

IN company blog post published Tuesday, Joel Kaplan, Meta’s global head of policy, calls its AI’s responses to questions about the shooting “unfortunate.” He says the Meta AI was originally programmed not to answer questions about the attempted attack, but the company removed that restriction after people began to notice it. He also acknowledges that “in a small number of cases, the Meta AI continued to provide incorrect answers, including occasionally claiming that the event did not occur — something we are working to quickly address.”

“These types of responses are known as hallucinations, which is an industry-wide problem that we see across all generative AI systems and poses an ongoing challenge to how AI deals with real-time events in the future,” continues Kaplan, who leads Meta’s lobbying efforts. “Like all generative AI systems, models can return inaccurate or incorrect results, and we’ll continue to address these issues and improve these features as they evolve and as more people provide feedback.”

Meta’s Not the Only One Involved: Google on Tuesday he also had to refute the claims that the autocomplete search function censored results about the attempted assassination. “Here we go again, another attempt at ELECTION TARGET!!!” Trump he said in the post on Truth Social. “GO TO META AND GOOGLE.”

Since ChatGPT came onto the scene, the tech industry has been grappling with the problem of limiting generative AI’s tendency to fall for falsehoods. Some players, like Meta, have tried to ground their chatbots in high-quality data and real-time search results as a way to compensate for hallucinations. But as this particular example shows, it’s still demanding to overcome what huge language models are inherently designed to do: invent things.