Meta's AI assistant recently misreported that the attempted assassination of former President Donald Trump did not occur. This mistake, identified by company executives, has been attributed to the underlying technology of its chatbot.
In a blog post published by Joel Kaplan, Meta’s global head of policy, he described the AI's responses regarding the incident as "unfortunate." Initially, Meta AI was programmed to avoid discussing the assassination attempt, but the restriction was lifted as user interest grew. Despite this, Kaplan noted that "in a small number of cases, Meta AI continued to provide incorrect answers, including wrongly asserting that the event didn’t happen." The company is addressing these inaccuracies promptly.
Kaplan highlighted that such erroneous responses are known as "hallucinations," a common issue across all generative AI systems and a significant challenge for handling real-time events. He emphasized, "Like all generative AI systems, models can produce inaccurate or inappropriate outputs, and we’re committed to improving these features as they evolve and incorporate user feedback."
This issue is not unique to Meta. Google also faced criticism regarding its Search autocomplete feature, which was accused of censoring information related to the assassination attempt. In response, Trump commented on Truth Social, alleging election manipulation and targeting Meta and Google for scrutiny.
Since the emergence of ChatGPT, the tech industry has been wrestling with the challenge of minimizing generative AI's tendency to generate falsehoods. Companies like Meta strive to ground their chatbots with reliable data and up-to-date search results, yet, as illustrated by this incident, overcoming the inherent propensity of large language models to fabricate information remains a complex endeavor.