Bard’s error teaches not to blindly trust AIs | Google Machine Learning Course | Benefits of Generative ai | Google Machine Learning Certification Free | Turtles AI

Have you heard about Google's recent AI disaster with Bard? Sure you do. In our previous article about Google's strategy (so far, a failure) to launch Bard, we mentioned the error made by the AI, but we didn't delve into the details. Many readers asked for a deeper analysis and here it is! Bard's story serves as a warning against blindly trusting AI and technology. We all know that knowledge and information accumulate over time, are passed down and, theoretically, shouldn't change. However, not all sources accurately report knowledge. This is the problem with widespread and collaborative sources, such as Wikipedia and the internet in general. Nevertheless, the plurality of sources allows us to "mediate" and fix errors made unconsciously. The danger with a system like a chatbot is that you might rely on it entirely for information. We now know that chatbots are trained on a large number of documents, but sometimes information can be misinterpreted. This is the basis of Bard's error, which, as we previously noted, cost Google $120 billion in market capitalization in just a few hours. In the demonstration of Bard, the chatbot was asked to describe, in words suitable for a 9-year-old child, what new discoveries can be attributed to the James Webb Telescope. After correctly describing some discoveries, the bot confidently stated a piece of incorrect information, saying that the telescope in question "took the first photos of a planet outside our solar system." Several astronomers immediately pointed out that this answer was widely inaccurate, since those photos were taken in 2004, when the James Webb was not yet in operation. The real problem, as I said at the beginning of the article, is the absolute veracity of a potentially incorrect answer, compared to a correct answer from a single source. We could say that Bard was a great liar... but he still lied. He did so "in good faith", certainly deriving the "false belief" from a wrong logical-verbal understanding of the documents used for training. The sentence at the basis of the error, in a NASA document, is the following: "For the first time, astronomers have used NASA's James Webb Space Telescope to take a direct image of a planet outside our solar system." In other words, instead of understanding (from the context) that the James Webb had, for the first time in its own history, taken a direct image, the chatbot distorted and reported the above. This highlights the absolute importance of rigorous - human and multiple - control of every answer given so far by a chatbot, and, plausibly, this will still be true for a long time. Moreover, it highlights the fundamental need for a plurality of sources. The concept is similar to reading multiple newspapers to get as close as possible to the truth about the same news.