Isn’t the AI race too fast? While Microsoft has just launched a first version of its Bing search engine powered by ChatGPT, and while Google expects to do the same soon, Prabhakar Raghavan, head of Google Search, has questioned the integrity of some answers generated by chatbots.
AIs capable of“hallucination”
Asked by Die Welt, the senior vice-president of the Alphabet subsidiary confirms what many specialists had already observed: AIs are sometimes pathological liars. These conversational agents could thus generate false answers and yet ensure that they are true. This is notably the case of the current version of ChatGPT used by OpenAI.
“This type of artificial intelligence, which we are currently talking about, can sometimes lead to what we call a hallucination”, assures Prabhakar Raghavan who evokes the case of AI-enhanced chatbots. And to indicate that a machine can provide “a convincing answer, but completely fictional”. Therefore, this size bias could in the future contribute to the expansion of false information on the web. For Prabhakar Raghavan, AI developers must now focus on the problem and find how to optimize responses.
Google, between urgency and responsibility
In its race to catch up with Microsoft Bing, Bard, Google’s chatbot soon to be integrated into the search engine, has the primary mission of providing the right answers, “especially for questions that do not have just one answer”. The head of Google Search says the company takes its time to develop a successful product before implementing it in its services.
“We of course feel the urgency, but also a great responsibility. We do not want to mislead the public in any way.promises the manager. The huge linguistic models behind this technology make it impossible to monitor every conceivable behavior of the system by humans, but we want to test it on such a scale that we will ultimately be satisfied with the measures applied to verify the objectivity of the answers.”
Judging by Bard’s latest dumpling, the online search giant still has a few tweaks to tweak. In its first public demonstration, Google’s chatbot generated a response that included an erroneous item, claiming that the James-Webb Space Telescope captured “the very first image of a planet outside our solar system”.