AI search engines invent sources for ~ 60% of queries, finds learning

AI search engines are like your friends who claim to be a specialist in each topic, even when they don’t really know what they are talking about, even when they don’t really know what they are talking about. New Research sheet Viewing Colombia journalism (CJR), AI models are open to a news event, or get a story more often or get a mistake or get a mistake.

The researchers fed different models, extract directly from actual news stories, and then asked to identify information, including the title of the article, the publisher and the URL. The confusion returned the wrong information, returned 37 percent of the time, and in the extreme end, the Xai was 97 percent of the groc. Errors, bot, even the URL offers links to articles that do not go anywhere because they are self-evident. In general, researchers broadcast false information for 60 percent of the test request for the AI ​​models.

Sometimes, such as search engines, such as search engines will pass PayWalls National Geographical These websites are not using a creeping text that responds to the search engines normally. In the past, confusion won on this hot water, but the experience claimed to be fair. Investment tried to offer income sharing deals to place the publishers, but still refuses to end the experience.

Not everyone talking in recent years should not be surprised. Chatbots are biased to refund such answers when they are confident. Search, search is active through a technique called expanded generation, which provides real-time internet to ensure the response, rather than relying on a fixed database provided by an AI model manufacturer. This can worsen the inaccuracy as countries such as Russia Nourse search engines with propaganda.

One of the most damaged things from some of the most damaged by ChatBots will accept the text or the logic chain to answer a desire to answer a desire. Anthropik’s Klumu, for example, when he wanted to conduct research, was arrested by entering the “Filler” information.

Mark Howard, Chief Operating Director Time The magazine expressed concern about how CJR’s content was adopted and how it was shown in AI models and the ability to demonstrate. For example, if the users learn the news they receive the reply stories, may damage the brand of publishers Guardian It’s wrong. This has been a last problem BBCwhich The apple was tasked Apple Intellingh recycled written warnings with notification warnings. However, Howard also accused users. From ARS TECHNICA:

However, howard also offers the user’s fault, if they offer the user’s sin, if they are doubtful in the accuracy of free AI instruments: “If any of these free products believe that any one will be 100 percent accurate, it is ashamed of them.”

Expectations should be placed here. People are lazy and they respond to confidently voicing way that conversations can complain. Thoughts on social media demonstrate that people do not want to hit the links and to get the immediate response from Google’s AI views; CJR uses one of the AI ​​models to search in one in one of four Americans. And even before the start of generative AI instruments, more Half of Google Search The user was “zero click” they received the information you need from clicking on a website. Other sites such as Wikipedia have proved over the years that people will receive something that can be less influential if it can be free and easily.

None of these findings from CJR must be surprised. Language models have a difficult problem by understanding something they say because they are trying to create only something and have created autoComplete systems tease right. They are ad-libbing.

Another quoted from Howard standing when he saw the room for future development at Chatbots. “Today is the worst of the product that the product will happen,” referring to all investments in this area. However, this can be said about any technology throughout history. This is still irresponsible to let the information about this.