Source: Finance Association
Author: Niu Zhanlin
① Liz Reed, head of Google's search business, admits that some strange, inaccurate, or unhelpful AI search results did appear;
② Experts warn that this kind of information retrieval should be Google's core business, and it cannot be rushed to hand it over to the AI model.
On Friday local time, tech giant Google announced that it had made “more than a dozen technical improvements and updates” to its AI system after discovering a series of misinformation in the “AI Overviews” (AI Overviews) function.
At the annual I/O developer conference in mid-May, the tech giant unveiled a number of AI-related products and services. One of the contents was to integrate AI into its world-leading search engine, the AI overview function, which was first launched in the US market. Simply put, when a user searches, the first thing that appears in front of them is no longer a web link, but a summary compiled by AI.
Soon after, however, social media users began sharing screenshots of a series of outrageous answers given by the AI Overview. Google defended this by saying that the AI overview function is generally accurate and has been extensively tested beforehand.
However, by Friday, Liz Reed, the head of Google's search business, admitted that some strange, inaccurate, or unhelpful AI search results did appear.
While many of these search results are stupid, there are also some that are dangerous or harmful lies. For example, applying glue to pizza to help the cheese stick, you should eat at least one small rock every day, etc. In addition, some netizens fueled the situation, and some false screenshots were also widely shared on social media, causing the incident to become more and more popular.
Reed said that more “trigger limits” have been added to the AI overview to improve the accuracy and reliability of answers to certain queries (such as health-related questions).
However, after Google announced the update and tested it, it was discovered that there were still quite a few issues. Google's AI Overview is designed to provide people with authoritative answers to the information they're looking for without having to click on a ranking list of website links, but it seems that the feature is still very flawed.
Some AI experts have also long warned Google not to hand over their search results to AI-generated answers, which may perpetuate prejudice and misinformation, and endanger those seeking help in emergency situations. The working principle of big language models is to predict which words will best answer people's questions based on the data they have been trained on, which causes them to often be fabricated, or AI illusions.
Reed said in this regard that AI overviews are generally not “illusory,” nor are they made up like other big language model products, because they are more closely integrated with Google's traditional search engine and only show content supported by the most authoritative or relevant web pages.
“When AI overviews go wrong, it's usually due to other reasons: like misunderstanding the query, misunderstanding the nuances of the language on the web, or not having much useful information,” she wrote.
However, computer scientist Chirag Shah, a professor at the University of Washington, pointed out that this kind of information retrieval should be Google's core business, and it cannot be rushed to hand it over to the AI model. “Because even if AI doesn't make up a mess, it will pass on misinformation to users, Google may have quite bad results if it does this.”
Editor/Jeffy