May 31, 2024 19:00:00

Google explains why the content of the 'AI Overview' search results becomes a mess

Google has acknowledged that its

AI Overview feature, which has been widely rolled out since May 2024 and uses AI to summarize search results and display an overview, had given strange answers such as 'using glue to stick cheese to pizza,' and has explained why it came to give such answers.

What happened with AI Overviews and next steps
https://blog.google/products/search/ai-overviews-update-may-2024/

Google explains AI Overviews' viral mistakes, defends accuracy

https://9to5google.com/2024/05/30/google-ai-overviews-accuracy/

Google is putting more restrictions on AI Overviews after it told people to put glue on pizza
https://www.engadget.com/google-is-putting-more-restrictions-on-ai-overviews-after-it-told-people-to-put-glue-on-pizza-011316780.html

Google explains how it is improving its AI Overviews
https://searchengineland.com/google-explains-how-it-is-improving-its-ai-overviews-442800

AI Overview is a feature that displays an AI-generated 'search result summary' at the top of the search results. For example, in response to a question such as 'How do I clean a fabric sofa?', the app will provide an overview of the search results in natural language, such as 'If you use a vacuum cleaner, do this, but if you want to focus on cleaning a specific area, do this.'

However, immediately after its release, AI Overview pointed out that the AI was giving clearly incorrect answers, such as 'suggesting to use glue to stick cheese to pizza,' 'claiming that dogs have played in the NBA,' and 'answering that 1919 was 20 years ago.'

It is revealed that Google's AI summary function gives strange answers such as 'glue is used to stick cheese to pizza' - GIGAZINE

In response, Google stated that it 'takes swift action to disable AI Overview answers for specific search terms where we deem appropriate under our content policies,' acknowledging that it does disable AI Overview on an individual basis.

Reports that Google is manually correcting erroneous AI searches, and a privately developed service that circumvents AI searches has also appeared - GIGAZINE

Furthermore, on May 30, 2024, Google published an official blog post stating, 'In light of the attention that AI Overview has received, we would like to explain what happened and the steps we have taken.'

In the official blog, Google explained, 'AI Overview does not simply generate output based on training data, but is integrated with the core web ranking system along with a customized language model to identify and output relevant, high-quality results from the index. Therefore, it emphasizes that ' hallucinations do not occur like other large-scale language model products.'

This time, Google analyzed the reason why strange answers such as 'using glue to stick cheese to pizza' were born 'because they were given meaningless queries and satirical content.' According to Google, even if you ask the question 'How many stones should I eat?' in a Google search, there is very little web content that can properly answer this question. These phenomena, where the amount of high-quality content on a topic is limited, are called 'data voids' or ' information gaps .'

In the case of the question 'How many rocks should you eat?', AI Overview relied on very little information - a website with satirical content on the topic and a blog post from a geological software provider - to create the overview, which resulted in this strange answer.

In addition, Google points out that 'we have also observed the launch of nonsensical websites that appear to be intended to generate false answers in AI Overview.'

Google's investigation also revealed that AI Overview picks up sarcastic content and trolls on discussion forums to generate answers. As a result, Google has worked to improve its AI Overview algorithm and establish a process to exclude answers that do not comply with its policies from the generated overview.

Specifically, the company has improved the detection mechanism for meaningless queries that should not display AI overviews, restricted the inclusion of satirical or humorous content in answers, and added trigger restrictions for queries that have proven to be less useful for AI overviews.

Google also revealed that it monitors user feedback and external reports and takes action against a small number of AI overviews that violate its content policies.