What is wrong with Google AI Overviews?


Deployed in the United States since May 14, 2024, the “AI Overviews” function summarizes web pages using Gemini artificial intelligence. She is strongly criticized by her first users, since some of her answers are completely false.

A revolution is coming to the world of search engines and Google does not want to take the risk of being overtaken. Faced with the OpenAI threat and the emergence of competitors who summarize web pages using AI, Google has been actively working for more than a year on its own version of AI summaries, until now only accessible by registration.

As of May 14, 2024, the AI ​​Overviews feature is available and deployed by default to all Google users in the United States. Billions of searches have already been summarized by the Gemini AI, which offers a written answer at the top of the results. The problem is that some of his answers frankly leave something to be desired, or even… are downright delusional.

For further

Sundar Pichai // Source: Nicolas Lellouche

Social networks massively mock AI Overviews

After the “black Nazis” controversy, which saw Gemini’s image generator modify the story to add inclusiveness, Google is in turmoil again. Many people, anonymous or self-confessed anti-Google, are having fun sharing screenshots of the AI ​​Overviews function in full drift. Some results make you smile, others are more worrying (and some are fakeintended to parody the errors of Google).

In this example, AI Overviews would recommend smoking while pregnant.In this example, AI Overviews would recommend smoking while pregnant.
In this example, AI Overviews would recommend smoking while pregnant. // Source: X
A bit like Grok, Google's AI can use social networks to respond. Here she would recommend suicide.A bit like Grok, Google's AI can use social networks to respond. Here she would recommend suicide.
A bit like Grok, Google’s AI can use social networks to respond. Here she would recommend suicide. // Source: X

Very often, AI Overviews seems to dig into answers on Reddit, instead of choosing more reliable sources. Google also doesn’t show the same results to everyone. Gemini can sometimes give bad health advice and sometimes confuse a joke with reality. Advising a depressed person to throw themselves off a bridge is not without a doubt not a good idea…

On Twitter, screenshots of AI Overviews’ failed responses were widely shared. Some are unverifiable, others are authentic. In any case, this is a bad spotlight for Google, which managed to associate its new innovation with a failed launch.

In a sarcastic tone, AI Overviews states that everything found on the Internet is true.In a sarcastic tone, AI Overviews states that everything found on the Internet is true.
In a sarcastic tone, AI Overviews states that everything found on the Internet is true. // Source: X
Out of 42 American presidents, 17 are white according to Google. In reality, Joe Biden is the 46th president… and the 45th white.Out of 42 American presidents, 17 are white according to Google. In reality, Joe Biden is the 46th president… and the 45th white.
Out of 42 American presidents, 17 are white according to Google. In reality, Joe Biden is the 46th president… and the 45th white. // Source: X

Should Google kill AI Overviews? Be careful not to get carried away

Aware of the problem, Google is acting to “cure” its algorithm by manually correcting each problematic answer (its quick response is to disable the function on questions that pose a problem). The web giant also offers a “Web” tab that hides AI, images and Maps, but it cannot be activated by default.

Is Google AI Overviews doomed to an early demise, as some seem to be looking forward to? It is important to qualify the controversy, remembering that billions of Google searches take place daily and that, a priori, the service has not been deserted since the appearance of AI Overviews. In most cases, IA0 summaries are probably not a problem.

Numerama, which was able to test AI Overviews during its last trip to the United States, experienced a relatively promising service. Bank card fees, ideas for things to do, general knowledge… AI Overviews is useful and saves time, while encouraging you to go further when it is too evasive. Google’s goal is likely to iron out snags, rather than immediately doom an entire system.

AI Overviews: some examples of research summarized by AI. The pages appear very far below. // Source: Numerama CapturesAI Overviews: some examples of research summarized by AI. The pages appear very far below. // Source: Numerama Captures
AI Overviews: some examples of research summarized by AI. The pages appear very far below. // Source: Numerama Captures

It is also important to note that Google, considered “too wokist”, is also the target of a very conservative part of Silicon Valley, which regularly protests against the practices of the tech giants. A number of anti-Google tweets come from these accounts.

The impact on the web, the other problem of AI Overviews

Either way, the initial user response to AI Overviews doesn’t look good. Many people regret that the option is activated by default, which forces them to navigate to the bottom of the page to display correct and verified results, such as those from recognized media for example.

Another unknown: what will happen to the traditional web if Google only selects sites arbitrarily for its answers, while hiding the results under a layer of blocks. The company says the impact is limited, but initial measurements reveal a 10% drop in audiences. This could become a problem for Google if certain sites disappear and its AI can no longer feed itself.

In the coming weeks, the web giant will certainly refine its system to prevent it from saying anything. The launch of AI Overviews nevertheless reminds us of an essential element: you should never blindly believe an AI.

If you liked this article, you will like the following: don’t miss them by subscribing to Numerama on Google News.



Source link -100