Forcing AI for everyone: Google has been rolling out AI Overviews to US users over the last several days. While the company claims that the AI summaries that appear at the top of search results are mostly correct and fact-based, an alarming number of users have encountered so-called hallucinations – when an LLM states a falsehood as fact. Users are far from impressed.
In my early testing of Google’s experimental feature, I found the blurbs more obnoxious than helpful. They appear at the top of the results page, so I must scroll down to get to the material I want. They are frequently incorrect in the finer details and often plagiarize an article word for word.
These annoyances prompted me to write last week’s article explaining several ways to bypass the intrusive feature now that Google is shoving it down our throats with no off switch.
And now that AI Overviews has had a few days to percolate in the public, users are finding many examples where the feature simply fails.
Social media is flooded with funny and obvious examples of Google’s AI trying too hard. Keep in mind that people tend to shout when things go wrong and remain silent when they work as advertised.
“The examples we’ve seen are generally very uncommon queries and aren’t representative of most people’s experiences,” a Google spokesperson told Ars Technica. “The vast majority of AI Overviews provide high quality information, with links to dig deeper on the web.”
While it may be true that most people have good summaries, how many bad ones are allowed before they are considered untrustworthy? In an era where everyone is screaming about misinformation, including Google, it would seem that the company would care more about the bad examples than patting itself on the back over the good ones – especially when its Overviews are telling people that running with scissors is good cardio.
AI entrepreneur Kyle Balmer highlights some funnier examples in a quick X video (below).
Google AI overview is a nightmare. Looks like they rushed it out the door. Now the internet is having a field day. Heres are some of the best examples https://t.co/ie2whhQdPi
– Kyle Balmer (@iamkylebalmer) May 24, 2024
It is important to note that some of these responses are intentionally adversarial. For example, in this one posted by Ars Technica the word “fluid” has no business being in the search other than to reference the old troll/joke, “you need to change your blinker fluid.”
The joke has existed since I was in high school shop class, but in its attempt to provide an answer that encompasses all of the search terms, Google’s AI picked up the idea from a troll on the Good Sam Community Forum.
How about listing some actresses who are in their 50s?
google’s AI is so smart, it’s honestly scary pic.twitter.com/7aa68Vo3dw
– Joe Kwaczala (@joekjoek) May 19, 2024
While .250 is an okay batting average, one out of four does not make an accurate list. Also, I bet Elon Musk would be surprised to find out that he graduated from the University of California, Berkley. According to Encyclopedia Britannica, he actually received two degrees from the University of Pennsylvania. The closest he got to Berkley was two days at Stanford before dropping out.
The same kind of error is unfortunately trivial to produce, eg, tech ceos who went to Berkeley. pic.twitter.com/mDVXT714C7
– MMitchell (@mmitchell_ai) May 22, 2024
Blatantly obvious errors or suggestions, like mixing glue with your pizza sauce to keep your cheese from falling off, will not likely cause anybody harm. However, if you need serious and accurate answers, even one wrong summary is enough to make this feature untrustworthy. And if you can’t trust it and must fact-check it by looking at the regular search results, then why is it above everything else saying, “Pay attention to me?”
Part of the problem is what AI Overviews considers a trustworthy source. While Reddit can be an excellent place for a human to find answers to a question, it’s not so good for an AI that can’t distinguish between fact, fan fiction, and satire. So when it sees someone insensitively and glibly saying that “jumping off the Golden Gate Bridge” can cure someone of their depression, the AI can’t understand that the person was trolling.
Another part of the problem is that Google is rushing out Overviews while in panic mode to compete with OpenAI. There are better ways to do that than by sullying its reputation as the leader in search engines by forcing users to wade through nonsense they didn’t ask for. At the very least, it should be an optional feature, if not an entirely separate product.
Enthusiasts, including Google’s PR team, say, “It’s only going to get better with time.”
That may be, but I have used (read: tolerated) the feature since January, when it was still optional, and have seen little change in the quality of its output. So, jumping on the bandwagon doesn’t cut it for me. Google is too widely used and trusted for that.