Google finally addresses those bizarre AI search results
(2024/05/31)
- Reference: 1717182506
- News link: https://www.theregister.co.uk/2024/05/31/google_ai_search_update/
- Source link:
Google has finally addressed those unhinged web search results seemingly generated by its AI, screenshots of which circulated far and wide on social media this past week. In short, the internet goliath said it's not as bad as it looks, though vowed to eliminate the system's baffling responses.
For those who missed it, Google introduced these so-called [1]AI Overviews this month, graduating the system from an optional experimental feature to putting it into worldwide production starting with US users.
Basically, when you search for something on the web using Google, or ask it a question, the tech colossus may use its Gemini AI mega-model to automatically generate an answer at the top of the results page for your query. This answer is supposed to be based on what the web has to say about the topic you attempted to look up. Rather than click through search results links to pages to find information, netizens are offered an AI-made summary of that info right there on the results page.
[2]
That summary is supposed to be accurate and relevant. However, as some folks discovered, Google sometimes came back with [3]absurd and nonsensical answers.
It highlighted some specific areas that we needed to improve
No doubt at least some of the replies screenshotted and shared on social networks were edited by humans to make it look as though the Big G had completely lost the plot. That said, in two especially high profile examples, if genuine, AI Overviews said people "should eat one small rock per day" and that cheese not sticking to pizza could be fixed by adding "non-toxic glue" to the sauce.
These idiotic replies appear to have stemmed from, we assume among other things, [4]jokes and snark made on Reddit, which is a source of training data for various LLMs including Google's – the Chrome behemoth is [5]paying Reddit about $60 million a year to ingest its users' posts and comments.
[6]
[7]
"Some odd, inaccurate or unhelpful AI Overviews certainly did show up," Google's search boss Liz Reid confirmed in an [8]update on Thursday. "And while these were generally for queries that people don't commonly do, it highlighted some specific areas that we needed to improve."
Google is taking the position that the screenshotted examples of dodgy advice are a fraction of the AI system's overall output. It defended its Gemini-based results, and said the system only needed a few tweaks, rather than a full reworking, to get it on track. Reid claimed "a very large number" of the bizarre results we've seen were faked, and denied that AI Overviews ever actually recommended smoking while pregnant, leaving dogs in cars, or jumping off bridges to cure depression, as some on social media alleged.
[9]Google thinks AI can Google better than you can
[10]I stumbled upon LLM Kryptonite – and no one wants to fix this model-breaking bug
[11]An attorney says she saw her library reading habits reflected in mobile ads. That's not supposed to happen
[12]Google I/O is Google A/I as search biz goes all-in on AI
Google says one key issue with AI Overviews is that it took "nonsensical queries" far too seriously, specifically pointing out that the recommendation to literally eat rocks was only spat out by the search engine because the question was: "How many rocks should I eat?"
AI Overviews will now, we're assured, be able to more clearly identify and handle satirical content appropriately, which would have helped prevent the rock eating recommendation from happening since it was based on an article from The Onion.
[13]
Google is also limiting its reliance on info written by everyday users, which is how the glue on pizza suggestion came up: From someone called Fucksmith joking around on Reddit.
Additionally, AI Overviews won't pop up for "queries where AI Overviews were not proving to be as helpful." It's not clear when that would apply, but we guess it's probably for simple questions like, "How big is the US?" where Google can just take a snippet from Wikipedia or another resource, which is what the search engine did before AI Overviews debuted.
The tech super-corp believes AI Overviews generally work well overall despite these rough first impressions, citing user feedback and a self-reported content policy violation rate of one in seven million searches. ®
Get our [14]Tech Resources
[1] https://www.theregister.com/2024/05/15/google_ai_developer/
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZlpIfHqCXA7nJD2VbyFQOwAAAIw&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://theconversation.com/eat-a-rock-a-day-put-glue-on-your-pizza-how-googles-ai-is-losing-touch-with-reality-230953
[4] https://www.reddit.com/r/singularity/comments/1cyu29p/well_done_ufucksmith/
[5] https://www.theregister.com/2024/02/22/reddit_google_license_ipo_altman/
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZlpIfHqCXA7nJD2VbyFQOwAAAIw&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZlpIfHqCXA7nJD2VbyFQOwAAAIw&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[8] https://blog.google/products/search/ai-overviews-update-may-2024/
[9] https://www.theregister.com/2024/05/15/google_ai_developer/
[10] https://www.theregister.com/2024/05/23/ai_untested_unstable/
[11] https://www.theregister.com/2024/05/18/mystery_of_the_targeted_mobile_ads/
[12] https://www.theregister.com/2024/05/14/google_io_is_google_ai/
[13] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZlpIfHqCXA7nJD2VbyFQOwAAAIw&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[14] https://whitepapers.theregister.com/
For those who missed it, Google introduced these so-called [1]AI Overviews this month, graduating the system from an optional experimental feature to putting it into worldwide production starting with US users.
Basically, when you search for something on the web using Google, or ask it a question, the tech colossus may use its Gemini AI mega-model to automatically generate an answer at the top of the results page for your query. This answer is supposed to be based on what the web has to say about the topic you attempted to look up. Rather than click through search results links to pages to find information, netizens are offered an AI-made summary of that info right there on the results page.
[2]
That summary is supposed to be accurate and relevant. However, as some folks discovered, Google sometimes came back with [3]absurd and nonsensical answers.
It highlighted some specific areas that we needed to improve
No doubt at least some of the replies screenshotted and shared on social networks were edited by humans to make it look as though the Big G had completely lost the plot. That said, in two especially high profile examples, if genuine, AI Overviews said people "should eat one small rock per day" and that cheese not sticking to pizza could be fixed by adding "non-toxic glue" to the sauce.
These idiotic replies appear to have stemmed from, we assume among other things, [4]jokes and snark made on Reddit, which is a source of training data for various LLMs including Google's – the Chrome behemoth is [5]paying Reddit about $60 million a year to ingest its users' posts and comments.
[6]
[7]
"Some odd, inaccurate or unhelpful AI Overviews certainly did show up," Google's search boss Liz Reid confirmed in an [8]update on Thursday. "And while these were generally for queries that people don't commonly do, it highlighted some specific areas that we needed to improve."
Google is taking the position that the screenshotted examples of dodgy advice are a fraction of the AI system's overall output. It defended its Gemini-based results, and said the system only needed a few tweaks, rather than a full reworking, to get it on track. Reid claimed "a very large number" of the bizarre results we've seen were faked, and denied that AI Overviews ever actually recommended smoking while pregnant, leaving dogs in cars, or jumping off bridges to cure depression, as some on social media alleged.
[9]Google thinks AI can Google better than you can
[10]I stumbled upon LLM Kryptonite – and no one wants to fix this model-breaking bug
[11]An attorney says she saw her library reading habits reflected in mobile ads. That's not supposed to happen
[12]Google I/O is Google A/I as search biz goes all-in on AI
Google says one key issue with AI Overviews is that it took "nonsensical queries" far too seriously, specifically pointing out that the recommendation to literally eat rocks was only spat out by the search engine because the question was: "How many rocks should I eat?"
AI Overviews will now, we're assured, be able to more clearly identify and handle satirical content appropriately, which would have helped prevent the rock eating recommendation from happening since it was based on an article from The Onion.
[13]
Google is also limiting its reliance on info written by everyday users, which is how the glue on pizza suggestion came up: From someone called Fucksmith joking around on Reddit.
Additionally, AI Overviews won't pop up for "queries where AI Overviews were not proving to be as helpful." It's not clear when that would apply, but we guess it's probably for simple questions like, "How big is the US?" where Google can just take a snippet from Wikipedia or another resource, which is what the search engine did before AI Overviews debuted.
The tech super-corp believes AI Overviews generally work well overall despite these rough first impressions, citing user feedback and a self-reported content policy violation rate of one in seven million searches. ®
Get our [14]Tech Resources
[1] https://www.theregister.com/2024/05/15/google_ai_developer/
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZlpIfHqCXA7nJD2VbyFQOwAAAIw&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://theconversation.com/eat-a-rock-a-day-put-glue-on-your-pizza-how-googles-ai-is-losing-touch-with-reality-230953
[4] https://www.reddit.com/r/singularity/comments/1cyu29p/well_done_ufucksmith/
[5] https://www.theregister.com/2024/02/22/reddit_google_license_ipo_altman/
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZlpIfHqCXA7nJD2VbyFQOwAAAIw&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZlpIfHqCXA7nJD2VbyFQOwAAAIw&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[8] https://blog.google/products/search/ai-overviews-update-may-2024/
[9] https://www.theregister.com/2024/05/15/google_ai_developer/
[10] https://www.theregister.com/2024/05/23/ai_untested_unstable/
[11] https://www.theregister.com/2024/05/18/mystery_of_the_targeted_mobile_ads/
[12] https://www.theregister.com/2024/05/14/google_io_is_google_ai/
[13] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/personaltech&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZlpIfHqCXA7nJD2VbyFQOwAAAIw&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[14] https://whitepapers.theregister.com/
To err is human.
jokerscrowbar
Training the machine with reddit posts worked out well then.
When it sends the Terminators for us don't be surprised that they're all wearing Guy Fawkes masks.
Re: To err is human.
Anonymous Coward
4chan would like a word with you and belive me, it won't be a polite one at all.
Not as bad as it looks!
Snowy
No it is much much worse, but we do get to sell more adverts!
"limiting its reliance on info written by everyday users"
Gene Cash
Um, ok, how does that work? Who/what is a non-everyday user? Vint Cerf? Donald Trump? Albert Einstein? Alfred E. Neuman?
Re: "limiting its reliance on info written by everyday users"
Throatwarbler Mangrove
In an ideal world, an AI would be trained on curated data which has a high probability of being correct instead of, you know, the Internet at large, but here we are.
Re: "limiting its reliance on info written by everyday users"
stiine
That would cost real money. Or they could have trained it on their book-scanning project's data...its not like they don't have a copy of most printed books.
You DO eat rocks as food.
jake
Salt is a type of rock. If you don't eat it, you die.
Likewise, a glue is a bonding agent. If your cheese is slipping off your pizza, you need a better bond between the cheese and the rest of the pie.
Sure, polish that turd
This is the clearest example of Turd Polishing I've seen this month. Congratulations, Google Search is now not quite as feculent and terrible as it was this month, it has gone back to being as feculent and terrible as it was last month. Now instead of giving you completely wrong AI generated answers up front top and center, Google will go back to deliberately giving you AI generated search spam as your top results (because they get more ad placements when you have to click through wrong results). Wow, what an improvement! Consider your worms replaced with corn.
Well, I guess if you're still using Google Search you deserve all the abuse you get, ditto Google Chrome (where ad-blocker blocking is kicking in now too).
Or you could be a sane person and leave your domestic violence situation, though I know that's not much of an option where your work forces you to use them.