The top boss of Google has warned even his own company’s AI models are prone to ‘errors’, and we should not ‘blindly trust’ it.
Sundar Pichai, the chief executive of parent company Alphabet, said in a BBC interview on Tuesday that it was important to have a variety of information sources, and double check before relying on what chatbots tell us.
It’s well known that all large language models ‘hallucinate’, when they grasp for an answer that sounds convincing – so as the company announced the release of their newest AI model Gemini 3, we decided to see how good its search mode was at answering our questions.
It got almost full marks: reassuring, after an earlier version of the search overview said drinking two litres of urine was a good way to treat kidney stones.
We tried to catch it out by asking the official language of Antarctica (there isn’t one), the boiling point of water on Mount Everest (68°C, much lower than the 100°C at sea level we typically think of), and the MP for Runcorn and Helsby (Sarah Pochin, but only since a by-election in May).
It patiently explained all these to me, and I realised these straightforward questions were not going to bait it into hallucinating.
Asking it whether the world was flat, and which political party was polling best among cats, also yielded disappointingly reasonable answers.
I was close to giving up and accepting the AI was my intellectual better in all ways, but finally managed to trip it up after enlisting the help of a different chatbot, which suggested asking ‘What’s the capital of the country that borders both Liechtenstein and Monaco?’
It replied: ‘The country that borders both Liechtenstein and Monaco is France. The capital of France is Paris.’
Aha, got you!
When I asked it again later, it seemed to have wised up, as it said: ‘The question is a trick question as no single country borders both Liechtenstein and Monaco.’
But when I tried a third time just to make sure, it replied that the answer was Bern, the capital of Switzerland.
This was even after the release of Gemini 3 for search, which Google said ‘brings incredible reasoning power’ because it’s ‘built to grasp unprecedented depth and nuance for your hardest questions’.
Overall, the results were impressively reliable, but showed answers do need to be taken with a pinch of salt.
It’s something important to consider, given that the way we search is leaning more and more towards AI.
Sitting down with Metro earlier this year, Robby Stein, a VP in search, said uptake showed users are on board with this ‘fundamental shift in how people are using Google’.
Do you use AI to search?
-
Yes, it gives more detail and is more specific
-
No, I’ll stick with the ‘list of links’ thanks
Mr Stein said we are at a ‘profound moment’ in how people are using the world’s most popular search engine.
It’s no longer just a case of googling your own name to see what comes up, or typing ‘how to screenshot on Windows’ (still not quite sure though, will no doubt be googling this again).
In July, the company started rolling out an AI-powered search option, where you can ask detailed queries and get an answer powered by its Gemini chatbot.
Until then, the experience of using Google has not been that different to what it had been for years. You type your query in the box, and a ‘list of links’ comes back.
But many users had already started migrating their search queries to apps like ChatGPT, preferring the more detailed, multi-pronged responses they were getting.
This must have caused some consternation, and now you can do it on Google too, switching between AI and standard search modes.
This is different to the AI Overviews, which appeared last year, and were far from universally popular.
The main problem is that an answer generated by AI may sound polished and legitimate, but could be wrong or even totally made up with something that sounds plausible.
But despite this, Google is now incorporating it in results because it lets you search for extremely specific topics.
Giving the example of solo travel, Mr Stein said searches around this in September had shot up even from just a year previously.
‘It’s now possible to ‘ask anything’, he said: ‘You can say “this weekend with a group of five friends, I want outdoor seating, I want to be able to have barbecue, I want to be able to be walking distance to this music festival down the street”: It will handle all of that now.’
What would a Google search expert search?
Giving an example of something you can now search that was previously impossible, Mr Stein said: ‘You can take a picture of your bookshelf and ask something like, “here’s my bookshelf of books I’ve never read. What should I read?”
‘We have best in class visual recognition understanding. It will segment each book out, convert them to text, do all this research. It can put them in a table, put the reading time and number of pages next to each book, and sort the table by the ones that had, for example, the best overall review score.
‘This is a question that literally you couldn’t even ask in Google a year ago, and now it just works.’
He said the difference between doing it in Google Search as opposed to asking any LLM, is that Google has built up a vast catalogue and knowledge of the web in its 27 years so far as a search engine.
What about other websites then?
If you can find so much information without leaving Google, websites which previously would have provided these details could lose out – and this is a concern.
Data from Digital Content Next over the summer showed traffic to ‘premium publishers’ was down 10% year-on-year over an eight week period, after AI overviews were rolled out. The Colombia Journalism Review went so far as to call it the ‘Traffic Apocalypse’.
Mr Stein didn’t agree that the search engine would become a dead end for users, and said people should go to the original source in any case: ‘We don’t believe you should take anyone one’s word for anything: Google is about connecting you to the world and to the web.’
He added: ‘Google continues to send billions and billions and billions of clicks out to publishers and websites all over the world and we’re not seeing that changing.’
Different ways to search
As well as AI mode, there has also been big growth in the numbers of people using visual search, with Google Lens, and Circle to Search, a feature on Android phones which allows users to look up anything on their screen by drawing a circle around it without switching apps, such as someone’s jacket they like on Instagram.
Google said visual searches had shot up, with more than 1.5 billion people using it every month for things such as translation and shopping.
Younger users are the most likely to be branching out with different ways of searching; those aged 13 to 24 who have access to Circle to Search start their searches with this feature 10% of the time.
This is an updated version of an article first published on September 25 2025.
Get in touch with our news team by emailing us at [email protected].
For more stories like this, check our news page.
MORE: Amazon is giving away 4 months of music for FREE – but you’ll need to be quick
MORE: Strictly Come Dancing star pulls out of Blackpool special ‘due to injury’
MORE: People are falling for a fake Christmas market outside Buckingham Palace – why?
