Google’s AI Overview is already becoming a little unhinged

Remember how excited everyone was when ChatGPT and its compadres arrived, in awe at their genius? But then, slowly, we began to notice that every so often they went a bit rogue? A bit unhinged?

Bing’s chatbot, powered by ChatGPT, started gaslighting users and complaining about them, in one case writing the person was ‘wasting my time and yours’ going on to call them ‘unreasonable and stubborn’.

Then in February, ChatGPT itself appeared to have a mini meltdown, mixing up languages and making up words. In one response, it wrote: ‘Let me encycloplease me si there’s more wonderenda tu articulation’s hungry for!’

Well, it seems Google’s AI, named Gemini, is also prone to the odd slip-up.

During the company’s biggest developer conference, Google I/O, CEO Sundar Pichai announced they would begin rolling out AI Overview, which tops search results with a written response generated by artificial intelligence. The aim is to save users time, so instead of clicking through various links, the answer to their question is right in front of them.

Unfortunately, there have been a few teething problems. 

As we’ve seen, AI can create some, shall we say, imaginative responses, and isn’t always quite factually accurate. Being trained from data on the internet, including that which is unsourced, a joke or just plain wrong, chatbots are susceptible to making mistakes – or just making stuff up entirely.

So far, the new technology has only been rolled out to non-beata users in the US, but has already thrown up some interesting results.

One user asked Google to help solve the issue of cheese slipping off a pizza.

Google suggested they try glue.

‘You can also add about 1/8 cup of non-toxic glue to the sauce to give it more tackiness,’ it wrote.

Incredibly, this wasn’t one instance of AI making something up. Instead, it had pulled the information from an 11-year-old Reddit post, in which the writer was obviously joking. Sadly, AI does not always get sarcasm. Yet.

Another query asked ‘how to pass kidney stones quickly’, to which Google recommended the unfortunate user drink urine.

‘Drinking plenty of fluids, such as water, ginger ale, lemon-lime soda or fruit juice can help pass kidney stones more quickly,’ it replied. ‘You should aim to drink at least 2 quarts (2 liters [sic]) of urine every 24 hours, and your urine should be light in colour’.

Firstly, if you’re drinking urine, your own urine won’t be light in colour.

Secondly, do not drink urine. 

In this case, the answer was generated while still in testing, having been posted on X, formerly Twitter, on May 5. Nevertheless, with only a week until the full US rollout, it didn’t bode well – and we can expect more where that came from.

With AI Overview not yet available in the UK, Metro.co.uk has not been able to independently reproduce any of the results, but other examples from legitimate sources include one response stating the Barack Obama was the first Muslim US president, and another that there have been 42 presidents. There have been 45, and 46 presidencies – Grover Cleveland served non-consecutive terms.

Realistically, we are less than two years into the widespread public use of large language model chatbots. Remember mobile phones two years after they became widespread? They’re more than a little different now, and in the same way, AI will continue to improve and refine – probably much faster than phones ever did.

For now though, it basically means all learning and testing is near enough happening live. For many, it’s fun to watch – and for those who worry their jobs may be at risk from AI, such inane guff is comforting.

However, there will be some who do not realise it is inane guff. Google must quickly get a grip on its AI Overview before someone follows a piece of poorly-written advice and suffers for it.

A Google spokesperson said: ‘The vast majority of AI Overviews provide high quality information, with links to dig deeper on the web. Many of the examples we’ve seen have been uncommon queries, and we’ve also seen examples that were doctored or that we couldn’t reproduce.

‘We conducted extensive testing before launching this new experience, and as with other features we’ve launched in Search, we appreciate the feedback.

‘We’re taking swift action where appropriate under our content policies, and using these examples to develop broader improvements to our systems, some of which have already started to roll out.’

For UK users, the wait to try out AI Overview continues, but you can be sure to expect some, erm, interesting results when it arrives.