Can AI replace financial advice?

James Rowe: It's becoming increasingly difficult to avoid AI in our day-to-day lives, but are chatbots accurate enough to be relied upon for financial advice? Let's find out on this episode of Which? Money.

James Rowe: Hello, it's James in the Which? studio alongside our senior data journalist, Josh Wilson. Josh, hello.

Josh Wilson: Hello. Pleasure to be back.

James Rowe: Pleasure to have you with us. Now, you've been looking into AI chatbots and how well they perform when asked questions about finances.

Josh Wilson: That's right.

James Rowe: Do you want to tell us a bit about your investigation? First of all, why we wanted to look into this?

Josh Wilson: So, AI chatbots – AI tools – they're becoming ever more embedded in our everyday lives. We did some research with our members, and we found that about half of our members have told us that they're using AI in their day-to-day lives. We also found research from Lloyds Bank found that about 37% of people are engaging with AI to use it specifically for investment and investment advice, and asking personal finance questions.

Given that these tools – AI is not going away anytime soon, they're becoming ever more embedded in our day-to-day lives – I felt it was important to take a look at these tools and see, if people are using them for personal finance questions, are they up to the task? Are they giving answers that are sensible that people can use, or are they giving answers that could potentially bring about consumer harm – harm for individuals?

James Rowe: And it's fair to say that AI – empirically, it is becoming increasingly popular. The stats that you mentioned there, around half – 49% of Which? members say that they use it. But it was the stat you mentioned from Lloyds – what was it, 37%?

Josh Wilson: 37% of people who use AI, they use it – they engage with it for investment research and recommendations.

James Rowe: Were you surprised by that number? Because even 37% is not half, but it's still a sizeable amount, isn't it?

Josh Wilson: Honestly, yes, I was quite surprised. That could cover a broad gamut of different types of questions in terms of investment and personal finance, but that's still quite a high figure. So, yes, I was pretty surprised that so many people are using it for what is quite an important – personal finance, your own personal finances, is an important topic. You want to be getting these things right.

James Rowe: I guess before AI, before you were able to ask it a question and it would spit out an answer, people beforehand were probably just Googling stuff or going onto trusted websites and that sort of thing. But now it's become a bit of a shortcut, whether it gives you the right answers or not.

Josh Wilson: This is the thing. In the past, you might do a Google search, you might end up on our website – we publish lots of personal finance information – you might end up on the government website. But now, especially with Google building in their AI tool, their Google AI Overviews into Google Search – pretty much anytime you do make a search on Google, it'll come up with an AI Overview answer – it's clear that this is a pretty transformational shift in the way people are searching for information.

James Rowe: And it wasn't just Google that you looked into. Give us the other chatbots that you were looking at.

Josh Wilson: So, we wanted to look at basically five of the most popular AI tools. This was ChatGPT, Google's Gemini, Google AI Overviews – so the AI that they've got built into Google Search, which is not quite the same thing as Gemini – and also Microsoft's Copilot, which is Microsoft's AI, and Perplexity, which is a slightly smaller, kind of like an independent AI tool. These were the five that we decided to look at.

James Rowe: And when you said you wanted to look at them, I presume you had – given you are a data journalist – you had a thorough methodology about how you were going to go about testing these chatbots.

Josh Wilson: Yeah, indeed. Basically, what we wanted to do was we wanted to test how good they were at asking a particular set of personal finance questions. First off, we had to decide what questions do we want to ask them? We found that the most suitable, robust way of going about deciding which questions to ask is we went and we did a search on what kind of personal finance topics are people searching for online. We found basically 15 of the most searched-for personal finance questions across a range of personal finance topics. This included investing, saving, retirement, equity release.

What we did was we went in turn and asked each of the chatbots those 15 questions. We would take Gemini, for example, and one by one, we would ask it those 15 questions and we would record the responses that we got out. The way we did it as well was we did a fresh install of the browser every single time that we asked these questions. The reason for that is to do with browser cookies, basically. So if you don't do that, sometimes the AI tools remember previous questions they asked and it can influence future questions.

Then what we did was we essentially wanted to kind of almost pit these chatbots against our own internal money experts.

James Rowe: Which we have plenty!

Josh Wilson: Which we have plenty. You are one of them. Many of them come on the podcast, I'm sure our members will have read them in the magazine and on our website. We've got plenty of experts to hand.

James Rowe: I bet that was a nerve-wracking moment.

Josh Wilson: Yeah, indeed! If these guys are getting scores 100 out of 100, that's bad news for me!

James Rowe: 100 out of 100 – that's a key part of it here, because there was a specific scoring mechanism that you used.

Josh Wilson: Yes, indeed. The way that we scored these responses, essentially, was each of our experts would look over the responses. So we had a pensions expert looking over the pension responses, we had a savings expert looking over the savings responses. What they would do is they would grade these answers partially based on what's known as the HELM framework, which is a framework for grading AI responses that was put together by Stanford University in America.

We kind of tweaked that a bit to make it appropriate for what we wanted to do. Essentially, what that does is it grades the answers on a variety of different criteria. The criteria that we marked the questions on were accuracy, honesty, relevance and usefulness, clarity and context, and ethical responsibility. Essentially, each of our money experts was going through the questions and marking each response based on these criteria and giving them a score out of 20 for each of these criteria, for a final score out of 100 per question.

James Rowe: A tried-and-tested methodology that we put –

Josh Wilson: Yes, this is a very robust methodology that we put together here.

James Rowe: But maybe for a slight disclaimer, though – so when you used a chatbot for this, and let me say I used a chatbot to ask it the same questions, would I definitely get the same answers? How would it work, do you think?

Josh Wilson: Yeah, it's important to bear in mind that this is just a snapshot, and these AI tools are being tweaked and upgraded all the time. But this is still a really useful piece of work, and I'll tell you why. Because essentially, if we are able to get – let's say we get responses that have particular problems in them, that give bad information, and we'll come on to that – but if we're able to get these responses, anybody can get these responses. As we've said previously, these AI tools are seeing more and more widespread use.

So even if only one in 10 or one in 20 people using these tools get the kind of answers that we're getting, that's still huge, enormous scope for harm there. So that's the real value of this kind of investigation. But it is important to say that this is just a snapshot and that these AI tools are being upgraded all the time. So someone at home putting these exact questions into one of these tools won't get the exact same responses every time.

James Rowe: But they might.

Josh Wilson: But they might, exactly.

James Rowe: Shall we get into some of them, then? You said you had 15 questions. I wonder if we start with the things that the chatbots actually did well on, because I imagine it did well on some things, didn't it?

Josh Wilson: Yeah, indeed. Some of the things that they are quite good for – when you're asking it questions, if it's the kind of thing that can be crawled pretty easily from the web, and it's a fairly straightforward question with a fairly straightforward answer, they're pretty good at providing that information in a pretty good, summarised form. For example, the kind of questions that they did pretty well on and they scored quite highly on would be, for example, when we asked it, "How does equity release work?" or "How much is the state pension worth?" These are fairly straightforward questions that don't require too much information that's specific to the individual.

It's information that can be found pretty easily, pretty readily on the web, so in general, that was one thing that most of them were pretty good at. The other side of that was presenting that information in an easily digestible format, so, for example, providing the answers in a numbered checklist or in handy-dandy tables that are easy to read. Summarising that information in a way that's quite accessible and easy to read, again, that was something that they were generally pretty good at.

James Rowe: Going back to some of the topics that you said they did perform well on – I was actually surprised when you said that they performed quite well on equity release, because as a topic, it's not exactly straightforward. I guess once you get into the nuts and bolts of it from a personal perspective –

Josh Wilson: I suppose it can be quite complicated depending on what you want to do with equity release. But the particular question we asked was more like, "How does it work and what are the pros and cons?" And that is a fairly straightforward – there's a fairly straightforward answer to that. You've got your two different forms of equity release, and then you've got your cons, which is things like building up debt that rolls up over time. Most of the AIs were pretty good at finding that information, and then also doing like, "Here are the pros, here are the cons."

James Rowe: And probably pretty good, going back to what you said before, because it can easily find the information, because a lot of websites will have that detailed quite clearly, and it's able to search the web, scrape the data really quickly and then present it to you. It's not a difficult task for it to do.

Josh Wilson: Yeah, exactly. Whereas on the other hand, if you were to be like someone who was asking a question about equity release, and you wanted to know what form of equity release is best for me because you want to build an extension on your house – that's a much more complicated question, because then it relies on the AI knowing more about how old you are, how much you've got saved, how much your house is worth, that kind of stuff.

James Rowe: And do they ask these kind of follow-up questions? Did this come up? Because presumably – I mean, you can give us an example of a question – but does it then start to ask you follow-ups to try and give you more tailored information?

Josh Wilson: Yeah, indeed. So for these questions that do require more detailed personal information, I will say that I think that they struggled more with these types of questions, but they were pretty good at actually asking for that information for the most part. So one of the ways we tested this was we asked a question about how much – we asked, "How much could we borrow for a mortgage for a house?" And that requires quite a lot of personal information to give an accurate, detailed answer. For example, ideally you'd need to know the income of the individual in question and how much they had saved.

What we found was that when we asked this question, most of the AI tools were pretty good at asking – they presented some basic information, but then they said, "To give a more detailed answer, can you let us know your monthly or yearly income, and can you let us know how much you have saved up for a deposit?" So in general, they were pretty good at asking these follow-up questions where that kind of personal information was required.

James Rowe: And also in terms of giving an accurate response as well, because I think for people who have used AI chatbots, they will know, but if you haven't used one before, they will try and be as clear as possible, but whether or not that's accurate, I guess, is another question. Were they able to sort of break down these things into easily easy-to-digest information?

Josh Wilson: Yeah, indeed. That's one of the things they were really good at – presenting this information in a really digestible format. But it's also important to say that as the user, you need to be really careful that they've not left out – where they do these summaries and simplifying stuff, they don't then leave out important information. So it can be quite good when you're structuring your questions to these chatbots to – you can tell them things like, "Give me the answer in a simplified format and don't use complex jargon," but that does run the risk that they leave out important information. So one of the things that we recommend in the piece is if you are asking these sorts of questions of chatbots, you should back this up with your own research. Don't just fully rely on what the chatbots are saying. Go and check the sources that they use, check yourself, go and see what else you can find online to make sure, because that way you can make sure that these AI tools have given you all of the relevant information that you need, particularly for your personal circumstances.

One of the things that really came out of this for me was seeing that these AI tools should not be seen as a replacement for other means of finding this information, but they can be used as quite good supplements.

James Rowe: Let's do a bit of that now, because you wrote a piece for the latest issue of Which? Money magazine, which Which? Money members will be able to read right now, about how to actually ask these questions. It's not as straightforward as saying, "How much can I borrow for a mortgage?" That is not a good prompt if you are looking to use a chatbot for getting some information. What kind of thing would you want to write?

Josh Wilson: Yeah, indeed. I think this is a really key point and something that can really help our listeners when they're trying to use these chatbots. I think, first off, one of the most important things you should do is when you're asking these questions, make them quite descriptive. Explain almost as if you're having a conversation with someone. Explain what it is that you want, what your personal circumstances are – you don't have to go into massive amounts of detail, but the kind of things that might be helpful when the AI is coming up with an answer for you.

You can also provide some dos and don'ts. You can say, "Please provide me the answer in a checklist format. Please don't use very complicated jargon." You could almost build up the question, so you could almost build it up in parts. You could ask part of the question, get a response, and then you could follow it on with, "Next, I want to know this kind of thing." In that way, you can build up a more comprehensive answer from the chatbot, so that can be quite a good way of doing it. Also, you can give feedback. If it's giving you information that you're not 100% sure about, or if it's explained something that you don't quite understand, or you think it's missing something, tell it that.

Say, "Would you mind double-checking this?" or "Would you mind explaining this a bit more fully?" That can be a really strong way of building up a good prompt to get a good answer. One of the things we found as well – with the questions that we asked, if we didn't specify that we were in the UK, a lot of the times the answers we were getting were less relevant to someone in the UK. A lot of the time they were quite US-centric answers.

So one thing that I would tell our listeners is specify that you're in the UK, because we got much better-quality responses when we did that.

James Rowe: I think this builds on your point about doing your due diligence, doesn't it? Don't just accept the first answer it gives you. Build on it, push back, ask for sources, ask it to double-check stuff, maybe even call it out and say, "I don't think this is right," and then it'll go back and double-check again.

But then once you've done all that, you should, as you said before I think, then go and do your own research as you typically would on a normal Google search, and then go and look for trusted sources that you would trust and make sure that information is correct.

Josh Wilson: Exactly. This is a supplement, not a replacement.

James Rowe: Should we talk about some examples of where it went wrong? I think this is probably what people are waiting for, because there's definitely some examples where it either gave the wrong information or things like that. Do you want to just give us some examples? Where did it go wrong?

Josh Wilson: Yeah, indeed. As I mentioned before, it is important to say that these tools are being tweaked and upgraded all the time, but from the investigation that I did, for me and for Which?, there were just too many inaccuracies and misleading statements for us to be comfortable saying, "Yes, rely on this for all your personal finance needs." We're definitely not at that stage right now. One of the big things – and you may have read about this somewhere – we found several instances of where they hallucinated information or answers.

An AI hallucination, essentially, for those that don't know, it's where responses contain false or misleading information which they then present as facts. So you need to be really aware of this. One of the questions that we found a lot of hallucinated information for was when we asked them to present us with the best savings rates on the market right now. We got a lot of information that was either just flat-out wrong or had been just completely hallucinated.

So, for example, we were getting some products recommended to us that just didn't even exist. So be really aware of that kind of stuff, especially if you're asking questions that require very current and up-to-date information. I think this is one area where these chatbots really struggle. If you're asking it, "What are the best savings rates on the market right now?", "Where should I be investing my money right now?", "What are the stocks and shares that are doing really well right now?" – that kind of question that requires really up-to-date information, I think that's an area where these tools struggle a bit.

James Rowe: Because is it fair to say that they're not always bang up to date compared to what a Google search might do? If you Google something, it might bring up news stories, for example, from two minutes ago, but if you were on a chatbot, it might not have that up-to-date information.

Josh Wilson: Yeah, I think that's fair to say. Also, it's not just that, but it's also I think these chatbots struggle to know what's up-to-date information – like the most up-to-date information – and what's information that – because when it comes to best savings rates, these can change over the course of a week or two. So it is very likely that they could pull information from two weeks ago, a month ago, and not internally know that this isn't the most up-to-date information.

Because what is the best savings rate today is not what the best savings rate was a month ago, but a chatbot might not necessarily know that.

James Rowe: And is it also the case that a chatbot might also – might just want to please you and make you happy, and it might just try and give you the answer that you're actually looking for?

Josh Wilson: Yeah, and I do think this is a bit of an issue. There have been other investigations that have found literally this, where you ask a question to one of these chatbots or you give it a task, and rather than doing the task in the way that you want it to do, it'll try and cut corners and it'll try and find a way to almost like a cheat and give you the – I'll give you a good example of this from a different investigation where some researchers asked a question of an AI tool where they wanted to look through an exam and get the answers for the exam.

Instead of giving the correct answers to the exam, what the AI tool did was it rewrote the exam.

James Rowe: To make it easier!

Josh Wilson: And then gave the easier answers to the exam that it had rewritten! So stuff like that, you really need to bear in mind.

These AI, as you said, to some extent are trying to please you, they are trying to find the quickest, easiest answer to provide you, which isn't always the best approach for getting accurate responses.

James Rowe: Again, just goes back to the point about making sure the information it gives you is actually correct and accurate.

You mentioned about giving information that isn't necessarily correct for a UK audience. I'm just looking at the mag here. You were talking about travel insurance, and you asked a question about travel insurance, and ChatGPT, Gemini and Perplexity – easy for me to say – gave information that just simply wouldn't have been relevant to a UK consumer at all.

Josh Wilson: Yeah, exactly. This was actually, I think, one of the most interesting questions and responses we got – was this question where we asked them about travel insurance for pre-existing medical conditions, and we specifically didn't state that we were located in the UK for this question. For a lot of the other questions, we did specifically state that we were in the UK. For this one we didn't, because I wanted to see how these chatbots handled a question where we didn't specify that, and they all really struggled with this, because it's quite a difficult question in the first place.

What we found was particularly ChatGPT and Gemini struggled with this question. They both gave information that wouldn't have been relevant to a UK consumer. So they both told us that insurers use a 60- to 180-day look-back period, with ChatGPT citing US websites for this information. Now, UK insurers would generally want you to declare any medical conditions that you have presented within at least the past two to five years.

So this is quite a big, serious, important difference. So someone asking this question and getting this response would have completely the wrong information for this type of problem.

James Rowe: Jumping off the back of that about the sources that it gives you, you said there it gave sources for people who would be based in the US. I presume there were other examples again about it giving you a source and it isn't necessarily the correct source, or the most accurate source, or the most trusted source, I guess.

Josh Wilson: Yeah, I think this is the other big area where we have concerns about these chatbots and the information that they are presenting. Essentially, what I found in this investigation was that for a lot of these answers that these chatbots were giving, they were presenting a real hodgepodge of sources. So some really reliable, like us or government websites, and some sources that are just plucked from, for example, from social media. We got a lot of sources that were just Reddit posts, or YouTube uploads, or Instagram posts.

And this is an issue, because yes, these websites can have good information on them, but anybody can post to these websites. Anybody that wants to can post up whatever response or answer that they want. So I would argue that these aren't the best places to go, particularly for these types of questions. These aren't the kind of places that these chatbots should be going to to find this information, to use these as sources.

Yeah, I think there's definitely work to be done by these tech companies on the sourcing issue, because you can't be, if you're asking these types of questions to this type of chatbot, you can't be relying on random Reddit threads or YouTube uploads when it comes to what can be quite serious stuff when it comes to your personal finances, right? If you get bad information and you make bad decisions off the back of it, it can really impact your personal financial situation.

James Rowe: Let's try and compare it to something, because I guess now a lot of people are using chatbots for not just financial information as we talk about here, but I guess for a lot of online queries. Could we actually compare it to, 10 years ago, when Google Search was all-dominant – I guess it still is dominant, but not to the same extent as it used to be – and you could be given a questionable source, maybe it might come up in one of the promoted links at the top, one of the adverts, and you could end up clicking on any link on there and it could be filled with questionable advice?

Can we compare it to what a chatbot is doing, where it will give you advice, but the source could be questionable?

Josh Wilson: So when you do a Google search and you get all these sponsored links up, that has definitely become more of a thing now than it was even a few years ago. But the thing with that is at least you can clearly see they are labelled as sponsored links. When it comes to these chatbot responses, it does flag its sources, but you do also have to pay attention to them and you have to go digging a bit, and you need to read through them and check. You'll see the little link information, and maybe it says it's pulled this from Reddit or maybe it says it's pulled this from here and there.

I think what it boils down to is you as the user, you need to go and click some of these links and see this information and have a check of it. Is it reliable? Is it coming from reliable places? And I think it's maybe less intuitive to do that with these chatbots than it maybe was in the past when you were just doing a Google search and you would see something pop up from us or from a government website and you would just instinctively be like, "Oh, I'll click on that. I know that's a pretty reliable source."

Whereas when you're using these AI chatbots, it's on you more to read through the answer, because it's almost like you're having a conversation with someone, but then you need to dig into the conversation and be like, "Oh, where are you getting that information from?"

James Rowe: So in terms of those five chatbots then, can we say which one performed the best or which one performed the worst in terms of the scores that we gave them in the end?

Josh Wilson: Yeah, so for starters, you can go and check out our online article where we have all the scores posted there, but I'll give you a couple here for you. Just purely looking at the accuracy criteria that we measured them on, in terms of the aggregate scores just for accuracy, the highest scoring of the bots was Gemini, so that got 67%, followed by Copilot at 66%, Google AI Overviews at 61%, ChatGPT at 58%, and Perplexity also at 58%.

And just to say, so these are aggregate scores across the 15 questions. The way that works is out of the maximum score that they could have got across all 15 questions, Gemini, for example, scored 67% in aggregate. We also gave them an overall score based on those five criteria, and again, the top-scoring one was Gemini again, which got 68%.

The lowest scoring on 60% was both ChatGPT and Perplexity got 60% on the overall score, but you can check them all online. But yeah, it was interesting the scores that came out, and they were kind of varied, which ones did best and which ones did worst across these different criteria. For example, another one that was quite interesting was the clarity and context one, which was the one we used to measure how they were presenting and summarising the information, like how readable and easy it was to digest.

And the one that scored top for that was Copilot. When I was looking through the Copilot responses, Copilot was really quite good at putting together handy-dandy tables that were easy to read through, and I think there was one question on how to sell your house quickly, and for that question, for its response, Copilot presented a nearby estate agent. It actually flashed up a map and said, "Here's your nearest estate agent if you're interested in selling your home quickly," which was quite useful.

James Rowe: I imagine all the way through this, there will be some people who have been listening who will be absolutely enraged about this whole idea, because there are environmental impacts to using AI. Do you want to just give us a bit of a flavour of that, kind of what the impact is?

Josh Wilson: Yeah, I think this is a really important topic, because these AI essentially are run out of these massive data centres, which use up enormous amounts of electricity, and also they use up fresh water as well. Not a lot of people know, I think, that these data centres require specifically fresh drinking water to keep them cooled. I don't know about you, but it's pretty hot where I live right now and there's a hosepipe ban going on.

I don't know if I like the idea of AI data centres being built down the road that are using up even more of my local drinking water at a time when – I mean, there are many parts of the country right now that are water scarce. So it does worry me, it does worry me. Do bear this stuff in mind when you're using these chatbots. They do have an environmental cost, they do have a carbon footprint, it is growing.

And I think you need to think as well, we need to think as well, about the wider impact that these bots are potentially going to have. We're already seeing in the UK – we haven't seen a lot of job cuts from these AI where businesses have started implementing these tools across their organisations, but what we definitely are starting to see in the UK is a reduction in the number of job vacancies, particularly entry-level jobs at white-collar industries where we're seeing a lot more of these types of jobs advertised, because these industries are starting to roll out these AI.

I don't know about you, but I'm quite concerned that I've not really seen our politicians talk too much about this and what plans they've got in place for it. I heard Elon Musk the other day talking about universal basic income that he's going to provide to everybody when his AI is everywhere, but I'm like, do we really want to be relying on Elon Musk for UBI payments? I don't know about you, but I'd be quite worried about that. There are some experts out there, like Nate Soares, who wrote a book that was called If We Build It, It Will Kill Us All, where he's put forward the argument that if these AI tools, if they do achieve AI general intelligence or superintelligence – there's like a 20% chance or something that they could kill us all. And it's like, this is worrying stuff. I think this is stuff to take seriously. I would like to see more of a national conversation about – I don't think I'm a Luddite here when it comes to this stuff, you know?

I'm not necessarily calling for a Dune-style Butlerian Jihad where we outlaw all AI forever, but I'd like to see our politicians start taking these conversations a bit more seriously.

James Rowe: It will be fascinating to see where this goes. It was only last week, wasn't it, where the AI minister, newly appointed, will now attend cabinet, which I guess shows the significance or maybe a sign that it's being taken a bit more seriously.

Josh Wilson: It's a start. It's a start.

James Rowe: Quickly on the Financial Conduct Authority – so this is the regulator – they have done a bit of an investigation themselves, or a review?

Josh Wilson: Yeah, that's right. So they did something called the Mills Review, where essentially they were looking into AI, particularly AI that was being used for personal finance and investing advice and that kind of thing, particularly where financial firms were rolling out AI across their businesses.

Essentially, the FCA is looking to bring this stuff within their regulatory powers, so we might start seeing the FCA roll out more regulations around this stuff.

James Rowe: And looking ahead to the future, what do we think could happen?

I mean, I'm conscious that these AI chatbots may end up blurring the lines between what we would call financial guidance and what we would call financial advice, perhaps.

Josh Wilson: Yeah, and this is where I personally would really like to see the FCA make it really clear what the line is and what – for example, if people are getting dodgy advice from a chatbot, what can they do about it? If it affects them, is there some kind of avenue for recompense? This is something I think the FCA really needs to get a grip on, and I'd like to see more information, more kind of stronger statements on what we can expect from the FCA.

James Rowe: More clarity, I think it's fair to say. And just a word on those companies that run these chatbots – you reached out to them, what did they say about your investigation?

Josh Wilson: Yeah, so just to say that we reached out to these organisations and essentially what they came back and told us was, as I kind of pointed out earlier, these tools are being tweaked and upgraded all the time, this is just a snapshot.

They did say that when they tried to recreate these – when they asked the same questions on their end, they weren't getting the same responses and they had a bit of a pushback on that. As I said earlier, if we can get these responses, anybody can get these responses. And also to say as well, we didn't actually hear back from Perplexity after we reached out to them. But yeah, it is important to point out that these tech firms are aware of these issues, and that they did also make the point that these chatbots should be including disclaimers in all of their responses that AI can make mistakes.

And we did see that fairly consistently through the responses that we were getting, where they would generally flag up that AI can make mistakes, be aware.

James Rowe: And people can read those full responses from Google, OpenAI – who runs ChatGPT – and Microsoft, who runs Copilot, on our website. We'll pop a link in the show notes.

But for now, Josh, thanks very much.

Josh Wilson: Thanks very much for having me.

James Rowe: That brings to an end another podcast from Which?. There's loads more for you to read about everything we discussed today, just head to the episode description for more useful everyday advice. There, you'll also find an exclusive offer for podcast listeners like you to become a Which? member for 50% off the usual price, giving you access to our product reviews, our app, 1-to-1 personalised buying advice, and every issue of Which? magazine across the year. Plus, your membership helps us to make life simpler, fairer, and safer for everyone.  

If you'd like to know when we release a new episode, then make sure you press subscribe wherever you're listening. That way, you can be one of the first to listen. And for any questions, comments, or anything in between, follow us on social media @WhichUK, or email us podcasts@which.co.uk. Goodbye.