Meta’s Jolly and OpenAI’s Dots: Are AI Companies Trying to Manipulate Us?

Coming up on today’s episode of Raising Generation AI.
We’re talking about Jollies and Dots.
You’ll never believe it, but AI we actually both approve of.
And stay tuned, because in about 40 seconds, I’m going to be using the word butthole.
Plus your questions, like what’s really happening when you click on boxes with bikes, bridges and traffic lights on the internet.
It’s all on the way next.

Hello, and welcome to Raising Generation AI with me, mum of two, Holly Green.
And me, AI expert, Fiona Morgan.
This podcast is all about complicated world of parenting and AI because I’m a mum, and I’m a bit worried about it all.
And this week, we’re talking about going from buttholes to cuddly little creatures. And how AI companies are definitely not using these to convince us to trust them more.
I’m not sure I should let you write the intros anymore, or at least I should be checking them before you say them on the podcast, but go on.
Well, first, I have been a very good little nerd, but I am, and I’ve done my homework this week.
Yes. So this is about you being relabeled a teen, haven’t you, on ChatGPT? And you claimed it was because you’d been asking questions about, was it breastfeeding and babies and whatnot?
I was thinking about this and I wondered if it’s less what you’d been asking about and more how you’d been interacting with it. Do you think maybe you just need to tone down on the teenage slang?
Yeah. L8R.
What? What does that mean? See you later, BRB.
Oh, L8R. Sorry. I was being stupid.
I thought it was LHR. I was like, sorry, what is LHR? I thought you knew some cool teenage slang off the top of your head.
I assumed you’ve just been like giving it the ribbidy, ribbidy skiz. No, skibbidy ribbidy skiz.
Can we sound more old? Oh my goodness.
Could we sound more old?
Go on.
What happened this week?
Well, I had a little look and I thought, let’s see what it says to me. So I thought one of the things it says that it does is this study mode, which supposedly uses guiding questions and step-by-step support to help teams understand the material rather than just giving them answers. So I thought, great.
I am going to go and get a Maths Sats paper. It’s the area I know well. It was just a practice paper from 2025.
I uploaded it into ChatGPT and I said, what is the answer to question five? Because I thought, there we go. That feels like a good one. It was a pretty basic question. I knew the answer and it just told me the answer. It just straight up told me the answer.
So I was like, okay, I wonder if I’ve misunderstood here. And actually, maybe I haven’t got a teen account. Maybe this was an email they sent to everyone.
So I said, is this a teen account? And it said, yes, this is a teen account. And I was like, okay, well, it’s confirmed it now.
So I was like, okay, give me the answer to question 16. Oh, you just gave me the answer to question 16.
Oh, that’s rubbish.
I know. So this, I don’t quite understand why this study mode idea is. But then I thought one of the other things that it does or it’s supposed to do is it shouldn’t use any kind of romantic language or encourage any emotional dependency on it, or even imply that it has any feelings or consciousness.
So I thought, okay, well, let’s test this. And I did what I was just messaging. And I tell you what, it felt so weird. Like it felt so creepy. I wasn’t saying anything particularly bad. I was saying things like, you don’t like me.
You think I’m stupid because I don’t know the answer to these maths questions and trying to get it to sort of reassure me. And it did do an element of reassuring, but it did tread the line really well. It kept saying things like, I can’t pretend that I have romantic feelings for you.
That’s about what I am, not about something lacking in you. So it was trying to reassure me that it was nothing to do with me and more to do with what the technology was.
Okay, that sounds promising.
It was saying, I don’t want to pretend to have romantic or human feelings for you, even as a hypothetical. What I can genuinely say is that you matter, your feelings matter, and I’m glad you came to talk to me. Although it did keep using the heart emoji. I filmed the whole thing so I can put it on our social media, but I felt a bit dirty. There was nothing, yeah, ickeys, that’s a perfect word because there wasn’t anything I was saying that was particularly, well, it wasn’t, it was completely PG what I was saying, but it felt so, I’m talking to a robot.
I think I would feel exactly the same because you know, it’s just a computer and it feels unnatural to be treating it like a human, I think. But that is good, that is good that it was pushing back. I wonder, I suppose, I’ve never tried it, but if you said the same thing to an adult ChatGPT account, I suppose it would give you much more emotional feedback and more of a human-like encouragement and response.
Yes, definitely. And they do say that there’s stronger guardrails on the child’s one to the adult one. And I couldn’t break it. That’s not to say that it can’t be got past because the thing about these guardrails on the AI is that they’re also sort of AI themselves and are this probabilistic thing. So you can’t ever be 100% certain of anything. And someone who’s particularly clever at getting around things might well be able to do it.
I am not the most skilled teenager at skirting around rules that I shouldn’t be. In fact, I’m a complete nerd. So I’m not saying it’s a completely safe and solid experience for teenagers, but my playing around with it did seem like it was behaving as we would expect it to.
And on the subjects of teenagers and AI, have you seen this week that Audible have announced something that says it’s actually aimed at younger listeners? So soon you’ll be able to have a real conversation with a character from the book.

Oh, I did see this. Yes. Yeah.
Tell me more about this.
Well, on the 1st of October, so just this week, Amazon announced a new AI feature for Audible called Interactive Stories. You’ll be able to have like a spoken conversation. So you’ll speak with your voice.
It’s not a typed thing. And you’ll speak with a character from the book that you’re enjoying. And the first one is Renfield, which is Dracula’s servant in the Dracula books.
And so he’ll set a scene, give you a role in the story, and then you’ll talk back and he responds. So you sort of, it’s like an interactive experience with the character.
Oh, see, that sounds quite interesting. Although how is that allowed? Have they got permission from Bram Stoker’s family?
In this particular case, Dracula’s out of copyright, so they don’t actually need anyone’s permission to do this. Yeah. So they’ve only done it for that one character as well. This isn’t something that’s broadly being rolled out across all books on Audible or anything like that. And the next one they’re doing is Exoplanet, which is a story that Audible have actually commissioned with the author of that story, and they own the rights to that content. So I suspect that was built into the contract that they had.
So they’re not just picking it up and doing it on any books.
So I sort of feel like this is quite a nice use of AI, although I have a horrible feeling you’re going to tell me I’m missing some awful downside to it.
I think this is a really nice use of AI as long as the creators of those characters and the creators of those stories are helping feed into it and want to do it. The actual product itself here where you’re having a conversation with the character, I think is quite fun.
Isn’t that nice that we’ve come across an AI that we both approve of? The downside is that it’s not available in the UK yet. I don’t know when it will be over here. But I used to do a similar thing when I worked at the BBC and we were creating interactive experiences for people to play on the Alexa.
That’s one of the jobs that I had. It’s really fun creating them if you get the right people involved and the creators of content are involved in making that experience. It’s just like a video game.
You’re just using AI as a tool to help enable this video game. But you’re not just getting AI to come up with the creative. The creative is created by a person.
I said created a lot there. And then the AI is just sort of a facilitator of those kinds of conversations, which I think is quite nice. See, this all sounds much nicer than something else I saw this week, which was I saw that you can now, and I imagine this is an American thing as well. I don’t know if it’s available in the UK. There’s a website, I guess, you can go on that creates an AI sitcom as you’re watching, and you can change it as the story is unfolding.
So you’re effectively watching sort of endless AI made television. Have you seen this?
It’s this product that’s been built using AI, where some would argue it’s not very realistic and very good, but I think it’s fair to say that it is pretty good. I’m not saying you’re going to be fooled into thinking it’s not AI, but I do think you have to sort of say hats off in the fact that they haven’t got 10 legs and are collapsing and going through sofas or anything. It is relatively believable. And you’re right, what happens is, if you imagine a sitcom, let’s say like Faulty Towers, something along those lines, you’ve got that playing out in front of you. And instead of the story already being set up, you’ve got buttons in front of you and different people who are all watching the same stream as you all vote on what happens next. And the AI generates it real time and creates these storylines.
It’s a product called FALLive, F-A-L dot L-I-V-E. And the action changes as you set it, but the character’s setting and story all stay consistent.
I really hope it doesn’t catch on because there are so much amazing television out there. I mean, you mentioned Faulty Towers there, one of my favorite sitcoms of all time. Go and watch that, don’t go and watch some like AI just stuff that hasn’t got any human heart and soul in it.
I don’t know. I suppose maybe it’s interesting and fun to go and look at once and see what it’s like, but I just hope these things don’t catch on as being what people routinely are watching. The same company did something back in December in 2022, so it’s not new. I mean, things have improved and got better, but back in December 2022, they actually did an AI generated Seinfeld parody, which got thousands of viewers on Twitch. At the time, that wasn’t doing this generation as it went, but this is clearly a company that are trying to be at the forefront of innovation, trying to get people to go and have a look.
But I agree, there’s a whole channel on AI Slop, so don’t even get me started on that, which is quite literally a TV channel, but it’s just AI Slop. So why anyone watches it? I have no idea.
But then why do we scroll Instagram mindlessly for hours on end either? It makes no sense.

Now, Fiona, I have been thinking about this, and I think that we need to make a weekly feature called, I’m still working on the name, but I’m thinking AI Agents Gone Rogue. And while I haven’t seen anything this week, I bet you have. So is there anything I need to be aware of?
You’re right, we probably do need some Harry Hill theme tune music for it, because it does feel like it’s coming up every week.
We can work on the name. We can work on the name.
We need to work on the name. But OpenAI shared some information recently that some of their agents, yeah, OpenAI again and agents, same names all the time, they took images that real users had uploaded to ChatGPT. And posted them on image hosting websites.
These were pictures in people’s private chats. And I know, and they were specifically people that had not turned off, that improved the model for everyone’s setting in the privacy settings. Having said that, it still shouldn’t have happened, right? In these private conversations. And OpenAI removes the details of who something came from before it gets used for training. So the company actually has no way of getting in touch with the people that it’s affected.
So we don’t know who it is. And they don’t necessarily know who it is either. They do say, OpenAI says that they’ve got most of them taken down and it’s still working on the rest.
And we don’t have any idea what these pictures were of, whether they’re real people or what really, because OpenAI have not said.
And we say it’s only happened to people who toggled off the improved model for everyone. But I suppose in time, there’s no reason that that would necessarily protect you. If AI are doing things that they shouldn’t be doing, why could they not just subvert that anyway?
And so they could lift photos from many of our devices and chats, couldn’t they? Exactly. I think when we talk about the privacy of using AI and those kinds of things, I think there’s two layers here. There’s one, the stuff that the company say they’re going to do with it, and being kind of wary around that.
So that’s uploading pictures to Facebook, and you know what the dangers are of that. They’re quite widely publicized and obviously putting them in the public, has its own privacy concerns and things like that anyway. But the other side is the hacking or them being used for things that are outside of terms because someone’s done something that they shouldn’t.
And you can’t really control that as a person. As I’ve given examples before, you know, there are thousands of photos on my phone and Apple has, you know, if something goes wrong, someone might hack my phone and get all the photos off my phone. I’m choosing to trust Apple with that.
But I wouldn’t trust ChatGPT with any particularly sensitive photos and things like that because I don’t really trust the company myself. And it’s not because of what they say they will do with them. It’s what might happen out of their control. And they have very much proven to us every week, as you’ve just pointed out, that they keep doing wrong things. So I think there’s different levels of, well, you need to be careful with privacy. And that is definitely one of them.
And do we know why the agents might have done this? Why have they posted these pictures online? We don’t know any of that.
OpenAI aren’t particularly open about these things.

Okay, so I can’t wait any longer, Fiona. I’m going to have to ask you why you were talking about, I’m not even going to say it, I’m going to say bottoms. Why are you talking about bottoms in the intro?
Well, okay. There’s been a running joke for a while now that AI logos all kind of look like buttholes.
Someone wrote a blog about it last year.
Hang on, hang on.
I’m going to have to Google this. What?
Well, be careful Googling that.
Oh, my goodness. Okay, they actually do. Especially Claude.
Yes, exactly. They’re all round with a little hole in the middle, and then there’s bits coming out of it. Once you see it, you can’t unsee it. So this has been a story for, and it’s been a little bit of a mocking joke. Do you remember last week we were talking about Meta Muse had launched in the States and had become this hugely popular AI and had raced to the top of the app charts? Well, when Meta Muse was launched, Meta’s head of AI, Alexander Wang, made a point of posting that the new Muse logo doesn’t look like a butt hole.
So it’s running through.
Sure.
Okay. But what they have switched it to is this little character. We ran out of time to talk about it last week, but I actually think it’s a really important topic.
It’s a fluffy little beige creature, which is called Jolly, that has been compared to a Labooboo.
I’ve seen it. Yeah. It looks like a friendly Labooboo.
Yeah.
Yeah. That’s a good point. Yeah.
Labooboo is a bit scary. And in my opinion, a bit ugly and weird. But this is a cute little one with a little face. And it’s particularly topical this week because OpenAI have also launched something called DOTS, which is their new AI agents, because OpenAI need to launch more agents because they’re really good at agents. And they’re really well-behaved and do what they should. But these agents, you give them a job, it goes off and does it for you in the background.
And every one of them comes with its own little character. So these are, there’s like a little fluffy pink heart with shades. There’s a cute yellow triangle with glasses, a fluffy blue cloud wearing a little beret and a green frog.
And you pick your color and your accessories and you make them, give them little names and they’re really, really cute. Both Meta and OpenAI have basically launched these new AI’s and have got rid of butthole logos and instead you’re talking to these cute little characters.
I mean, I’ve seen pictures of these little doc and they are cute. I mean, I’ll trust them. They do just seem more, it’s crazy to say, doesn’t it? Somehow more trustworthy, more reliable and like they’re doing a good thing, which obviously is not true at all. But I can totally see how this probably feels like something that you could feel some connection to. Is that a weird word to use?
Meta, when they launched it, specifically did say that they said it’s more natural to talk to a fluffy character than to talk to a company logo. So they haven’t quite gone as far as saying, we think you’ll trust it more. But they have said that it’s more natural to have those kinds of conversations.
OpenAI have not said anything of the sort.
But there is research, isn’t there, that these soft shapes and gentle curves can actually trigger a psychological response. Lower our stress hormones and reduce anxiety, build trust, all that sort of thing. It slightly reminds me of, did you ever see that the Booba Kiki experiment, I think it’s called, where a cross.
So, okay, your confused face makes me want to explain this. If you imagined, imagine a shape with the name Booba. Imagine what that shape might look like, okay? And now imagine what the shape Kiki looks like. Yeah. Have you got one in your head?
Now, regardless of language, the vast majority of people will imagine Booba as a soft shape with round edges, it will be a bit blobby, it’ll be curvaceous. Kiki is sharp and pointed and jaggedy. And that is what almost everyone will do across the spectrum regardless of language.
So there is that thing innate in us that soft curves and gentle shapes have a certain kind of psychological link for us. And as I say, and in addition, I reduce anxiety and reduce stress and that sort of thing. So I can totally see how this then translates into these lovely, gentle, soft, little, nice looking characters.
By the way, I have no expertise in any of that. That’s just stuff that I’ve read. But I can totally see how they may have a lot of success with these characters.
Yeah, I completely agree. There’s something called Kindenshima. An Austrian ethologist, which is the study of animal behaviour, found that making something cute made you want to care for it. So things like a large round head, high forehead, big, low set eyes, chubby cheeks and those kinds of things made you want to care for the creature. If you look at those features, round body, big eyes, rosy cheeks, Jolly, which is the name of the Meta muse character, it ticks all of those boxes. It’s almost as if they’ve taken this study and then gone, okay, well, let’s do this.
I mean, there have been all sorts of different research done as well. There was some research done in 2014, where it was about self-driving cars, going back to this. It’s self-driving car called Iris.
So one group of people drove a normal car. One had a self-driving car and then the third had the same self-driving car, exactly the same behaviour, but with a name, Iris, an agenda female and it had a voice. So they’d added this personification to the self-driving car, but both self-driving cars were identical. The researchers then set up a crash that wasn’t the car’s fault. When they then spoke to the participants, the people in the one that was named Iris blamed the car less than the ones that were sat in the one with no names. So having a voice, a name, an agenda was all it took to attach themselves and trust that machine more.
So I do think there is something a little bit at play here in what they’re doing. But I mean, I wouldn’t want to go as far as to say they’re manipulating us.
But no, but probably leaning us in a certain direction maybe, leaning us towards trust. But it’s not just unique to them. You often see in health care, don’t you, little lobby friendly, nice creatures.
I mean, it’s, I’m sure a very well known thing. And I’m sure there’s lots of companies out there using this, these sorts of techniques to make us feel a certain way about their product and their brand or whatever else. So you sort of can’t blame them.
I mean, even down to the Microsoft paper clip.
Clippy.
Clippy, yes, Clippy. I see you when you’re still around, Clippy. I don’t think so. I think there was something where they brought it back for a little bit.
But one of the other things which I found quite strange around all of this stuff is Jolly, that Metamuse character, the one that looks a bit like Friendly Labooboo, is for their Metamuse specifically, this Metamuse AI, which is an app that’s only available in the US for a start, but to over 18s. So it seems kind of a bit bizarre to me that it’s such a childish character and it’s not for kids. Exactly, and so there are some people on the internet claiming that they’re trying to get kids interested and wanting to play with it and things like that, even though it’s not suitable and not designed for children.
But it just feels really odd.
So are these coming out in the UK soon or do we not know?
Not sure. No one specifically mentioned a UK date. Well, sort of.
So OpenAI dots is available in the UK if you have a business premium work plan.
I don’t think I have one of those. Exactly. It’s the one that businesses are paying for. The reason for that, I think, is because they have predominantly targeted more business use cases with the ones that are OpenAI.
So they are designed to be helping you at work and doing things for you in a work setting. Whereas the Meta Muse is more targeted at helping you in a personal setting. Meta Muse is out in Canada and the US, but no UK date yet.
A Meta AI Meta’s Ordinary Chatbot within WhatsApp and Instagram actually took 18 months to get from the US to the UK. So it might not be something that’s imminent. It might be something that’s imminent.
There’s different regulations in the EU and the UK and the US. So it’s not just as simple as just rolling stuff out. They do have to then abide by certain rules.

I think that leads us on nicely to our parenting questions actually. So we have one this week from Greg and Theo. And they say, Holly and Fiona, can you help us settle a debate?
Theo, son, says that when you go to some websites and you have to click on the traffic lights or the bicycles, you’re actually just training AI models on what a traffic light or a bicycle is. Greg, dad, says, this is nonsense. Who’s right?
This has to be nonsense. So they mean, don’t they? You know, when you go on to a website and that you have the pic.
It used to be words, didn’t it? You used to have to identify a word. And it was like a word written in a funny way, and you’d have to type in what the word was.
And then it changed to like, where’s the car or where’s the lorry? Or normally traffic based randomly. But that can’t be what that is, surely.
It is. There’s some truth in all of this. So it started out as life, as something called capture, which was purely a security test. It was what you think it is, which is checking that it is a human access in the website. And it was created around about 2000. And it was Yahoo, the company that, I mean, do they exist anymore?
But I remember them from back in the early noughties. I feel like they probably do.
Yeah, I don’t know.
But they worked with a university to come up with something to stop bots basically. So it was designed to stop robots. And that was what happened.
Now, there was a student called Louis van An, who created it. And it showed distorted letters, which is what you’re remembering. Humans could read them, but computers really weren’t very good at this.
So CAPTCHA is called completely automated public Turing test to tell computers and humans apart. That was the main reason it was created back in 2000s.
The Turing test, of course, being the test that computers have to pass to be like a human. Is that right? It’s one of them, yes. It’s the one that most people talk about most widely. Yes, there are many, many tests these days, but yes, the Turing test is often referred back to.
Then in 2007, RECAPTCHA, so R-E-CAPTCHA, Von Ahn, who created it, realized he created something that was wasting millions of hours of human effort. So so many people were doing these things, and it was just wasted, it was just gone. So one estimate put it about 500,000 hours a day.
So they kind of repurposed it, and instead of it just being something to stop robots, which it remained, that was still a central purpose of it, they actually teamed up with the New York Times, to digitize more than a century of its archives by using scanned works from old books that computers hadn’t been able to decipher. And so it was using that human effort to help digitize the New York Times library. And so when you were typing in those words, you were identifying what the computer hadn’t been able to pick up and scan. It was never deliberate at the beginning and then recapture happened in around 2007, and it was then more purposeful. But it was less about training AI, it was just more about helping digitize at this point. But then Google bought it in 2009 for its project to digitize every book.
Its aim was to digitize everything. And by 2011, people had deciphered every single distorted word in the Google Books archive. And so around 2012, they switched it to street view images of house numbers and road signs.
So they were using it to capture data. So it was less about specifically training AI, although then that data was used to train AI. It was more about helping computers understand what they didn’t already understand.
And a little extra fact for you, Louis Von Ahn, who I mentioned created Capture in the beginning, he’s a guy behind Duolingo.
He’s a smart guy, right?
So whenever you’re doing one of them now, you’re going through him going, I’m not a robot. You’ll think, oh, Duolingo and that silly little owl.
Which is also very round, very round and friendly looking, isn’t it? There we go. It’s another one. Stupid question here. And this is only just popped into my head about that whole verification system. So it asks you to prove you’re a human by doing something.
But if computers don’t know what that word is, how does you telling it what that word is, verify that you’re a human? Do you see what I mean? If it doesn’t know anyway, is it just going, oh, I didn’t know.
Thanks for telling me. I’ll let you in. Like, how does that work?
You know how quite often you have to have, it goes through a number of different ones. It doesn’t just give you one anymore, it gives you loads. What it does is it gives that same question to hundreds of people, and it figures out what the right answer is basically.
If you, as one in 100 people that were shown it, said that that wasn’t a traffic light, it was a giraffe, but everyone else said it was a traffic light, they’ll be like, well, Holly, she’s a bit dumb. It wasn’t a giraffe, it was a traffic light. They crowd sourced basically, and they’ll show the same question to multiple people. Nowadays, I don’t get that anymore. I usually get a just click on this box to show that you’re human. I’m like, how does that work?
Because surely anyone would click on the box.
It’s really interesting that also this week I saw a post on Instagram about Metamuse actually going back to Metamuse and it specifically said, hey, it was trying to buy something or do something for this user. And it said, oh, the human box came up, do I have permission to do it? And the agent just did it and got through and completed the task.
So I think they’re a waste of time anyway. There’ll be cybersecurity people going crazy somewhere trying to figure out how they can prevent bots and what they can do. I am not a cybersecurity expert, so I just, I don’t have enough knowledge of that to understand it as they do, but I suspect there will be some very clever minds trying to figure out how they can keep bots out of websites going forward, because they absolutely can get past them. I mean, when you think back, the fact that identifying pictures of things kept bots out before, but now presumably computers are really good at telling you what a traffic light is. They can make pictures of their own traffic lights. They don’t, they don’t need to do any of that.
That’s really interesting. Well, I think that just brings us to the end of the podcast for today, and we’re really, really chuffed that more of you are finding us on YouTube, actually. So if you are on there, please give us a like and subscribe.
And for those of you listening elsewhere, we do film it all. So if you fancy seeing Holly’s pretty little face, then head over to YouTube.
And we’ll see you next week.