Happy Friday!
How have we reached October already? Half term is already creeping up on us and it feels like they just went back to school. Maybe it’s the constant stream of the same types of AI stories that make it feel like we’re living in a strange kind of infinite loop.
This week I dive a bit into the new parental controls for teens on WhatsApp. Specifically why I think we’re being too accepting of the very little tech companies are doing to give parents the knowledge and control they need to best support their kids.
On the podcast this week we’re talking about whether the cute AI characters might make you trust them a little more
AI isn’t the first technology that has been accused of using manipulation tactics to get customers to stay longer and spend more – social media is entirely built on this premise – but the latest move from Meta Muse and OpenAI is bringing that conversation back.
Last week we spoke about Meta Muse, and in this very newsletter I wrote about how the character is a cute little fluffy creature, with an adorable face. Well this week OpenAI launched dots – and you guessed it, they’re cute little fluffy creatures with adorable faces. So we’ve moved from butthole logos (go on, look at the logos of Gemini, ChatGPT and Claude and tell me they don’t look just like buttholes!) to cute, fluffy creatures.
Is this all a plan to get you to trust their AI more, give it more data and allow it further into your life? There’s real research that backs this – so we’ll have a good ol’ natter about it in the podcast this week.
Why AI is potentially so much more dangerous than traditional software.
Yesterday the headlines announced “New parental controls for teens on WhatsApp” and so I thought I’d have a little dig through what they were. Obviously, given the name of the podcast, I was most interested in what was there about the Meta AI integration and I was pleased to see there was a section on that.
I was less pleased to see that it was pretty light on detail. You can pick between “standard” and “limited”. Ok, I thought, let’s find out more about the topics that these two options allow to be discussed. But nothing. I looked for more than an hour to find details on this and there is nothing.
I’d like to say this surprises me, but it really doesn’t. I suspect the reason they don’t list the topics explicitly is because an LLM’s guardrails are not 100%. So if they say it won’t talk about XYZ and I, as a pretty experienced user of LLMs, managed to make it talk about XYZ then there’s a problem. But if they’ve not actually outlined the topics then it’s a bit more of a grey area. Of course, listing the banned topics will also potentially spur more people to try, including the very teens the guardrails are there to protect.
But all this brings me on to why LLMs are a new type of worry for parents. Typical computer systems, including those in your children’s toys, are deterministic – which means that they behave in exactly the same way every time. You press a button and it plays one of three farmyard noises. AI is probabilistic – which means that it behaves differently each time, based on a probability. The safety issue with this is that it cannot be comprehensively tested because we cannot replicate all the potential situations that might apply.
Think of it like this. If you had a cute little farmyard toy for a 2 year old. Each time you press the button it plays a farmyard sound with the idea that your child can guess the animal. If this was built in the traditional way there would be a bank of noises built into the toy and it would randomly deliver one. That might feel probabilistic, but it’s not really – we know that it will play one of the noises in the bank every single time. So the testers of the toy can keep hitting the button until they’ve tested and experienced all the noises and they are happy. And you, as a parent, can test it yourself too.
If that same toy is built using AI technology then it might take into account the time of day, the location, whether it’s flat on a surface, the age of the child, whether the child responds well to getting things wrong, the excitement level of the child’s voice…. I could go on, but I think you understand. With all these different variables it’s impossible to replicate every possible situation that the toy could find itself in. Which means most of the time that toy plays a sound it’s doing it for the first time, because the chances of that exact scenario having been replicated in testing is very slim.
What this means is that the experience your child gets from AI-driven toys, software, games etc. is unique to them. And this means that no one can be 100% certain that what they experience is good, or safe. There are guardrails in place to try and prevent bad things from happening, but these are also not 100% because they often use the same type of technology. So it’s impossible to test the experience your kid is going to get when it comes to AI.
I fear that the wording Meta has decided to use for these parental controls suggests that parents should be confident in their guardrails. I’m not saying that their guardrails aren’t good enough, they’re a massive company with a lot of money and a decent amount to lose, so I suspect their guardrails are some of the best. But if they are driven by AI then they cannot have been tested for every scenario. And I think parents deserve to know the truth about that, and not just assume that everything will always be fine.
Have a lovely weekend
I hope you have a great weekend and we’ll see you on Monday for the next episode of Raising Generation AI.
Any questions you have please hit reply on this email or get in touch on any of our social accounts: Instagram, Facebook, TikTok and YouTube.
Cheers!
Fiona