Paul Murphy Challenges Shift on Recommender Algorithms, AI Risks

Paul. Thanks a lot, and thanks a lot for all the presentations. Just to start with you, Mr. Godfrey, could I ask why Commissioner Man dropped the requirement for social media companies to turn off recommender algorithms? So it was in the draft online safety code published in December 2023, but by the time the final version was published in May 2024, it was gone. Yeah, so in the draft code, we didn't actually have a recommendation to turn off recommender systems. What we had considered or consulted on was the idea that we might have greater transparency about recommender systems. Just can I quote it to you? It quotes, measures to ensure that recommender algorithms based on profiling are turned off by default. That was in the draft. By default, yeah. So I think when we got a whole lot of feedback and we looked at the interaction between the online safety code and the OSMR Act and the Digital Services Act, and essentially the conclusion we reached was that recommender systems were better dealt with under the Digital Services Act. That was a better legal basis for it. And that feedback that you got, did that come from big tech companies? I think we had some feedback from big tech companies. We had feedback from NGOs. We had feedback from the European Commission. We also had extensive legal advice. And the feedback from the NGOs, there's a joint submission by 60 civil society groups who welcomed the inclusion of the call for a ban on recommender or turned off by default recommender systems. So the feedback from the tech groups was against and the European Commission was also against? So I think there's a lot of questions about the interaction between what can be done under the online safety code, which was there to implement the Audiovisual Media Services Directive, what can be done under the DSA, which is a maximum harmonization measure. And as I say, the conclusion was that in order to get the code in place, get it through the TRIS process, to avoid the prospect of having things struck down because they were inconsistent with the DSA, it was better to use the Digital Services Act as the mechanism for addressing recommender system issues. And the industry submissions claimed that, quote, recommender systems play an important role in ensuring a safe, predictable and trusted online environment by ensuring that users are connected to relevant and high quality information. And that goes strongly against the evidence that we've heard as a committee so far in terms of people being pushed down far-right pipelines, anti-immigrant pipelines, eating disorder pipelines, incredibly quickly upon setting up a blank social media account. Would you agree with that claim from the industry? So I think you're quite right that the way recommender systems work can have very negative effects on users, particularly children. As I said earlier, there are now much more extensive guidelines about how recommender systems intended for where children are using them need to be configured to meet the obligation to keep children safe online under Article 28 of the Digital Services Act. And so it's important, I think, that we make those things happen. So recommender systems obviously can have, do have some utility for users, but when they are configured with the aim of maximising engagement and only maximising engagement, then they don't keep users safe. They are, as we've seen, they promote rage bait, they promote echo chambers, they can promote, you know, the content that users engage with and are attracted by is not necessarily content that is healthy. Thanks a lot. If I could turn to Dr McQuillan and thank you, first of all, for your blunt assessment in terms of the problems of AI. I think it's a necessary part of the conversation and an important contribution. Just what you spoke about, we're currently witnessing the EU walking back the flagship AI Act in the face of pressure from Trump and big tech. Could you expand on that a little bit? Well, at the moment, the leaked propositions are actually to walk back both the AI Act and some of the GDPR and the, well, presumably, I mean, this isn't stated in those leaked proposals, but presumably under pressure from big tech and most directly from the Trump administration. We know how closely they work in coordination with each other. If I may just try to also bridge between that very issue and some of the conversations around social media and the question of AI and the state, which I understood from the chair was also the core issue of the committee, it's very interesting to hear about social media and disinformation and the influence of the platforms because these are exactly the same issues that are before, I would say, before the committee in terms of AI. You have the corporate presence and power, which to some extent is already beyond regulation and demonstrates that repeatedly, but more fundamentally, and this is in relation to the technology itself, you have a form of technology, the AI algorithmic systems, which are essentially going to introduce deep fakes and ranking systems of a character and type into the very functioning of the state. So it seems to me that the same issues that you're now reflecting, as you say the proof of the pudding is in eating, we've eaten it in terms of social media, and the question is whether you're going to eat it in terms of the rest of the system that you're in. You're going to introduce the stuff into the very system itself. Not only that, if I just finish briefly, you're going to do that by shoveling a lot of public money at it at a time when the companies themselves are talking about the fact that the whole AI industry is greatly valued, and you'll do that also at the considerable cost to climate impacts. Thanks a lot. I was struck by the point you made about large language models are sold as learning accelerators but substitute slot for critical thinking. They're becoming the first port of call for everything from essays to relationship advice. There are now ads across Dublin City and presumably across the country for ChatGPT, which has a couple standing beside a car that's clearly broken down and supposedly using ChatGPT to try and fix their car. I mean, I don't know if they're legally responsible if they get the wrong hallucination and they blow themselves up or whatever. There was a piece on the radio yesterday, there was a columnist speaking about how she uses ChatGPT or some AI, some LLM for medical advice. And there was a doctor in studio who I didn't miss, I missed the doctor I got home by then who was presumably horrified by this. But you also, when you go on Twitter, you know, people treating Grok as a font of wisdom and truth, whereas you're making the point that this is not, they're not about finding truth, they're about finding plausible rather than factual things. What are the kind of wider implications of this for our democracy? I mean, if you could ask ChatGPT for how you can fix your car, how you should fix your broken toe, well then presumably you can also ask Grok or ChatGPT who you should vote for. What's the impact in terms of the wider of your character sphere? Presumably people are doing that. And I think Grok is an interesting case because it's slightly less nuanced than a lot of the other systems, mainly because of the character who's in charge of it. So we did, in fact, see a situation where, which I'm sure the committee are already aware of, where Grok suddenly started recommending that people find out about the Great Replacement Theory just at the time when Trump was offering asylum to white South Africans because it had been directly interfered with by its owner. Now that's a very florid example and easy to dismiss what I'm really concerned about. And this is where I bring up the issue of education. As an educator, I can tell you that universities are at the front line of this very issue, that the constant narrative about how safe these systems are being made are really narratives about how to make this exact effect just more subtle. So really what you're introducing are very powerful influencer systems, which seems to be a common concern of members of the committee. You're introducing a very powerful but also very opaque influencing system into pretty much every question that everybody has about daily life. Thanks very much. I'll try and make it back for a second round if we're on to something. Thank you. Thank you.