AI Interrogator: The Legal Challenges and Considerations of AI Development with Lorna Woods

Kate Bevan: Hello. And welcome to this edition of the Infosys Knowledge Institute's podcast on all things AI, The AI interrogator. I'm Kate Bevan of the Infosys Knowledge Institute and my guest today is Professor Lorna Woods of Essex University, which is in the East of England. Lorna is an expert on internet law. She has been involved in policymaking, across a wide range of technologies, including social media, data protection and communications. And she's now looking closely at AI regulation. Lorna, welcome. Thank you very much for joining us.

Lorna Woods: Well, thank you for inviting me.

Kate Bevan: I'm delighted to have you. Given your background and working with so much internet regulation, generally, where do you think we are with AI regulation? Are we going in the right direction with it?

Lorna Woods: Well, I suppose, it depends, who you mean by, "We," because we can already see that there are some different ap-proaches. I think everybody's a little bit behind the curve. I think we could have learned from our experience with social media, with the development of the big tech companies we've got already, that you can't rely on self-regulation and good intentions. I think there has to be a clear intent to developing safe products. And that, given the pressures of the market, given the pressures of shareholders, that doesn't happen, unless they have a clear incen-tive, through the application of a legal regime. Obviously, getting that balance right, between a legal regime, which sort of safeguards the public interest, while not stifling the market, stifling innovation, completely, is easier said than done.

Kate Bevan: How much of a problem is it, also, then, that we have lots of different regulatory regimes. We've got... I don't think you have data protection. We've got the GDPR in Europe. America's a real patchwork of different regulations. And then, you've got a completely different approach in places like China.

Lorna Woods: I think that's going to be a challenge. And you've given an example of jurisdictional differences, But, actually, even within the same jurisdiction, you're going to be looking at lots and lots of rules intersecting. If we take a look at the EU, because they've just agreed the AI Act, that's got to sit against the Digital Services Act and the GDPR and proba-bly some consumer laws. And, although, the introductory provisions, the recycles, always say that they're going to work together and sit against the existing rules, that's more difficult, when those rules weren't initially drafted to take into account the particular technological developments.

So, there is going to be an issue in terms of understanding how rules link together, within one jurisdiction, even be-fore we start taking into account the difficulties we get when you've got multiple jurisdictions and pan-national companies that have got to make a decision about how they engage with those, against a background where they are based somewhere. There will be a corporate culture and a particular orientation, and that may not sit as well with certain jurisdictions, as with others.

So, I think there is a tension there. And, particularly, there's going to be a tension, where we are looking at more general legal rules. I mean, it's one thing to say, "Oh. We've got rules about AI and everybody understands that AI is difficult and it is big companies," but if you start looking back at, say, data protection or consumer protection, those are much more based in the actual legal systems, the various legal systems, much closer to Joe Bloggs, Josephine Bloggs, on the High Street and what people expect of their day-to-day lives. So, I think it might be easier to talk about, "Oh. Well, let's have international standards, when we're just talking about corporate stuff," much less so, when it is showing a real direct impact on people.

Kate Bevan: So, when do you think, this is going to show itself, as a problem? What are the problem areas?

Lorna Woods: Well, we're already starting to see some of them. I mean if we're talking AI, obviously, there's the different stages. So, you've got the initial design stage, you've got the creating the data set and the training of the model, and then you've got the deployment of the model. So, they're all different stages. If we just look at model creation, it's, "Where is that data coming from?" And so, certainly, the large language models, they've just gone off and done mass scraping, and there are issues, even if we say that the sets are accurately, I know, labeled and cleaned up, there's the copyright problem, and then there's the data protection problem, and then there's the illegal content problem. So, I think there's three areas, where they're starting to share

Kate Bevan: The copyright problem is one, of course, that's quite high and on everybody's minds, at the moment, with all the cases happening. What's your view on that? I mean, I feel very, I mean, a little sympathy for the content creators, who are saying, "Look, you've just scraped my content and you've fed it into this giant sausage machine and it's churning stuff out." But, at the same time, I'm also thinking, "How can you prove, as a content creator, that your work has been remixed to produce that output?" What's your take on that?

Lorna Woods: It is actually quite a difficult area. I mean, some of the big data sets have said where they've got their content from, but there is going to be, I think, increasingly a problem, around that. There's going to be a tension between pres-sures towards transparency, as to sourcing and I suppose a technical response, on the part of companies creating these massive models, that it may help them to obscure the issues, just in terms of the basic proving in court. I mean, obviously if, just as a matter of copyright law, if you are starting to get outputs that look like particular copy-right issues, I think it would then be hard for a user or a creator of a AI tool, say, like ChatGPT or something, to sort of push against the inference that that's where it had come from.

And I think there's probably a difference in terms of if you have the unfortunate experience that the output actually matches a copyright work. That, I think, is a completely different situation, from the creation of the data set, in the first place, even if it is technically an act of copying the original source. I think there are other issues, when we con-sider different forms of content, though, say, if we're looking at confidential content or content that is either per-sonal data, from the perspective of individuals, all falling within various privacy protections. Some countries have personality protections. For others, you might start looking at passing off, if, say, advertisers are using cloned voices, all the like. So, I think there are at least a slalom of legal issues to be negotiated.

Kate Bevan: As we start to understand, I think that what is going to be the useful generative AI tools are going to be the small ones, the ones developed by corporates, that are just for managing your own database of work. Is generative AI going to get more useful or less useful, then, as we go forward?

Lorna Woods: I suppose, it depends, what format it ends up taking. I have questions about whether general, not particularly tar-geted forms of gen-AI tools we'll have, there are also many issues with them, around accuracy. I think, in particular, I think that the more targeted, the fine-tuned models, are going to be, for day-to-day businesses, more effective, more reliable. I suppose, that then leaves us with the problem of people who don't care about accuracy. And maybe this is the big market for the gen-AI, the large-language models. You know? The people who are generating clickbait.

You know? We still have the clickbait farms and the modern versions of, I guess, astroturfing or cybersquatting, where you are getting very low-quality, AI-generated journalism, just to get the ad revenue. That's possibly the be-nign end. You have disinformation campaigns, where what is important is persuasiveness and not accuracy. And I think that's where the problem lies, where users don't have a reputation to lose and they're just worried or aiming at either volume, preferably with persuasiveness. And you can see concerns about, also, scam artists, as well, in that bucket of issues.

Kate Bevan: Particularly, as we go into a really important year for elections and quite a lot of economies. I mean, not just the UK and the US, but there are other countries as well. I think India's got a big election this year. What concerns me, how can we regulate for that? Where does the moral come into this?

Lorna Woods: I suppose part of it is, if you want to get there, I probably wouldn't start from here. There is a point at which, if you've had some of the models, let loose the open-source models, then it is more difficult to control them. What we see is, I think, an emphasis on watermarking and labeling and that sort of thing, on the assumption that that will help, I suppose, people with their media literacy efforts. I think that the evidence, so far, on how effective inocula-tion or counter narratives are with fighting misinformation, I don't think it's really clear, as yet. I think there's differ-ent methods with that.

I do worry that if you've got hyper-partisan audiences, you actually don't even need to get as far as a sophisticated LLM gen-AI generated piece of fakery. A quite unsophisticated Photoshop will do the job, for some audiences. I think there's an issue about emotionality and belief coming in there. So, I don't think we're quite the logical people, but I think the media literacy narrative assumes. I mean, obviously, it does something, but I question how effective it's going to be, especially if we set it against the existing social media networks that we've got, at the moment, that, to a large extent, are optimized for engagement, for virality, for emotion and the like. So, yeah, it'll be interesting to see how that pans out. I do have a concern.

Kate Bevan: Maybe I'm being a bit Pollyannaish here, but I feel slightly optimistic that we have had these learning experiences, from the past couple of US elections, from getting a better understanding of the impacts of social media. And there are lots of great brains, like yours, like many other women's, particularly, getting out there and thinking hard about the potential harms and dangers coming. Do you think we're getting out ahead, in front of it?

Lorna Woods: We have learned, and, again, it depends, who the, "We," we are talking about are, because you can see that the UK government has taken quite a pro-innovation, "Let's," in a way, "Worry about it later." I have some sympathy for the idea that you should apply existing law and that some of this will fall within the remit of some of the relevant regula-tors.

So, yes, there is something there, but given we're only just now in the UK and in the EU, starting to really implement those social media rules, which have been around for the last, well, nearly two decades, I'm slightly concerned about an approach that says, "Wait and see," because it is building a lot of delay into the timeline. Obviously, if you move too fast, you see the problem that the AI Act ran into, that, until quite the last minute, it didn't really think too much about foundational models and the particular issues that gen-AI brings up. So, there is possibly some advantage to waiting, though, not too long.

I don't think we've got a handle on how the previous elections were really manipulated. I don't think we know, real-ly, what happened still. There was a real issue about data access. Then, there's a certain amount of over-claiming, from some of the people involved. A certain amount of denial about other actors who participated in the process. So, I think it's actually hard to get any clear lessons in terms of solutions, from, say, the 2016 elections, the 2020 elec-tions.

Kate Bevan: If you could make one provision in legislation and regulation with AI, what would it be?

Lorna Woods: I would say, product safety applies. I think it is easier to try to create something that's, for want of a better word, safe, when you're at the design stage than to retrofit plasters, Band-Aids. And I've said this before, I've said it in rela-tion to social media, and I say it here, I think we're looking at a more thoughtful design process, a more human-focused design process, and a design process that involves a greater diversity, in being heard about what concerns are. And I think it has to be an iterative process, "Let's talk to civil society and then go off and forget about them." Because I think the one problem, obviously, with design process, is that things change, so you need to design and revise. You can't just go, "Okay. We've done our design thing, we fixed the problem." It's got to be ongoing.

Kate Bevan: I'm going to finish up with a question I ask everybody. Do you think the AI is going to kill us all?

Lorna Woods: Well, possibly, but not the sort of dramatic existential, "Computers take over the world," sort of response. I think we need to recognize the impact on our environment of these tools that require an awful amount of computing power, an awful amount of electricity, and a significant amount of raw materials. Just burning up the Earth's resources to generate memes, I don't think is sensible, so that's one form of risk. And the other risk is what I was talking about, that we might get disinformation campaigns that stoke tensions. It won't be a problem on our own, I don't think, but in combination with existing geopolitical tensions, then, they could be problematic.

Kate Bevan: Professor Lorna Woods, thank you very much, indeed.

Lorna Woods: And thank you for inviting me. Been a pleasure.

Kate Bevan: I'm so delighted to have you.

The AI Interrogator is an Infosys Knowledge Institute production, in collaboration with Infosys Topaz. Be sure to fol-low us, wherever you get your podcasts. And visit us on infosys.com/IKI. Yulia De Bari and Christine Calhoun pro-duced the podcast. Dode Bigley is our audio engineer. I'm Kate Bevan, of the Infosys Knowledge Institute. Keep learning, keep sharing.