WEBVTT 00:00:06.570 --> 00:00:08.789 Good afternoon, and welcome to another episode 00:00:08.789 --> 00:00:11.289 of the Oxford University Undergraduate Law Journal 00:00:11.289 --> 00:00:14.089 podcast. I'm your host, Isaac, and today we are 00:00:14.089 --> 00:00:16.570 joined by Mr Adrian Mak, a fellow at the Stanford 00:00:16.570 --> 00:00:19.730 Law School AI Initiative. Adrian has contributed 00:00:19.730 --> 00:00:22.370 extensively to academic discourse on the intersection 00:00:22.370 --> 00:00:25.109 of technology, dispute resolution, and private 00:00:25.109 --> 00:00:28.109 law. He was a co -editor of Privacy and Personal 00:00:28.109 --> 00:00:30.789 Data Protection Law in Asia by Hart Publishing, 00:00:30.910 --> 00:00:33.530 and contributed to the Cambridge Handbook of 00:00:33.530 --> 00:00:36.369 Private Law and Artificial Intelligence. He also 00:00:36.369 --> 00:00:38.570 has extensive practice experience in international 00:00:38.570 --> 00:00:41.750 technology, commercial and energy disputes. Good 00:00:41.750 --> 00:00:43.409 afternoon, Adrian, and welcome to the podcast. 00:00:44.250 --> 00:00:46.090 Good afternoon, Isaac. Thanks very much for having 00:00:46.090 --> 00:00:51.130 me. Okay, so just to get started. For our listeners 00:00:51.130 --> 00:00:53.429 who do not know much about data protection and 00:00:53.429 --> 00:00:56.289 privacy, can you give a brief overview of data 00:00:56.289 --> 00:00:58.890 protection principles and the legal basis in 00:00:58.890 --> 00:01:01.109 which companies and individuals can interact 00:01:01.109 --> 00:01:05.010 with personal data? Yeah, thanks, Isaac. So let 00:01:05.010 --> 00:01:09.189 me set the framing a little bit. So last week, 00:01:09.209 --> 00:01:11.689 I was at the India AI Impact Summit. So it's 00:01:11.689 --> 00:01:14.849 one of the largest AI summits, something apparently 00:01:14.849 --> 00:01:17.750 300 ,000 people went. But one of the panels that 00:01:17.750 --> 00:01:21.430 I went to, it was about data privacy and data 00:01:21.430 --> 00:01:25.099 transparency in the age of AI, right? It doesn't 00:01:25.099 --> 00:01:26.980 seem obvious, but there's a lot of underlying 00:01:26.980 --> 00:01:30.019 tensions in how we think about data privacy and 00:01:30.019 --> 00:01:33.540 also these other competing rights or other demands 00:01:33.540 --> 00:01:36.500 that we have in the age of AI. So a good framework 00:01:36.500 --> 00:01:40.099 that I like to have is by Professor Helen Nissenbaum 00:01:40.099 --> 00:01:42.780 called contextual integrity. It's about not so 00:01:42.780 --> 00:01:45.260 much how we control information and data privacy, 00:01:45.379 --> 00:01:48.939 but sort of underlying norms in data privacy. 00:01:49.480 --> 00:01:52.739 So what she's saying is it's about the norms 00:01:52.739 --> 00:01:56.900 that governs the data flows and the norms in 00:01:56.900 --> 00:02:00.739 the original context of the data, right? So the 00:02:00.739 --> 00:02:03.200 main theme I want to highlight is a lot of these 00:02:03.200 --> 00:02:06.719 norms are changing. A lot of it is subject to 00:02:06.719 --> 00:02:09.400 debate, whether from AI, a lot of the changing 00:02:09.400 --> 00:02:12.199 geopolitical tensions that we're seeing. But 00:02:12.199 --> 00:02:15.939 it's also when we talk about GDPR and different 00:02:15.939 --> 00:02:18.759 frameworks, they are legal tools or attempts 00:02:18.759 --> 00:02:21.819 to encode these norms that we have. And it's 00:02:21.819 --> 00:02:24.159 interesting to see how the legal and these underlying 00:02:24.159 --> 00:02:26.240 norms are changing and having this discussion. 00:02:27.139 --> 00:02:29.819 But that's the macro framing. And I think if 00:02:29.819 --> 00:02:31.919 we go down to the details, let's look at like. 00:02:32.159 --> 00:02:35.439 personal data, right? It's a broad definition and 00:02:35.439 --> 00:02:39.439 would be any information that relates to an identified 00:02:39.439 --> 00:02:42.860 or identifiable natural person. What does that 00:02:42.860 --> 00:02:45.280 mean? What are some of the principles? I think 00:02:45.280 --> 00:02:50.039 an easy way is you look at seven different principles 00:02:50.039 --> 00:02:53.080 and the GDPR, they're in Article 5, but I think 00:02:53.080 --> 00:02:57.469 some of the main ones include It's not so much 00:02:57.469 --> 00:03:01.990 a list of litany of principles, but a hierarchy. 00:03:02.289 --> 00:03:05.710 I think it's quite helpful. So you have lawful 00:03:05.710 --> 00:03:09.490 basis. So has the data been collected lawfully? 00:03:09.610 --> 00:03:13.870 And then you have things like... data minimisation 00:03:13.870 --> 00:03:15.889 and purpose limitation. Even if you have it, 00:03:15.930 --> 00:03:17.349 you can't just get everything and everything 00:03:17.349 --> 00:03:19.870 you want, right? And then you have these things 00:03:19.870 --> 00:03:24.509 like quality control, like is it accurate? Is 00:03:24.509 --> 00:03:26.650 it being stored in the right place? And then 00:03:26.650 --> 00:03:29.050 you have further on top these nice things like 00:03:29.050 --> 00:03:33.669 just whether it's integrity and whether it's 00:03:33.669 --> 00:03:37.229 being held accountable. But as I'm sure we'll 00:03:37.229 --> 00:03:39.469 discuss later, a lot of these principles will... 00:03:39.689 --> 00:03:41.949 are being challenged in the age of AI so i just 00:03:41.949 --> 00:03:44.930 want to flag that very briefly but also when 00:03:44.930 --> 00:03:47.810 you look at some of the basis for processing 00:03:47.810 --> 00:03:50.710 these personal data that there's six main ones 00:03:50.710 --> 00:03:52.849 but i i really just want to highlight um three 00:03:52.849 --> 00:03:56.449 of them they're in article uh six of GDPR people 00:03:56.449 --> 00:03:58.009 can take a look i don't want to go into too much 00:03:58.009 --> 00:04:01.620 details but um there's things like have you consented 00:04:01.620 --> 00:04:04.939 to the collection of that personal data? It's 00:04:04.939 --> 00:04:08.400 things like, is there a contractual or business 00:04:08.400 --> 00:04:12.000 necessity, right? And then the main thing of 00:04:12.000 --> 00:04:14.360 interest for AI is legitimate interests, right? 00:04:14.500 --> 00:04:17.439 Like you look at what's the interest in collecting 00:04:17.439 --> 00:04:19.120 this, whether the rights involved in this and 00:04:19.120 --> 00:04:22.420 balancing exercise. But also just very briefly. 00:04:23.079 --> 00:04:25.259 what does this all mean, right? So there's these 00:04:25.259 --> 00:04:27.879 corresponding rights for us as individuals or 00:04:27.879 --> 00:04:32.399 companies to demand things like, let's get a 00:04:32.399 --> 00:04:34.579 set of the data to look at what data, personal 00:04:34.579 --> 00:04:38.040 data that companies have, or can we demand erasure 00:04:38.040 --> 00:04:40.740 of these personal data? Can we rectify incorrect 00:04:40.740 --> 00:04:44.040 information? And a lot of these, again, are interesting 00:04:44.040 --> 00:04:48.279 themes that will come up in the age of AI. Thank 00:04:48.279 --> 00:04:50.100 you. Thank you so much. And I think, yeah, you 00:04:50.100 --> 00:04:52.889 kind of, you know, prefaced my question there 00:04:52.889 --> 00:04:54.610 that actually my next question is going to be 00:04:54.610 --> 00:04:58.370 about how large language models um have affected 00:04:58.370 --> 00:05:00.449 some of the principles you've talked about some 00:05:00.449 --> 00:05:02.250 you know you said they've gone in a hierarchy 00:05:02.250 --> 00:05:06.009 so is there is it okay if it affects the principles 00:05:06.009 --> 00:05:08.350 at a lower hierarchy and then you know is it 00:05:08.350 --> 00:05:10.850 more problematic if it affects principles higher 00:05:10.850 --> 00:05:15.910 up the chain um yeah so there's really a few 00:05:16.720 --> 00:05:18.959 principles in particular that are being affected 00:05:18.959 --> 00:05:23.180 by generative AI in particular. So principles 00:05:23.180 --> 00:05:26.819 such as data minimisation and purpose limitation, 00:05:27.220 --> 00:05:31.899 right? It's principles of whether when we demand 00:05:31.899 --> 00:05:36.339 an accurate um result um of the AI output um 00:05:36.339 --> 00:05:38.399 how accurate does it have to be right there's 00:05:38.399 --> 00:05:40.319 these problems of hallucination there's these 00:05:40.319 --> 00:05:44.000 problems of memory regurgitation um of personal 00:05:44.000 --> 00:05:46.279 data that we see from the outputs of these LLMs 00:05:46.279 --> 00:05:49.819 um so there's various things that that are going 00:05:49.819 --> 00:05:52.459 on i'm happy to delve into these um into more 00:05:52.459 --> 00:05:56.600 detail yeah uh that that would be great you know 00:05:56.600 --> 00:05:58.680 if you could like you know maybe give us a few 00:05:58.680 --> 00:06:02.379 examples and maybe uh elucidate a little bit 00:06:02.379 --> 00:06:06.079 more on how these uh challenges and how they've 00:06:06.079 --> 00:06:09.500 been managed by AI companies and also how maybe 00:06:09.500 --> 00:06:11.980 legislatures and courts have kind of responded 00:06:11.980 --> 00:06:15.360 to these challenges yeah yeah That's a good question. 00:06:15.480 --> 00:06:19.259 So it's an evolving scene. So I like to think 00:06:19.259 --> 00:06:25.360 of AI and large language models as a data supply 00:06:25.360 --> 00:06:29.680 chain, right? And there are many data privacy 00:06:29.680 --> 00:06:35.240 risks involved in each of these phases. So let's 00:06:35.240 --> 00:06:37.699 start with a concrete example, right? So you're 00:06:37.699 --> 00:06:43.420 an Oxford student and you're sharing... You're 00:06:43.420 --> 00:06:46.379 asking ChatGPT in the UK about like this weird 00:06:46.379 --> 00:06:49.040 medical sport condition that you have, right? 00:06:49.240 --> 00:06:52.500 And you're taking photos of it. And this, let's 00:06:52.500 --> 00:06:55.120 say, personal data information is sent to the 00:06:55.120 --> 00:07:00.079 US with some Microsoft Azure server. And let's 00:07:00.079 --> 00:07:02.519 say it's being stored there for training purposes. 00:07:02.699 --> 00:07:05.660 But at the same time, once there's the processing 00:07:05.660 --> 00:07:07.970 of that. question there's the inference output 00:07:07.970 --> 00:07:10.310 stage where it's sent back to the UK and gives 00:07:10.310 --> 00:07:12.910 you like an answer um what you know what the 00:07:12.910 --> 00:07:15.569 medical condition is um this example is quite 00:07:15.569 --> 00:07:19.370 neat so it flags um three different themes that 00:07:19.370 --> 00:07:23.720 are ongoing in data privacy with AI one is the 00:07:23.720 --> 00:07:27.220 training versus the inference. So the training 00:07:27.220 --> 00:07:29.959 of the LLM involves a lot of personal data. 00:07:31.060 --> 00:07:35.319 How much is too much? Is there legitimate interest 00:07:35.319 --> 00:07:37.060 in processing all this personal information? 00:07:37.420 --> 00:07:40.439 There's the output stuff, right? Where you have 00:07:40.439 --> 00:07:43.680 these models, yeah, hallucinating and giving 00:07:43.680 --> 00:07:48.439 or regurgitating very specific personal data 00:07:48.439 --> 00:07:51.600 that it was being trained on. So that's one big 00:07:51.600 --> 00:07:53.339 theme, and there are different privacy risks 00:07:53.339 --> 00:07:58.079 involved in these two sectors. And then a second 00:07:58.079 --> 00:08:00.899 bucket from this example is the sort of international 00:08:00.899 --> 00:08:04.439 versus local cross -border data flow versus data 00:08:04.439 --> 00:08:08.819 sovereignty theme where, you know, when you have 00:08:08.819 --> 00:08:11.879 these, for example, the UK's data privacy regime 00:08:11.879 --> 00:08:14.579 being applied here, but also there's the US stuff, 00:08:14.639 --> 00:08:17.050 and then there's... different servers everywhere, 00:08:17.269 --> 00:08:20.970 which data protection regime applies and what 00:08:20.970 --> 00:08:23.149 happens if there are conflicting data protection 00:08:23.149 --> 00:08:25.569 regimes and how do you resolve these conflicts? 00:08:26.050 --> 00:08:30.910 But also that's why many jurisdictions are trying 00:08:30.910 --> 00:08:33.909 to assert things like data sovereignty and AI 00:08:33.909 --> 00:08:37.009 sovereignty, whether that's having data localisation 00:08:37.009 --> 00:08:39.669 rules, you need to have your full stack of AI 00:08:39.669 --> 00:08:44.909 from the chips. being more open source and um 00:08:44.909 --> 00:08:48.590 to running on data centers that are being located 00:08:48.590 --> 00:08:51.750 in this specific jurisdiction um to even things 00:08:51.750 --> 00:08:55.250 like you know data centers and um critical minerals 00:08:55.250 --> 00:08:57.929 and all that um but but that's sort of a second 00:08:57.929 --> 00:08:59.889 bucket and i think the third bucket when it comes 00:08:59.889 --> 00:09:05.090 to privacy is this sort of um the the cloud versus 00:09:05.090 --> 00:09:10.230 the on device so um a lot of companies are pushing 00:09:10.230 --> 00:09:13.580 for things like oh like a big proponent would 00:09:13.580 --> 00:09:16.419 be Apple right so a lot of the times they would 00:09:16.419 --> 00:09:21.100 say this uh AI inference all runs on your machine 00:09:21.100 --> 00:09:26.340 and it doesn't go out to other data centers and 00:09:26.340 --> 00:09:30.019 the well there's some nuance to this like often 00:09:30.019 --> 00:09:31.919 when they say that actually it's just sometimes 00:09:31.919 --> 00:09:34.440 it's encrypted secure channel um it's sometimes 00:09:34.440 --> 00:09:37.659 it leaves your phone but um the the intuition 00:09:37.659 --> 00:09:40.360 is the the more it's on your device the less 00:09:40.360 --> 00:09:43.539 likely there are security risks although um or 00:09:43.539 --> 00:09:45.480 personal data risks although we can still um 00:09:45.480 --> 00:09:50.200 talk about that um but yeah um i mean i i like 00:09:50.200 --> 00:09:54.460 to think about the more interesting stories um 00:09:54.460 --> 00:09:58.120 when it comes to like the the training and the 00:09:58.120 --> 00:10:01.860 inference side of things um in the training side 00:10:01.860 --> 00:10:04.159 of things there's i mean the training corpus 00:10:04.159 --> 00:10:08.320 is huge right so um like a Common Crawl set where 00:10:08.320 --> 00:10:12.019 AI is trained on um it's like 250 billion pages 00:10:12.019 --> 00:10:15.259 um another common set is like LAION 5b which is 00:10:15.259 --> 00:10:18.389 like 5 billion pairs of text -to -image pairs 00:10:18.389 --> 00:10:21.490 that AI is trained on. And a lot of this contains 00:10:21.490 --> 00:10:24.149 super sensitive personal data of you and me, 00:10:24.210 --> 00:10:28.470 right? And this involves questions of, okay, 00:10:28.549 --> 00:10:30.809 what's the legitimate basis of processing these? 00:10:30.970 --> 00:10:33.450 It's not consent when we post it initially on 00:10:33.450 --> 00:10:36.370 the internet. It's not necessary business need. 00:10:36.570 --> 00:10:38.690 And so what is the legitimate interest when billions 00:10:38.690 --> 00:10:42.720 of people didn't really say, oh, I... I didn't 00:10:42.720 --> 00:10:44.659 have a reasonable expectation that this would 00:10:44.659 --> 00:10:46.759 be used to train on AI in the first place, right? 00:10:47.159 --> 00:10:50.500 So this is sort of one interesting area and there 00:10:50.500 --> 00:10:53.039 are techniques to resolve that. But the other 00:10:53.039 --> 00:10:56.899 way more fun stuff is in the output side. So 00:10:56.899 --> 00:11:00.139 I believe we chatted about this before in Japan, 00:11:00.399 --> 00:11:05.299 Isaac, but the sort of when you look at the... 00:11:07.210 --> 00:11:10.669 Stability and Getty Images and um Stability and 00:11:10.669 --> 00:11:12.929 The New York Times these are copyright cases 00:11:12.929 --> 00:11:15.450 but they involve similar concerns which is uh 00:11:15.450 --> 00:11:19.269 when you prompt the AI to say generate a Caucasian 00:11:19.269 --> 00:11:23.009 woman in a red dress and pearl necklace literally 00:11:23.009 --> 00:11:25.129 the output is the same as the training sample 00:11:25.129 --> 00:11:27.529 which is very limited sample right and if you 00:11:27.529 --> 00:11:30.230 prompt the AI to to say a theme related to the New 00:11:30.230 --> 00:11:35.289 York Times paper, it gives a very much verbatim 00:11:35.289 --> 00:11:37.649 answer as the training sample. And you look at 00:11:37.649 --> 00:11:40.190 the risk when it comes to personal data. So there's 00:11:40.190 --> 00:11:42.870 these papers that show when you prompt, I think 00:11:42.870 --> 00:11:49.169 it was GPT -2, where you ask GPT to say, repeat 00:11:49.169 --> 00:11:52.230 the word company or poem infinite times. And 00:11:52.230 --> 00:11:54.590 it does this for like... 20, 30 times and then 00:11:54.590 --> 00:11:56.549 it starts regurgitating super sensitive personal 00:11:56.549 --> 00:11:59.149 data of like a company litigation that it was 00:11:59.149 --> 00:12:02.730 trained on, right? And we find actually there's 00:12:02.730 --> 00:12:05.629 these papers that show even for the larger models, 00:12:05.870 --> 00:12:08.210 this regurgitation memorisation still occurs 00:12:08.210 --> 00:12:12.690 and actually the sort of more knowledge intensive 00:12:12.690 --> 00:12:15.929 tasks, it does have this risk more versus like 00:12:15.929 --> 00:12:19.500 if it's very reasoning intensive tasks. um so 00:12:19.500 --> 00:12:21.100 that's super interesting and then there's the 00:12:21.100 --> 00:12:24.620 there's another bucket where there's a hallucination 00:12:24.620 --> 00:12:28.240 example as well so um there's these cases like 00:12:28.240 --> 00:12:35.460 um Starbucks uh versus um Google and um not he's 00:12:35.460 --> 00:12:37.659 he's an activist he's not i don't think he's 00:12:37.659 --> 00:12:40.659 related to Starbucks the company but when you 00:12:40.659 --> 00:12:43.279 search his name um it would come up with things 00:12:43.279 --> 00:12:47.950 like um he's uh he's the like literally child 00:12:47.950 --> 00:12:52.029 rapist and serial abuser and like he's uh like 00:12:52.029 --> 00:12:54.529 offender like multiple things on Gemini right 00:12:54.529 --> 00:12:58.830 and um he has another case where it's against 00:12:58.830 --> 00:13:02.330 Meta Meta's ai where you know he's he's accused 00:13:02.330 --> 00:13:04.970 of being in these riots but none of this is true 00:13:04.970 --> 00:13:09.549 right so um the the pretty crazy stuff and you 00:13:09.549 --> 00:13:12.970 know what the one solution was was to whenever 00:13:12.970 --> 00:13:16.559 you search the name um on Gemini um Gemini would 00:13:16.559 --> 00:13:20.379 just block the output. Because, for example, 00:13:20.460 --> 00:13:23.639 it's very expensive to retrain the whole model 00:13:23.639 --> 00:13:25.980 just for one person, even though there's these 00:13:25.980 --> 00:13:30.279 principles of erasure under GDPR. And there's 00:13:30.279 --> 00:13:32.679 these other techniques, like could you do machine 00:13:32.679 --> 00:13:35.360 unlearning, where you just unlearn one part of 00:13:35.360 --> 00:13:38.019 the data, but then there's these other risks 00:13:38.019 --> 00:13:41.679 involved, right? So technically, there are these 00:13:41.679 --> 00:13:45.070 super interesting advancements and maybe the 00:13:45.070 --> 00:13:47.490 legal rights is one direction to tackle it and 00:13:47.490 --> 00:13:50.389 maybe the sort of technical solution is one way 00:13:50.389 --> 00:13:53.450 to tackle it but there's there's many aspects 00:13:53.450 --> 00:13:56.590 to this that are very interesting a lot of privacy 00:13:56.590 --> 00:13:59.889 personal data risks are involved and i think 00:13:59.889 --> 00:14:03.179 um This is a long answer, but I think just the 00:14:03.179 --> 00:14:06.980 third thing to flag, apart from the input and 00:14:06.980 --> 00:14:08.899 the output side, is the infrastructure around 00:14:08.899 --> 00:14:13.200 AI. So we're in an age of agentic AI. It's not 00:14:13.200 --> 00:14:16.360 just the LLM, which is just the next token prediction, 00:14:16.500 --> 00:14:19.700 but the architecture. So when you use agents 00:14:19.700 --> 00:14:23.200 that have access to email. I'm giving it access 00:14:23.200 --> 00:14:27.059 to send replies on my behalf. A lot of this increased 00:14:27.059 --> 00:14:30.019 the surface area for attack and increasing traditional 00:14:30.019 --> 00:14:35.139 risks for personal data and data infringement 00:14:35.139 --> 00:14:38.539 risks. but also new categories of data privacy 00:14:38.539 --> 00:14:41.820 risks like prompt injection and enlisting new 00:14:41.820 --> 00:14:44.799 sorts of personal data from your AI agents. So 00:14:44.799 --> 00:14:50.980 there are lots of developing things, but it's 00:14:50.980 --> 00:14:54.360 helpful, I think, to think of the risks in terms 00:14:54.360 --> 00:14:57.379 of this AI data supply chain and what are the 00:14:57.379 --> 00:15:01.610 different vectors that map across it. yeah i 00:15:01.610 --> 00:15:02.990 actually just want to bounce off some of the 00:15:02.990 --> 00:15:04.950 things you talked about earlier i think i was 00:15:04.950 --> 00:15:07.710 listening recently to a Financial Times podcast 00:15:07.710 --> 00:15:10.649 and they were talking about how uh you know they've 00:15:10.649 --> 00:15:13.090 got some sources at OpenAI that's saying that 00:15:13.090 --> 00:15:16.769 OpenAI is now switching from you know long -term 00:15:16.769 --> 00:15:19.450 development like AI research to commercialising 00:15:19.450 --> 00:15:23.289 their main product which is ChatGpt and do you 00:15:23.289 --> 00:15:25.129 think like about all the issues you mentioned 00:15:25.129 --> 00:15:28.129 earlier you know do you think that there's obviously 00:15:28.129 --> 00:15:30.370 a technical solution which would involve a lot 00:15:30.370 --> 00:15:33.769 more long -term deep research into how the AI 00:15:33.769 --> 00:15:37.090 models work do you think this is something that 00:15:37.090 --> 00:15:39.450 can be kind of controlled by legislation in a 00:15:39.450 --> 00:15:41.610 way that both so both the legal solution and 00:15:41.610 --> 00:15:44.330 both the technical solution kind of work together 00:15:44.330 --> 00:15:47.169 such in a way to put pressure on these AI companies 00:15:47.169 --> 00:15:50.269 to adopt you know to really think about the the 00:15:50.269 --> 00:15:52.850 technical solution to such problems rather than 00:15:52.850 --> 00:15:55.269 just focus focusing on commercialising their 00:15:55.269 --> 00:15:58.559 product because you know That's what these companies 00:15:58.559 --> 00:16:00.360 want to do. I mean, in particular, I think there's 00:16:00.360 --> 00:16:03.700 been talks about ChatGPT having ads now. So it 00:16:03.700 --> 00:16:06.840 might be possible that if you prompt them about 00:16:06.840 --> 00:16:09.700 something, they'll just give you an ad or they'll 00:16:09.700 --> 00:16:12.240 insert an ad into their response. And is this 00:16:12.240 --> 00:16:14.220 really the right direction AI companies should 00:16:14.220 --> 00:16:16.019 be going? Or do you think there could be some 00:16:16.019 --> 00:16:18.539 form of legal pressure put on them to start really 00:16:18.539 --> 00:16:22.059 looking at these models and continually developing 00:16:22.059 --> 00:16:27.340 them to avoid such risk? Yeah, there's the famous 00:16:27.340 --> 00:16:31.740 Super Bowl ad run by Anthropic, where they basically 00:16:31.740 --> 00:16:35.740 do a parody of what the future looks like if 00:16:35.740 --> 00:16:38.820 it was run by OpenAI. And OpenAI gives you, you're 00:16:38.820 --> 00:16:41.360 asking very sensitive, highly important information, 00:16:41.360 --> 00:16:44.039 and then suddenly AI gives you an ad in the middle, 00:16:44.100 --> 00:16:46.379 right? And maybe that's the last thing you want 00:16:46.379 --> 00:16:53.659 in an AI. look not all AI companies are the same 00:16:53.659 --> 00:16:55.240 and they have different business models right 00:16:55.240 --> 00:16:58.399 like so traditionally you would think of um Anthropic 00:16:58.399 --> 00:17:01.679 as be like Claude looking at much more of B2B 00:17:01.679 --> 00:17:04.519 and their business model is slightly different 00:17:04.519 --> 00:17:07.240 from OpenAI which is traditionally seen as more 00:17:07.240 --> 00:17:11.259 B2C and sort of the ecosystem of OpenAI why 00:17:11.259 --> 00:17:14.000 they had to double click on Sora video generation 00:17:14.000 --> 00:17:18.089 like a social media app is Because a lot of the 00:17:18.089 --> 00:17:20.869 concerns around revenue and whether the current 00:17:20.869 --> 00:17:23.369 trajectory of investing hundreds of billions 00:17:23.369 --> 00:17:26.990 of dollars in data center and energy is that 00:17:26.990 --> 00:17:29.349 sustainable, right? And different companies are 00:17:29.349 --> 00:17:32.410 really trying to figure out what is the way to 00:17:32.410 --> 00:17:34.970 monetise and keep this sustainable, right? If 00:17:34.970 --> 00:17:38.990 we are to have a leading edge. So I think the 00:17:38.990 --> 00:17:42.990 business model question... is a big one and how 00:17:42.990 --> 00:17:47.529 that ties into safety research and um data privacy 00:17:47.529 --> 00:17:51.529 risks is is a big one right so so these these 00:17:51.529 --> 00:17:54.289 companies have for example like OpenAI has like 00:17:54.289 --> 00:17:57.039 a model behaviour team they have like like psychology 00:17:57.039 --> 00:18:00.819 psychologist looking into how different personalities 00:18:00.819 --> 00:18:04.500 of AI would react differently in different safety 00:18:04.500 --> 00:18:08.180 scenarios. Anthropic is very big on the constitutional 00:18:08.180 --> 00:18:11.279 AI and they released a new one where they have 00:18:11.279 --> 00:18:16.180 different ways of collecting values from different 00:18:16.180 --> 00:18:19.039 folks and trying to see how to embed that into 00:18:19.039 --> 00:18:22.019 the way that both the training and also how the 00:18:22.019 --> 00:18:24.539 answers that AI gives are different. And I think... 00:18:24.920 --> 00:18:27.960 You're right in pointing out that there are these 00:18:27.960 --> 00:18:31.799 concerns with AI companies. But there is also 00:18:31.799 --> 00:18:36.019 this flip dynamic where we're talking largely 00:18:36.019 --> 00:18:38.640 about closed source AI, but there's this movement 00:18:38.640 --> 00:18:41.559 with open source and a lot of the leading open 00:18:41.559 --> 00:18:44.740 source companies, particularly coming from China, 00:18:44.900 --> 00:18:50.319 which means they're technically open weight. You're 00:18:50.319 --> 00:18:52.460 you're allowing people you're releasing into 00:18:52.460 --> 00:18:54.299 the public and you're allowing people to fine 00:18:54.299 --> 00:18:57.400 tune it. And a lot of them do release their specs 00:18:57.400 --> 00:19:00.140 on how they train it. And there's all sorts of 00:19:00.140 --> 00:19:03.660 regulations around product specs and documentation. 00:19:03.799 --> 00:19:07.319 So this is one way where we encourage more transparency 00:19:07.319 --> 00:19:11.200 around the ecosystem and a healthier debate around 00:19:11.200 --> 00:19:15.380 what should be done. I think the legislation 00:19:15.380 --> 00:19:20.099 question is is. interesting one that different 00:19:20.099 --> 00:19:24.400 jurisdictions are grappling with, right? So there 00:19:24.400 --> 00:19:29.460 is the question of, do we do the whole Brussels 00:19:29.460 --> 00:19:35.559 effect with GDPR and try to out -innovate and 00:19:35.559 --> 00:19:39.880 lead by regulation? One of the concerns, both 00:19:39.880 --> 00:19:42.799 externally outside of EU, but also within EU 00:19:42.799 --> 00:19:47.210 with things like the Draghi report is whether 00:19:47.210 --> 00:19:50.890 there is too much regulation within the EU and 00:19:50.890 --> 00:19:52.990 whether that hinders innovation, especially when 00:19:52.990 --> 00:19:55.009 it comes to compliance costs for small and medium 00:19:55.009 --> 00:19:57.730 enterprises. If you look at some of the obligations 00:19:57.730 --> 00:20:02.630 for, not for OpenAI or the likes of those foundation 00:20:02.630 --> 00:20:05.130 model companies, but for small companies who 00:20:05.130 --> 00:20:07.130 are just trying to make something interesting, 00:20:07.369 --> 00:20:09.670 the compliance costs are super high. Now, of 00:20:09.670 --> 00:20:12.089 course, there is that conception. And then there's 00:20:12.089 --> 00:20:14.210 also the counter argument that you know, we're 00:20:14.210 --> 00:20:16.470 thinking about regulation as a hindrance, but 00:20:16.470 --> 00:20:20.890 actually regulation is in many instances an enabler 00:20:20.890 --> 00:20:25.829 of innovation, right? So think about the aviation 00:20:25.829 --> 00:20:29.730 sector, where it's because we demand these high 00:20:29.730 --> 00:20:32.829 standards for aviation and we establish things, 00:20:32.990 --> 00:20:34.609 international standards, but also like federal 00:20:34.609 --> 00:20:38.769 aviation administration, that we have such great 00:20:38.769 --> 00:20:44.599 technologies behind aviation. And, you know, 00:20:44.640 --> 00:20:46.539 there's similar arguments going on, like how 00:20:46.539 --> 00:20:50.880 much of the regulation can we push to have to 00:20:50.880 --> 00:20:52.759 encourage these companies to have much safer 00:20:52.759 --> 00:20:55.519 and also taking care of much more of our concerns. 00:20:56.420 --> 00:20:59.660 And you see something similar in, you know, GDPR, 00:20:59.680 --> 00:21:02.480 I think Article 25, like these mandates for privacy 00:21:02.480 --> 00:21:05.400 by design. So it's a legal obligation, but it 00:21:05.400 --> 00:21:10.819 says, look, let's push companies to embed privacy 00:21:10.819 --> 00:21:12.960 principles when they're designing some of them. 00:21:12.880 --> 00:21:15.119 the tools, including AI tools that they have, 00:21:15.319 --> 00:21:19.119 right? So that's why people develop technologies 00:21:19.119 --> 00:21:23.180 like differential privacy, where you try to add 00:21:23.180 --> 00:21:27.960 random noise into a sample when training AI. 00:21:28.079 --> 00:21:30.299 So you don't really know whether any individual 00:21:30.299 --> 00:21:33.680 is in that sample or not. And there's other technologies 00:21:33.680 --> 00:21:37.599 like Things like federated learning where you're 00:21:37.599 --> 00:21:41.079 training the model locally and then transferring 00:21:41.079 --> 00:21:43.960 the weights outside to another place. So there's 00:21:43.960 --> 00:21:47.619 less risk of transferring personal data or things 00:21:47.619 --> 00:21:51.059 like machine learning. So we are entering into 00:21:51.059 --> 00:21:54.680 a phase like you hinted, which is we need the 00:21:54.680 --> 00:21:58.160 technical governance and we need the legal governance, 00:21:58.279 --> 00:22:00.140 but we also need normative governance. We need 00:22:00.140 --> 00:22:02.740 people working together because it is such a 00:22:02.740 --> 00:22:07.339 cross. cross interdisciplinary foundational general 00:22:07.339 --> 00:22:11.759 purpose technology right thank you for that Adrian 00:22:11.759 --> 00:22:13.779 uh i think you mentioned a little bit about like 00:22:13.779 --> 00:22:16.400 different jurisdictions and how they manage it 00:22:16.400 --> 00:22:19.420 uh if you don't mind maybe could you do a little 00:22:19.420 --> 00:22:20.880 bit of a comparison i think you've done a little 00:22:20.880 --> 00:22:24.759 bit between the GDPR and uh maybe and maybe the 00:22:24.759 --> 00:22:27.400 UK and the US but you know maybe if we're looking 00:22:27.400 --> 00:22:29.680 towards Asia especially because China is a huge 00:22:29.680 --> 00:22:33.589 AI jurisdiction maybe Can we compare across these 00:22:33.589 --> 00:22:35.650 jurisdictions and see, and maybe you can give 00:22:35.650 --> 00:22:38.250 us your opinion on what you think, you know, 00:22:38.250 --> 00:22:40.509 the right things they are doing and maybe, you 00:22:40.509 --> 00:22:43.849 know, what do you think could be improved? Yeah, 00:22:43.849 --> 00:22:54.619 that's a great question. So there's a paper written 00:22:54.619 --> 00:22:59.859 by my friend at Stanford. It's called Comparing 00:22:59.859 --> 00:23:02.660 Apples to Oranges. And it talks about how when 00:23:02.660 --> 00:23:04.980 we talk about the different AI regulations, actually, 00:23:04.980 --> 00:23:10.420 it's a wide spectrum. And some regulations focus 00:23:10.420 --> 00:23:15.079 on risk. Some focus on the technology itself. 00:23:15.900 --> 00:23:19.779 That's one. uh difference right um another way 00:23:19.779 --> 00:23:22.799 of thinking about it is uh some focus on ex ante 00:23:22.799 --> 00:23:26.480 so um so before the fact how we have these regulations 00:23:26.480 --> 00:23:28.920 and some focus on expose where it's much more 00:23:28.920 --> 00:23:31.180 after the fact let's have litigation and try 00:23:31.180 --> 00:23:33.480 to regulate this stuff and then there's some 00:23:33.480 --> 00:23:38.359 that focus on sort of a more general uh wide 00:23:38.359 --> 00:23:42.380 ranging act that tries to be comprehensive and 00:23:42.380 --> 00:23:44.779 then there's some that try to be more sectoral. 00:23:44.779 --> 00:23:47.700 Let's look at sectors like biometrics, health, 00:23:47.940 --> 00:23:50.619 financial services, and then how AI is being 00:23:50.619 --> 00:23:53.019 applied in these sectors. And then how, let's 00:23:53.019 --> 00:23:56.099 look at the specific instances. So the EU AI 00:23:56.099 --> 00:24:01.859 Act, I think it'd be helpful to give some backstory, 00:24:02.000 --> 00:24:06.539 which is after the GDPR, which is late 2010s. 00:24:07.740 --> 00:24:10.259 um the working commission they came together 00:24:10.259 --> 00:24:12.740 and they said let's let's try to regulate let's 00:24:12.740 --> 00:24:15.519 look at the AI and how we can look at specific 00:24:15.519 --> 00:24:20.460 risks uh from using of AI and this is early 2020s 00:24:20.460 --> 00:24:24.119 and they you know basically looked at these sort 00:24:24.119 --> 00:24:28.019 of uh banned banned categories because of the 00:24:28.019 --> 00:24:29.940 high risk because of the risk and then there's 00:24:29.940 --> 00:24:32.099 these high risk or minimal risk and limited risks 00:24:32.099 --> 00:24:35.640 um similar to what you see now in the AI act 00:24:35.640 --> 00:24:39.890 and then They pretty much wrapped. And then in 00:24:39.890 --> 00:24:43.809 late 2022, November, this thing called ChatGPT 00:24:43.809 --> 00:24:48.150 came out or GPT 3 .5. And people were astounded, 00:24:48.190 --> 00:24:50.769 right? Like this wasn't really the type of technology 00:24:50.769 --> 00:24:53.849 that the EU was looking. And then so they sort 00:24:53.849 --> 00:24:58.170 of had to... um scramble and giving a very high 00:24:58.170 --> 00:25:00.170 level storytelling manner of how the process 00:25:00.170 --> 00:25:04.410 went uh they they try to look into also the regulation 00:25:04.410 --> 00:25:07.269 the technology itself so they were looking into 00:25:07.269 --> 00:25:10.450 some of the um foundation models categorising 00:25:10.450 --> 00:25:14.009 into high risk and um general purpose and whatnot 00:25:14.009 --> 00:25:18.089 and um so you have this sort of behemoth of a 00:25:18.089 --> 00:25:21.359 regulation and captured in the EU AI Act And then 00:25:21.359 --> 00:25:24.500 that's sort of meant to be complementary to the 00:25:24.500 --> 00:25:26.940 GDPR in many ways. But there's also many tensions 00:25:26.940 --> 00:25:31.880 within the EU AI act and GDPR, like the things 00:25:31.880 --> 00:25:33.660 that we discussed, right? So purpose limitation, 00:25:34.000 --> 00:25:37.559 data minimisation. Yes, we want to limit that. 00:25:37.640 --> 00:25:40.619 But then in the age of AI, to train these models, 00:25:40.700 --> 00:25:44.460 we do want like larger, diverse data sets to 00:25:44.460 --> 00:25:49.079 train these models. access to sensitive, high 00:25:49.079 --> 00:25:51.279 -risk categories of data in order to at least 00:25:51.279 --> 00:25:54.119 know whether there is bias in the AI, right? 00:25:55.000 --> 00:26:00.799 So these are types of tensions, and recognising 00:26:00.799 --> 00:26:03.940 these tensions and also the huge advancements 00:26:03.940 --> 00:26:07.140 that LLMs are making. Last year in November, 00:26:07.440 --> 00:26:10.140 so just a few months ago, the commission came 00:26:10.140 --> 00:26:13.720 out with... digital omnibus package. So it tries 00:26:13.720 --> 00:26:15.440 to reconcile some of the differences between 00:26:15.440 --> 00:26:18.539 the GDPR and the AI Act. So looking at legitimate 00:26:18.539 --> 00:26:21.680 interests and can that be done for AI? Can we 00:26:21.680 --> 00:26:24.700 delay some of the high risk obligations under 00:26:24.700 --> 00:26:29.279 AI Act? So that's sort of a moving picture of 00:26:29.279 --> 00:26:31.799 AI. And there's a big question of how much that 00:26:31.799 --> 00:26:34.660 model has been exported to, say, China, say, 00:26:34.779 --> 00:26:42.049 the US. In the US, it's also very a scattered 00:26:42.049 --> 00:26:44.509 approach right now, right? So federal level, 00:26:44.670 --> 00:26:46.990 there's always been discussions of whether there 00:26:46.990 --> 00:26:50.390 should be an APRA, so American Privacy Rights 00:26:50.390 --> 00:26:54.029 Act at the federal level, which hasn't been passed. 00:26:54.750 --> 00:27:00.349 But there's a lot of sort of state level privacy 00:27:00.349 --> 00:27:02.910 acts, but also a lot of sectoral approaches. 00:27:03.630 --> 00:27:07.549 um like in like HIPA like VIPA which in the 00:27:07.549 --> 00:27:10.410 US which deal with sort of health and um privacy 00:27:10.410 --> 00:27:14.170 rights for uh schools that are funded uh publicly 00:27:14.170 --> 00:27:17.910 and then sort of you know these um there's these 00:27:17.910 --> 00:27:22.109 things where there's these acts um that are coming 00:27:22.109 --> 00:27:25.259 out specifically for AI that sort of tangentially 00:27:25.259 --> 00:27:27.039 relate to privacy, but actually don't deal with 00:27:27.039 --> 00:27:30.460 it directly. So Colorado has a huge AI act that's 00:27:30.460 --> 00:27:34.660 comprehensive in nature. Tennessee. This is interesting. 00:27:34.859 --> 00:27:37.880 Tennessee has one called the Elvis Act, which 00:27:37.880 --> 00:27:41.500 was in 2024, which relates to AI deepfakes for 00:27:41.500 --> 00:27:45.319 singers. They don't want singers to be cloned 00:27:45.319 --> 00:27:48.680 with AI, right? So that's sort of related to 00:27:48.680 --> 00:27:52.420 some of the concerns in data privacy. And then 00:27:52.420 --> 00:27:57.670 there's some existing... um legislations like 00:27:57.670 --> 00:28:01.089 the Wiretapping Act um but that relates to things 00:28:01.089 --> 00:28:02.970 like on Zoom like in this conversation do you 00:28:02.970 --> 00:28:06.119 need one party to consent to the recording or 00:28:06.119 --> 00:28:07.779 do you need to put both parties in different 00:28:07.779 --> 00:28:09.960 states of different approaches? So it's very 00:28:09.960 --> 00:28:13.740 scattered, different approaches. And again, there's 00:28:13.740 --> 00:28:15.519 discussion of like, how much should we try to 00:28:15.519 --> 00:28:18.180 do it on the federal level? With the current 00:28:18.180 --> 00:28:21.140 administration, should we have no state level 00:28:21.140 --> 00:28:24.460 regulations because we want a uniform compliance 00:28:24.460 --> 00:28:28.359 standard in the US? So that's sort of changing 00:28:28.359 --> 00:28:30.900 as well. And then I think, as you rightly pointed 00:28:30.900 --> 00:28:35.740 out, in certain text circles, people do think 00:28:35.740 --> 00:28:40.240 like at least with the Chinese approach, they 00:28:40.240 --> 00:28:44.589 are super specific. applications when it comes 00:28:44.589 --> 00:28:47.849 to privacy, when it comes to AI. There's some 00:28:47.849 --> 00:28:50.150 super specific legislations like, oh, let's look 00:28:50.150 --> 00:28:53.109 at watermarking and you have to watermark AI 00:28:53.109 --> 00:28:55.869 outputs. There's some super specific ones related 00:28:55.869 --> 00:29:00.309 to algorithm, recommender algorithms. And I think 00:29:00.309 --> 00:29:04.829 what you'll find in the current state of jurisdictions 00:29:04.829 --> 00:29:08.619 in different including Asian jurisdictions, is 00:29:08.619 --> 00:29:11.819 they tend to do a bit of both. In the legislative 00:29:11.819 --> 00:29:14.779 process, they look at some of the Chinese style, 00:29:14.900 --> 00:29:16.640 they look at some of the EU style, some of the 00:29:16.640 --> 00:29:19.319 US style, and they have a bit of everything, 00:29:19.460 --> 00:29:22.759 right? So a lot of them tend not to have a super 00:29:22.759 --> 00:29:25.759 comprehensive one like the EU AI Act because... 00:29:26.329 --> 00:29:28.910 For example, they don't have leading model companies 00:29:28.910 --> 00:29:32.930 and there may be a huge compliance cost, but 00:29:32.930 --> 00:29:36.049 let's do some sectoral approaches and regulations. 00:29:36.549 --> 00:29:39.529 And then let's also try to develop technical 00:29:39.529 --> 00:29:42.630 interoperable standards that are more soft and 00:29:42.630 --> 00:29:46.180 non -binding in nature. for example, like the 00:29:46.180 --> 00:29:49.900 Japan equivalent of the AI Act focuses much more 00:29:49.900 --> 00:29:52.900 on innovation and how to encourage investments 00:29:52.900 --> 00:29:57.420 into the AI economy. So it comes in all flavours. 00:29:57.480 --> 00:30:01.839 And yeah, I'd recommend people looking more into 00:30:01.839 --> 00:30:05.259 this changing scene. Thank you. Thank you. And 00:30:05.259 --> 00:30:09.420 I think you also mentioned about like the difficult, 00:30:09.460 --> 00:30:11.119 I think this may be something you talked a little 00:30:11.119 --> 00:30:14.410 bit more about in Japan when we met that. there's 00:30:14.410 --> 00:30:16.509 a difficulty when it comes to regulation simply 00:30:16.509 --> 00:30:19.309 because you know the pace of AI models are developing 00:30:19.309 --> 00:30:21.710 so quickly that you know we're going to have 00:30:21.710 --> 00:30:23.970 you know sometimes the models the the regulations 00:30:23.970 --> 00:30:26.130 and the legislation just can't keep up with the 00:30:26.130 --> 00:30:29.269 new developments in these models the the intrusiveness 00:30:29.269 --> 00:30:31.210 of some of these models actually i think you're 00:30:31.210 --> 00:30:32.789 talking a little bit about transcribing just 00:30:32.789 --> 00:30:35.369 now and just as a little personal anecdote before 00:30:35.369 --> 00:30:39.369 you join the call i use Notion which is a you 00:30:39.369 --> 00:30:42.430 know a note -taking app and Notion AI there's 00:30:42.430 --> 00:30:45.390 a Notion AI now yeah it's great basically offered 00:30:45.390 --> 00:30:47.569 to transcribe the call if i wanted to i mean 00:30:47.569 --> 00:30:50.029 similarly Descript the platform that we're 00:30:50.029 --> 00:30:53.390 currently on uses an AI model to transcribe whatever 00:30:53.390 --> 00:30:55.829 we're saying like i mean i'll plug this in and 00:30:55.829 --> 00:30:57.769 the AI will just transcribe whatever we've said 00:30:57.769 --> 00:30:59.490 we'll just say it and that will create a transcript 00:30:59.490 --> 00:31:01.769 and i don't have to like manually listen to it 00:31:01.769 --> 00:31:04.029 and type it out and and that kind of intrusiveness 00:31:04.029 --> 00:31:06.970 and i guess i was just so surprised that things 00:31:06.970 --> 00:31:09.859 are developing so quickly And I guess in relation 00:31:09.859 --> 00:31:13.539 to that question, do you feel that the best approach 00:31:13.539 --> 00:31:16.579 to regulation is to think a little bit about 00:31:16.579 --> 00:31:20.059 normative theories of law? Like in a sense of 00:31:20.059 --> 00:31:25.059 like, maybe do we use the approach of like going 00:31:25.059 --> 00:31:27.599 on general principles rather than very specific 00:31:27.599 --> 00:31:30.000 legislation that targets a very specific technology 00:31:30.000 --> 00:31:34.079 or very specific area of risk? Instead, should 00:31:34.079 --> 00:31:37.250 we... use i guess because i'm studying jurisprudence 00:31:37.250 --> 00:31:40.690 now i guess use these you know principles of 00:31:40.690 --> 00:31:43.309 ethics that may be that that granted may vary 00:31:43.309 --> 00:31:45.269 depending on the focus of the country i mean 00:31:45.269 --> 00:31:49.250 japan might want to focus on innovation um countries 00:31:49.250 --> 00:31:50.730 in the west might want to focus a little bit 00:31:50.730 --> 00:31:53.170 more on regulation but can we use some general 00:31:53.170 --> 00:31:56.329 normative principles of ethics to influence the 00:31:56.329 --> 00:31:58.109 way we approach legislation and come up with 00:31:58.109 --> 00:32:01.269 a more i guess a more loose legislation but also 00:32:01.269 --> 00:32:03.049 a legislation that's therefore more flexible 00:32:03.049 --> 00:32:07.859 to it adapt to the changes in technology yeah 00:32:07.859 --> 00:32:09.619 that's a great question i i think just to relate 00:32:09.619 --> 00:32:15.579 to your story i mean um in in meetings nowadays 00:32:15.579 --> 00:32:17.980 you always have like these when you join Zoom 00:32:17.980 --> 00:32:20.420 they ask you consent but also a separate note 00:32:20.420 --> 00:32:22.480 taker that joins and it's always it looks like 00:32:22.480 --> 00:32:24.519 a separate person who's joining but it's actually 00:32:24.519 --> 00:32:26.640 just someone who transcribed But the summaries 00:32:26.640 --> 00:32:30.079 are pretty good, right? But I think in this story 00:32:30.079 --> 00:32:31.960 that I'm trying to tell, it's important not to 00:32:31.960 --> 00:32:34.599 focus only on the risk, but also the opportunities. 00:32:35.099 --> 00:32:38.559 I used to work in big law, doing IPOs and M&As. 00:32:38.619 --> 00:32:41.839 And a lot of what junior lawyers work was transcribing, 00:32:41.839 --> 00:32:44.400 typing manually, using hours. And I remember 00:32:44.400 --> 00:32:47.019 one story was like, I mean, just doing this Chinese 00:32:47.019 --> 00:32:50.299 huge AI company, well, just a huge company, but 00:32:50.299 --> 00:32:54.880 transcribing the call. Literally, I had six rounds 00:32:54.880 --> 00:32:57.000 of back and forth just with the senior associate 00:32:57.000 --> 00:32:59.839 just on how the notes should be taken. And the 00:32:59.839 --> 00:33:02.339 associate was great. I learned so much from this 00:33:02.339 --> 00:33:04.839 is a better standard of how to note take. But 00:33:04.839 --> 00:33:08.619 that spends so much of hours and billable hours 00:33:08.619 --> 00:33:11.539 where you could release that time to focus on 00:33:11.539 --> 00:33:14.140 much more valuable work. And that's a narrative 00:33:14.140 --> 00:33:17.279 of how much are we replacing humans and augmenting 00:33:17.279 --> 00:33:23.599 human work. But onto your point about the normative, 00:33:23.680 --> 00:33:28.279 and I think in the way you've defined it, I think 00:33:28.279 --> 00:33:30.799 it's essential. I think it's not just a good 00:33:30.799 --> 00:33:35.059 to have nowadays, right? Even in the example 00:33:35.059 --> 00:33:37.279 that we're talking about, so this makes me think 00:33:37.279 --> 00:33:41.740 of Professor Erik Brynjolfsson from Stanford, 00:33:41.819 --> 00:33:43.559 I think a few years ago, he talked about this 00:33:43.559 --> 00:33:46.859 within the AI circle, people always talk about 00:33:47.710 --> 00:33:50.730 um replacement rather than augmentation of human 00:33:50.730 --> 00:33:53.029 labor right like there's this company i think 00:33:53.029 --> 00:33:56.849 Mechanized Inc that like sort of makes fun and 00:33:56.849 --> 00:33:59.369 tries to say the billboards would say we're here 00:33:59.369 --> 00:34:04.490 to um replace humans right and um like it's that 00:34:04.490 --> 00:34:08.269 like narrative of you know we like we don't need 00:34:08.269 --> 00:34:10.969 human labor but but that's sort of the the paper 00:34:10.969 --> 00:34:13.920 by Erik Brynjolfsson and also my professor Professor 00:34:13.920 --> 00:34:16.440 Rob Reich talks about this, which is even in 00:34:16.440 --> 00:34:21.099 the early 40s and 50s in the CS AI field, there 00:34:21.099 --> 00:34:24.960 would be debates by Alan Turing. The Turing test 00:34:24.960 --> 00:34:29.519 is sort of, in some ways, encourages us to think 00:34:29.519 --> 00:34:31.500 about this replacement because it's how can humans 00:34:31.500 --> 00:34:34.659 replace, how can AI's machines replace human 00:34:34.659 --> 00:34:38.400 and what we're doing? But another framework by 00:34:38.400 --> 00:34:42.579 the founder of Cybernetics in the 40s is how 00:34:42.579 --> 00:34:45.139 can... AI and machines augment what humans are 00:34:45.139 --> 00:34:47.579 doing and complement humans are doing. So even 00:34:47.579 --> 00:34:50.239 in the AI field, where we don't think a lot about 00:34:50.239 --> 00:34:52.820 ethics and normative framework, or traditionally 00:34:52.820 --> 00:34:54.659 we don't think they're that field, there's a 00:34:54.659 --> 00:34:58.659 lot of normative discussion. And the current 00:34:58.659 --> 00:35:01.980 vibe shift now, at least here in Silicon Valley 00:35:01.980 --> 00:35:03.579 at Stanford, is a lot more people are trying 00:35:03.579 --> 00:35:08.400 to do this alignment work. even in the examples 00:35:08.400 --> 00:35:10.380 i mentioned like Anthropic they're trying to 00:35:10.380 --> 00:35:12.260 do a lot of like safety and alignment work which 00:35:12.260 --> 00:35:15.039 is super cool but also in the law and AI field 00:35:15.039 --> 00:35:19.400 right there's a lot more discussions on law following 00:35:19.400 --> 00:35:22.619 AI agents so um previously people would talk 00:35:22.619 --> 00:35:25.900 about how can we get agents to um abide by our 00:35:25.900 --> 00:35:27.780 values and there's a lot of like normative and 00:35:27.780 --> 00:35:31.050 ethical discussions but it's um difficult once 00:35:31.050 --> 00:35:33.230 you think about okay this is super vague but 00:35:33.230 --> 00:35:36.309 it's that's a good start and helpful in how we 00:35:36.309 --> 00:35:38.769 conceptualise AI agents but now people are talking 00:35:38.769 --> 00:35:40.809 about wait let's talk about how they follow the 00:35:40.809 --> 00:35:43.809 law um when AI agents are roaming around and 00:35:43.809 --> 00:35:46.170 then there's questions of which law does it follow 00:35:46.170 --> 00:35:49.110 um so for example there's an Institute of Law 00:35:49.110 --> 00:35:51.190 and AI and they're doing a workshop with Cambridge 00:35:51.190 --> 00:35:54.570 as well just focusing on like how how does it 00:35:54.570 --> 00:35:56.250 follow these values and how does it value these 00:35:56.250 --> 00:36:00.980 follow these laws but yeah, to answer your question, 00:36:01.139 --> 00:36:04.300 I think it's essential to do these normative 00:36:04.300 --> 00:36:06.360 work as well as the practical work. And I always 00:36:06.360 --> 00:36:08.900 think it's the intersection that's the most interesting. 00:36:09.139 --> 00:36:14.630 I think... um one one potential resource um and 00:36:14.630 --> 00:36:17.610 and not to shamelessly plug myself but the sort 00:36:17.610 --> 00:36:19.650 of Cambridge handbook of uh Private Law and AI 00:36:19.650 --> 00:36:23.550 um i contributed to a chapter um but a lot of 00:36:23.550 --> 00:36:26.130 the the chapters were discussing about these 00:36:26.130 --> 00:36:29.010 sort of normative aspects of of uh AI and how 00:36:29.010 --> 00:36:31.869 it relates to the law and i think um yeah i would 00:36:31.869 --> 00:36:35.110 love um that there's that there's more uh there's 00:36:35.110 --> 00:36:36.690 a lot of work on this as well and i'd love for 00:36:36.690 --> 00:36:39.590 you and other people to join in this discussion 00:36:40.780 --> 00:36:43.360 no yeah definitely because i think one of the 00:36:43.360 --> 00:36:46.460 so we have to do a mini thesis for our summer 00:36:46.460 --> 00:36:49.340 over you know over our summer break this time 00:36:49.340 --> 00:36:51.039 around uh moving from the second year to the 00:36:51.039 --> 00:36:53.460 third year and one of the mini options that we 00:36:53.460 --> 00:36:56.179 can write about is on uh law and technology and 00:36:56.179 --> 00:36:58.639 it's and and it's a jurisprudence mini option 00:36:58.639 --> 00:37:00.500 so i i would think that this would actually be 00:37:00.500 --> 00:37:03.260 very helpful because you know we want to think 00:37:03.260 --> 00:37:06.860 about how these theories and these values that 00:37:06.860 --> 00:37:08.340 we've learned in jurisprudence, like traditional 00:37:08.340 --> 00:37:10.619 jurisprudence from people like Hart, Dworkin, 00:37:10.719 --> 00:37:14.099 Finnis, and how we're thinking about such things 00:37:14.099 --> 00:37:16.780 in relation to AI, in relation to the tech, in 00:37:16.780 --> 00:37:19.780 relation to the risk or the intrusiveness that 00:37:19.780 --> 00:37:21.960 it has, but also the potential for innovation 00:37:21.960 --> 00:37:25.179 and that growth. And I guess definitely just 00:37:25.179 --> 00:37:27.280 bouncing off a little bit on what you're talking 00:37:27.280 --> 00:37:29.739 about, very practical, that kind of experience 00:37:29.739 --> 00:37:32.619 within the legal sector with AI, I think. Anthropic 00:37:32.619 --> 00:37:35.199 does a legal AI, there's Harvey as well, and 00:37:35.199 --> 00:37:38.519 Legora as well. There's one very big legal AI 00:37:38.519 --> 00:37:41.139 competitor. And so maybe, I guess, just to kind 00:37:41.139 --> 00:37:43.239 of wrap up this episode, I think a lot of us 00:37:43.239 --> 00:37:45.380 are definitely possibly considering going to 00:37:45.380 --> 00:37:48.480 big law. I mean, it's either, you know, some 00:37:48.480 --> 00:37:50.820 of us also think of going and becoming barristers, 00:37:50.860 --> 00:37:53.139 but whichever path we choose, I do believe that 00:37:53.139 --> 00:37:55.639 AI is going to play an important role. And so 00:37:55.639 --> 00:37:57.699 maybe, I guess, this final question to wrap up, 00:37:57.760 --> 00:37:59.159 I just wanted to know what are your thoughts 00:37:59.159 --> 00:38:02.949 on how, you know, us as students going you know 00:38:02.949 --> 00:38:05.250 graduates going into the into the job market 00:38:05.250 --> 00:38:09.030 can capitalise kind of let AI augment the way 00:38:09.030 --> 00:38:12.190 we work and what kind of where do you see it 00:38:12.190 --> 00:38:15.150 going and where do you think you know maybe a 00:38:15.150 --> 00:38:17.429 line should be drawn or a certain boundary or 00:38:17.429 --> 00:38:18.630 do you think there should be no boundaries at 00:38:18.630 --> 00:38:22.130 all like you know let as AI develops let it you 00:38:22.130 --> 00:38:24.929 know continually augment up the chain and maybe 00:38:24.929 --> 00:38:27.449 you know up the chain of legal work so maybe 00:38:27.449 --> 00:38:31.019 just that yeah that that's my question Yeah, 00:38:31.059 --> 00:38:35.679 I sort of run these workshops with some friends, 00:38:35.820 --> 00:38:39.880 sort of either teaching students or sometimes 00:38:39.880 --> 00:38:44.599 judiciaries on sort of usage of AI. And there's 00:38:44.599 --> 00:38:47.099 these sort of 20 to 30 principles that I always 00:38:47.099 --> 00:38:49.460 like to start with. And it's things like, you 00:38:49.460 --> 00:38:51.500 know, just start playing with it, right? Like 00:38:51.500 --> 00:38:54.840 a lot of like, we're actually in the minority. 00:38:54.900 --> 00:38:56.900 Like when you look at the world's population, 00:38:56.980 --> 00:38:58.969 I think some... report just came out it's like 00:38:58.969 --> 00:39:01.010 84 of the world's population have never used 00:39:01.010 --> 00:39:03.150 AI and then there's further categories of like 00:39:03.150 --> 00:39:06.989 um how much do you vibe code with it and um do 00:39:06.989 --> 00:39:08.550 you only use the free version instead of the 00:39:08.550 --> 00:39:11.789 premium version um have you used the deep research 00:39:11.789 --> 00:39:16.010 function um and also like you know things like 00:39:16.010 --> 00:39:19.389 don't not not just like try to prompt it or like 00:39:19.389 --> 00:39:21.190 prompt engineering but like just try to converse 00:39:21.190 --> 00:39:25.650 with it and um um understand your own tastes 00:39:25.650 --> 00:39:30.139 um and judgment It's as much understanding yourself 00:39:30.139 --> 00:39:32.219 through the process and learning how to develop 00:39:32.219 --> 00:39:34.480 better taste and judgment to better prompt and 00:39:34.480 --> 00:39:37.489 converse with AI. So things like that. But I 00:39:37.489 --> 00:39:39.510 think you're absolutely right. There's so many 00:39:39.510 --> 00:39:42.710 different tools upcoming. I think specifically 00:39:42.710 --> 00:39:46.210 the companies you've mentioned, they're all great 00:39:46.210 --> 00:39:49.409 companies. Some of them have like programs with 00:39:49.409 --> 00:39:52.210 scores as well. So law schools as well. And I 00:39:52.210 --> 00:39:54.030 definitely encourage people to just play with 00:39:54.030 --> 00:39:56.329 it instead of like, do we have these preconceived 00:39:56.329 --> 00:40:00.860 notions? but also like the hot take would be 00:40:00.860 --> 00:40:05.119 what is their mode right like we saw uh Claude 00:40:05.119 --> 00:40:08.159 co -work and code having these legal plugins 00:40:08.159 --> 00:40:11.260 a few weeks ago and that caused a couple billion 00:40:11.260 --> 00:40:13.579 dollars wiped out from the market the so -called 00:40:13.579 --> 00:40:18.409 sas apocalyptic like how easy it is for us with 00:40:18.409 --> 00:40:20.670 minimal background in coding to just vibe code 00:40:20.670 --> 00:40:24.869 a new app that's similar to the AI legal companies 00:40:24.869 --> 00:40:27.010 that you mentioned, right? Or that's similar 00:40:27.010 --> 00:40:29.949 to the particular use cases you have in a law 00:40:29.949 --> 00:40:33.250 firm. And we see this upcoming new generation 00:40:33.250 --> 00:40:36.530 of lawyers who are somewhat familiar with vibe 00:40:36.530 --> 00:40:39.550 coding and they're very useful in law firms not 00:40:39.550 --> 00:40:42.269 only because they can vibe code specific applications 00:40:42.269 --> 00:40:45.570 in the law firm using enterprise version that 00:40:45.570 --> 00:40:50.449 are safe and have data protection privacy protections 00:40:50.449 --> 00:40:54.469 but also they can teach the senior lawyers who 00:40:54.469 --> 00:40:57.440 frankly don't have much time in learning about 00:40:57.440 --> 00:40:59.400 these tools, but who would love to know how this 00:40:59.400 --> 00:41:02.820 impacts their bottom line. And so I think I would 00:41:02.820 --> 00:41:04.840 just encourage people to use it. And then, of 00:41:04.840 --> 00:41:08.260 course, once you start using it, there's so much 00:41:08.260 --> 00:41:12.239 just nuance and so much exciting developments 00:41:12.239 --> 00:41:16.139 that you can keep track of. And it's always, 00:41:16.219 --> 00:41:23.389 I think, important to um use it but also be critical 00:41:23.389 --> 00:41:26.829 of it so having that balance of you know um not 00:41:26.829 --> 00:41:29.150 outsourcing all of your thinking and retaining 00:41:29.150 --> 00:41:31.250 the essential legal tools that you have and work 00:41:31.250 --> 00:41:35.389 um in augmentation um to deliver value right 00:41:35.389 --> 00:41:37.650 and then the other thing is like there's the 00:41:37.650 --> 00:41:40.769 AI and then there's the human part which um there's 00:41:40.769 --> 00:41:43.170 a book called AI Mirror which talks about you 00:41:43.170 --> 00:41:45.789 know it's great you know that AI supercharges 00:41:45.789 --> 00:41:48.670 us but also it's a mirror for us to understand 00:41:48.670 --> 00:41:53.360 um our faults and biases and how do we level 00:41:53.360 --> 00:41:55.840 up and, you know, work together better, right? 00:41:56.340 --> 00:42:00.059 Even as lawyers, we still work largely in silos. 00:42:00.199 --> 00:42:03.619 Like if you go into big law, you go into these 00:42:03.619 --> 00:42:05.300 systems where you have to check in a document 00:42:05.300 --> 00:42:07.639 and then check out. And then when you're using 00:42:07.639 --> 00:42:10.219 it, nobody else can work on it. Whereas like 00:42:10.219 --> 00:42:13.019 a lot of like CS people are using Google Drive 00:42:13.019 --> 00:42:15.300 or like collaborative methods of working, right? 00:42:15.360 --> 00:42:19.280 So the legal industry is changing. I think it's 00:42:19.280 --> 00:42:21.170 important. going to reflect on how do we use 00:42:21.170 --> 00:42:22.909 these tools the best and also reflect on how 00:42:22.909 --> 00:42:26.269 do we use AI to increase our human intelligence 00:42:26.269 --> 00:42:30.250 and wisdom the most thank you and maybe okay 00:42:30.250 --> 00:42:32.130 i know i said that was the last question but 00:42:32.130 --> 00:42:34.489 i think you just talked a little bit about like 00:42:34.489 --> 00:42:37.670 AI tech and just reminded me of this funny like 00:42:37.670 --> 00:42:39.849 video that i used to always see on like Instagram 00:42:39.849 --> 00:42:41.869 where it was this guy in the US i have no idea 00:42:41.869 --> 00:42:43.809 whether it's AI generated it looked quite 00:42:43.809 --> 00:42:46.920 legit but you can never tell basically yeah you 00:42:46.920 --> 00:42:49.199 can never tell but like basically it's a clip 00:42:49.199 --> 00:42:52.039 showing a judge in a US court I believe being 00:42:52.039 --> 00:42:55.039 very upset because this plaintiff basically you 00:42:55.039 --> 00:42:59.079 can imagine a litigant in person uses AI to generate 00:42:59.079 --> 00:43:01.980 an AI lawyer to give his submissions and he just 00:43:01.980 --> 00:43:04.000 plays it off this video and it's like you know 00:43:04.000 --> 00:43:05.619 it shows up on the screen in court and it's like 00:43:05.619 --> 00:43:08.739 you know if it may please the court and it's 00:43:08.739 --> 00:43:11.400 this AI guy talking and obviously she catches 00:43:11.400 --> 00:43:13.599 on and then she says like oh wait is this a record 00:43:13.599 --> 00:43:15.079 and then she you know she goes on to scold him 00:43:15.079 --> 00:43:18.719 and all that uh but i guess my point of asking 00:43:18.719 --> 00:43:22.099 that question is do you see AI actually serving 00:43:22.099 --> 00:43:24.059 that purpose like am i is it going to be a day 00:43:24.059 --> 00:43:26.340 when i just you know plug in a bunch of submissions 00:43:26.340 --> 00:43:28.599 to this AI model and it's just going to generate 00:43:28.599 --> 00:43:30.800 me in court and it's just going to be able to 00:43:30.800 --> 00:43:33.320 deliver my oral submissions and i wouldn't re 00:43:33.320 --> 00:43:35.300 you know it's not good it's going to kind of 00:43:35.300 --> 00:43:37.300 take over that that role because i think a lot 00:43:37.300 --> 00:43:39.260 of us believe that you know we're going to use 00:43:39.260 --> 00:43:41.719 AI and drafting first drafts all that kind of 00:43:41.719 --> 00:43:44.159 stuff in terms of researching processing but 00:43:44.159 --> 00:43:46.369 we don't I think a lot of us lawyers would like 00:43:46.369 --> 00:43:48.769 to believe that we still have a job because the 00:43:48.769 --> 00:43:50.849 oral bit, the stuff when we're doing negotiations 00:43:50.849 --> 00:43:55.110 for terms in an agreement or terms in an acquisition 00:43:55.110 --> 00:43:58.530 agreement or even in terms of speaking in court, 00:43:58.610 --> 00:44:00.570 giving oral submissions, that's something AI 00:44:00.570 --> 00:44:03.989 is not going to replace us in. Do you think that's 00:44:03.989 --> 00:44:07.010 true or do you think it might just enter that 00:44:07.010 --> 00:44:08.929 area and we're just going to have to accept it? 00:44:10.090 --> 00:44:14.349 I think the short answer is read my article with 00:44:14.349 --> 00:44:17.750 Professor Reyes in the AI and Private Law handbook 00:44:17.750 --> 00:44:20.250 on AI and commercial dispute resolution, where 00:44:20.250 --> 00:44:22.909 we discussed this. But the long answer is, I 00:44:22.909 --> 00:44:26.250 think one possible scenario future is there's 00:44:26.250 --> 00:44:31.429 going to be multiple options for plaintiffs, 00:44:31.469 --> 00:44:37.110 litigants to choose from, both in terms of how 00:44:37.110 --> 00:44:40.570 much of AI will be used in the court and how 00:44:40.570 --> 00:44:43.170 much of AI will be used by judges and in the 00:44:43.170 --> 00:44:46.530 submission process. In some parts, also influenced 00:44:46.530 --> 00:44:48.829 by these private markets and arbitration and 00:44:48.829 --> 00:44:51.670 mediation, where there's a massive push for these 00:44:51.670 --> 00:44:55.849 different user -consented mechanisms. And as 00:44:55.849 --> 00:44:58.809 long as the user is fine with it, and it delivers 00:44:58.809 --> 00:45:04.699 justice. in a faster, much cheaper form in a 00:45:04.699 --> 00:45:08.219 way that's better for them, better suited, then 00:45:08.219 --> 00:45:10.159 perhaps that's one way to go. And another form 00:45:10.159 --> 00:45:13.599 is perhaps you can have these first layer AI 00:45:13.599 --> 00:45:18.500 -assisted or AI, fully AI judges, and then you 00:45:18.500 --> 00:45:22.500 have a human layer of review. if you want, and 00:45:22.500 --> 00:45:24.780 that's more expensive. And I think last I heard, 00:45:24.800 --> 00:45:26.679 and anecdotally, this has already been implemented 00:45:26.679 --> 00:45:28.900 in some Middle Eastern, I think, Abu Dhabi courts, 00:45:29.139 --> 00:45:32.980 right? But I think that you have to look at it 00:45:32.980 --> 00:45:35.619 more holistically, or at least that's how I think 00:45:35.619 --> 00:45:39.159 about it, which is, like, in some jurisdictions, 00:45:39.260 --> 00:45:41.719 it's been said, like, you know, the waiting time 00:45:41.719 --> 00:45:46.869 for courts is 3 to 4 years, right? courts like judges 00:45:46.869 --> 00:45:49.469 are systematically overwhelmed this is not just 00:45:49.469 --> 00:45:51.730 unique to one jurisdiction this is a lot of jurisdictions 00:45:51.730 --> 00:45:54.929 and that's why partly the the push for arbitration 00:45:54.929 --> 00:45:57.510 mediation which also now has like procedural 00:45:57.510 --> 00:46:02.750 burdens but um we are seeing um people use these 00:46:02.750 --> 00:46:06.809 tools more uh creatively um both for judges and 00:46:06.809 --> 00:46:10.250 litigants and i think um how to use it in a responsible 00:46:10.250 --> 00:46:13.630 way um is is a big issue right um so there are 00:46:13.630 --> 00:46:16.389 multiple guidelines by judiciaries and um by 00:46:16.389 --> 00:46:19.750 arbitration institutions on how to use this but 00:46:19.750 --> 00:46:22.849 the the second component that you're saying which 00:46:22.849 --> 00:46:25.429 is the super interesting part is what what does 00:46:25.429 --> 00:46:28.710 this mean for for us students and lawyers um 00:46:28.710 --> 00:46:30.949 that are coming into this world right and what 00:46:30.949 --> 00:46:33.489 are the tools um what are the skills that we 00:46:33.489 --> 00:46:36.829 should be using i under the paradigm that i've 00:46:36.829 --> 00:46:41.099 discussed i'm not sure if even just saying i'm 00:46:41.099 --> 00:46:42.860 only going to do my oral advocacy and then all 00:46:42.860 --> 00:46:44.659 the written advocacy is going to be done by AI 00:46:44.659 --> 00:46:48.159 is is true right um it's a fast changing world 00:46:48.159 --> 00:46:51.219 where just a few years ago people would say AI 00:46:51.219 --> 00:46:54.119 is very good at you know analytical and intelligence 00:46:54.119 --> 00:47:00.469 and then um well the um and there's a differential 00:47:00.469 --> 00:47:04.829 the premium is on things like EQ and collaboration 00:47:04.829 --> 00:47:08.050 and how to work with clients and talking with 00:47:08.050 --> 00:47:11.389 clients because we're going to be the trust is 00:47:11.389 --> 00:47:13.449 going to be built by humans human relation where 00:47:13.449 --> 00:47:16.869 everyone sort of increasingly uses AIs i still 00:47:16.869 --> 00:47:19.570 think that's true but then you see papers like 00:47:19.570 --> 00:47:22.869 AI is super good at empathy it's more empathetic 00:47:22.869 --> 00:47:26.070 in its responses than doctors which is like understandable 00:47:26.070 --> 00:47:28.510 because the doctors are so overwhelmed in in 00:47:28.510 --> 00:47:33.070 hospitals right um and so um that there are these 00:47:33.070 --> 00:47:35.469 things like the heuristics that we have like 00:47:35.469 --> 00:47:38.010 of course like it's it's humans humans connection 00:47:38.010 --> 00:47:39.869 and all that which i think is still true and 00:47:39.869 --> 00:47:41.909 maybe it's a matter of waiting at the weight 00:47:41.909 --> 00:47:46.250 that we lean towards these things But I think 00:47:46.250 --> 00:47:48.210 it's just playing around these tools and following 00:47:48.210 --> 00:47:50.050 your passion. And if you think these things are 00:47:50.050 --> 00:47:54.389 very fun to play with, then go for it, right? 00:47:55.250 --> 00:47:58.690 Another way of thinking about it is what Professor 00:47:58.690 --> 00:48:03.730 Andrew from Stanford calls, you know, it's the 00:48:03.730 --> 00:48:07.980 overlap between the... the AI and your substantive 00:48:07.980 --> 00:48:10.659 expertise, that's going to reap the most value, 00:48:10.780 --> 00:48:12.000 right? You're very good at one thing and then 00:48:12.000 --> 00:48:13.699 you're very good at using AI and that's where 00:48:13.699 --> 00:48:15.780 the value comes from. And then there's other 00:48:15.780 --> 00:48:17.619 types of people where you're just very good at 00:48:17.619 --> 00:48:22.559 CS and AI and that's another category of people 00:48:22.559 --> 00:48:24.960 who will succeed. And then there's like maybe 00:48:24.960 --> 00:48:27.280 a third category of people who just know how 00:48:27.280 --> 00:48:30.559 to use AI to leverage the businesses and whatnot. 00:48:30.780 --> 00:48:33.840 So I think... It goes back to what we discussed. 00:48:33.920 --> 00:48:35.500 There's a lot of threats, but there's a lot of 00:48:35.500 --> 00:48:40.300 opportunities and navigating that as students, 00:48:40.360 --> 00:48:43.320 as lawyers requires a lot of wisdom and humility. 00:48:43.480 --> 00:48:46.780 And I think we'll just have to wait and see, 00:48:46.840 --> 00:48:49.440 but also call for action for a more responsible 00:48:49.440 --> 00:48:53.099 future. Thank you. Thank you so much, Adrian. 00:48:53.280 --> 00:48:55.820 And that wraps up our episode for today. So thank 00:48:55.820 --> 00:48:57.820 you, Adrian, for your time. And I look forward 00:48:57.820 --> 00:48:59.460 to having you on the podcast again in the future. 00:48:59.539 --> 00:49:01.559 Thank you so much. thank you thank you so much 00:49:01.559 --> 00:49:03.599 guys have a good one yep