1 00:00:04,200 --> 00:00:07,950 Sean Aylmer: Welcome to the Fear & Greed Daily Interview. I'm Sean Aylmer. AI (Artificial Intelligence) 2 00:00:07,950 --> 00:00:11,880 Sean Aylmer: in the form of ChatGPT or Google's offering, Bard, has 3 00:00:11,880 --> 00:00:15,389 Sean Aylmer: taken off in recent months. In fact, ChatGPT had one 4 00:00:15,389 --> 00:00:19,079 Sean Aylmer: of the fastest tech adoptions ever with tens of millions 5 00:00:19,079 --> 00:00:22,650 Sean Aylmer: of people using the technology to create written content. Other 6 00:00:22,650 --> 00:00:26,520 Sean Aylmer: platforms allow a similar experience with images, but the explosion 7 00:00:26,520 --> 00:00:29,640 Sean Aylmer: in the use of AI has created a legal minefield, 8 00:00:29,880 --> 00:00:33,030 Sean Aylmer: not just over the intellectual property created and used in 9 00:00:33,030 --> 00:00:36,059 Sean Aylmer: the process, but a range of other issues as well. 10 00:00:36,269 --> 00:00:39,390 Sean Aylmer: And it's something businesses need to be aware of. Simon 11 00:00:39,390 --> 00:00:44,638 Sean Aylmer: Newcomb is the technology and intellectual property partner at Clayton Utz. Simon, 12 00:00:44,639 --> 00:00:45,690 Sean Aylmer: welcome to Fear & Greed. 13 00:00:46,320 --> 00:00:47,549 Simon Newcomb: Thanks for having me. Great to be here. 14 00:00:48,659 --> 00:00:54,960 Sean Aylmer: Let's start with governance, AI governance. What AI governance is 15 00:00:54,960 --> 00:00:59,160 Sean Aylmer: needed? What's the framework we should be thinking about to 16 00:00:59,160 --> 00:01:00,540 Sean Aylmer: address all these issues? 17 00:01:01,680 --> 00:01:04,229 Simon Newcomb: So look, there's a lot of benefits of generative AI 18 00:01:04,230 --> 00:01:06,900 Simon Newcomb: and people are really focusing on that and exploring how 19 00:01:06,900 --> 00:01:09,628 Simon Newcomb: they can use it in their business. But there's also 20 00:01:09,719 --> 00:01:13,830 Simon Newcomb: many risks and people are also concerned about that as 21 00:01:13,830 --> 00:01:16,740 Simon Newcomb: they can see how AI can affect people's lives. And there's 22 00:01:17,160 --> 00:01:20,429 Simon Newcomb: a broad range of concerns really across fairness, bias, and 23 00:01:20,429 --> 00:01:24,599 Simon Newcomb: discrimination from the training data, particularly where AI is used 24 00:01:24,599 --> 00:01:31,860 Simon Newcomb: to make decisions. Privacy, security, accuracy, transparency, and accountability. There's 25 00:01:31,860 --> 00:01:34,560 Simon Newcomb: a thing called the black box problem where it's hard 26 00:01:34,560 --> 00:01:37,290 Simon Newcomb: to tell how an AI is making a decision. So it's 27 00:01:37,410 --> 00:01:40,469 Simon Newcomb: not very transparent. And of course, safety all the way 28 00:01:40,469 --> 00:01:44,730 Simon Newcomb: from having reliable information up to existential risks from super 29 00:01:44,730 --> 00:01:48,059 Simon Newcomb: intelligent AIs. So I guess one way to think about 30 00:01:48,059 --> 00:01:52,410 Simon Newcomb: governance is that really there's a need for businesses and 31 00:01:52,830 --> 00:01:57,059 Simon Newcomb: large organisations, all organisations to have a social license to 32 00:01:57,059 --> 00:02:00,179 Simon Newcomb: use AI to address those concerns. And there are some 33 00:02:00,179 --> 00:02:03,659 Simon Newcomb: good frameworks to help with AI governance. The Australian government 34 00:02:03,719 --> 00:02:08,609 Simon Newcomb: has published Australian AI ethics frameworks, and there are some other frameworks 35 00:02:08,610 --> 00:02:10,470 Simon Newcomb: from other governments and NGOs. 36 00:02:11,310 --> 00:02:13,410 Sean Aylmer: Okay, so there are frameworks, and we talk about a 37 00:02:13,410 --> 00:02:18,419 Sean Aylmer: social license, which is not necessarily regulated. Do we need 38 00:02:18,419 --> 00:02:21,900 Sean Aylmer: AI specific regulation? Because certainly there are calls out there 39 00:02:22,199 --> 00:02:25,860 Sean Aylmer: for specific regulation. If we do, what would it look like? 40 00:02:26,430 --> 00:02:31,110 Simon Newcomb: Yeah, so I guess people are really getting concerned that 41 00:02:31,110 --> 00:02:34,828 Simon Newcomb: the existing laws don't go far enough, and self- governance 42 00:02:34,830 --> 00:02:37,950 Simon Newcomb: is really not enough. There's a race by the AI 43 00:02:37,950 --> 00:02:42,388 Simon Newcomb: companies to develop and release powerful new models. And so 44 00:02:42,480 --> 00:02:46,950 Simon Newcomb: there are calls for AI specific regulation from governments. Governments 45 00:02:46,950 --> 00:02:49,440 Simon Newcomb: around the world are including ours, are looking at this. 46 00:02:49,440 --> 00:02:53,758 Simon Newcomb: And almost counterintuitively, I think the AI industry is also 47 00:02:53,758 --> 00:02:56,609 Simon Newcomb: calling for it because they see it as an important 48 00:02:56,610 --> 00:03:00,900 Simon Newcomb: part of breeding trust in the technology. If we can 49 00:03:01,110 --> 00:03:04,950 Simon Newcomb: look around the world at what's going on, the Europeans 50 00:03:04,950 --> 00:03:07,500 Simon Newcomb: are a fair way down the track with this regulation 51 00:03:07,500 --> 00:03:10,290 Simon Newcomb: called the AI Act. And the way that works is 52 00:03:10,290 --> 00:03:13,260 Simon Newcomb: that it classifies AI systems by the level of risk. 53 00:03:13,500 --> 00:03:17,969 Simon Newcomb: So systems with unacceptable risks are prohibited. For example, social 54 00:03:17,969 --> 00:03:21,870 Simon Newcomb: scoring systems that lead to discrimination like the one that 55 00:03:22,020 --> 00:03:25,470 Simon Newcomb: has been reported on in China. And then there are 56 00:03:25,470 --> 00:03:28,440 Simon Newcomb: high risk AI systems that are permitted, but then they 57 00:03:28,440 --> 00:03:32,939 Simon Newcomb: have regulations that require things like registration of those systems, testing 58 00:03:33,179 --> 00:03:36,990 Simon Newcomb: and human oversight and accountability. So examples of those sort of high 59 00:03:36,990 --> 00:03:42,510 Simon Newcomb: risk systems are things like critical infrastructure, education assessment, biometrics, 60 00:03:43,110 --> 00:03:48,029 Simon Newcomb: use in recruitment, immigration, law enforcement, and in the courts. 61 00:03:48,780 --> 00:03:52,379 Simon Newcomb: What that act also does interestingly, is it requires a 62 00:03:52,379 --> 00:03:54,870 Simon Newcomb: system of any level of risk to disclose if it's 63 00:03:54,870 --> 00:03:58,950 Simon Newcomb: interacting with a human, that it's an AI. So you 64 00:03:58,950 --> 00:04:01,320 Simon Newcomb: have to know that you're dealing with an AI and 65 00:04:01,740 --> 00:04:06,270 Simon Newcomb: this regulation like the GDPR will probably become a global 66 00:04:06,270 --> 00:04:10,830 Simon Newcomb: standard because it's got extraterritorial application and very large fines 67 00:04:10,830 --> 00:04:11,429 Simon Newcomb: for breach. 68 00:04:12,510 --> 00:04:15,059 Sean Aylmer: Just on that, it's important. I mean, I think of accounting 69 00:04:15,060 --> 00:04:16,710 Sean Aylmer: standards and you have two kind of main sets of 70 00:04:16,710 --> 00:04:20,489 Sean Aylmer: accounting standards, which always cause friction. In AI, is this 71 00:04:20,490 --> 00:04:23,610 Sean Aylmer: an opportunity to get one overarching set of rules? 72 00:04:24,059 --> 00:04:27,899 Simon Newcomb: Look, I imagine that's sort of holy grail of regulation. 73 00:04:28,260 --> 00:04:30,178 Simon Newcomb: It doesn't often tend to happen that way because you 74 00:04:30,180 --> 00:04:34,110 Simon Newcomb: have different people with different interests and views, and you're 75 00:04:34,110 --> 00:04:36,900 Simon Newcomb: right that you do often end up having to manage 76 00:04:36,900 --> 00:04:39,299 Simon Newcomb: to a patchwork of regulation. And that's certainly happening in 77 00:04:39,300 --> 00:04:43,709 Simon Newcomb: privacy globally now. And I guess there is a fair 78 00:04:43,710 --> 00:04:47,068 Simon Newcomb: potential for that to happen with AI specific regulation as well. 79 00:04:48,000 --> 00:04:50,219 Sean Aylmer: Stay with me, Simon. We'll be back in a minute. 80 00:04:56,580 --> 00:05:00,178 Sean Aylmer: My guest this morning is Simon Newcomb, Technology and Intellectual 81 00:05:00,178 --> 00:05:03,960 Sean Aylmer: Property Partner at Clayton Utz. Okay, so we've sort of 82 00:05:03,960 --> 00:05:06,508 Sean Aylmer: been talking big picture. Let's talk about some of the 83 00:05:06,510 --> 00:05:10,589 Sean Aylmer: particular issues that affects people using AI. IP is a big one, 84 00:05:10,589 --> 00:05:15,150 Sean Aylmer: intellectual property. So who owns the AI generated content? 85 00:05:15,928 --> 00:05:19,079 Simon Newcomb: Well, in Australia at the moment, no one. The thing 86 00:05:19,080 --> 00:05:22,260 Simon Newcomb: is that under our copyright legislation, and that's the type 87 00:05:22,260 --> 00:05:26,370 Simon Newcomb: of IP right that protects texts and images and music 88 00:05:26,370 --> 00:05:28,950 Simon Newcomb: and so on, there's a requirement for a human author. 89 00:05:29,550 --> 00:05:33,570 Simon Newcomb: And where the content's generated by an AI, there's no 90 00:05:33,570 --> 00:05:35,790 Simon Newcomb: human author, and so there's not going to be any copyright. 91 00:05:36,240 --> 00:05:39,690 Simon Newcomb: Now you might say, well, what if I put a 92 00:05:39,690 --> 00:05:43,650 Simon Newcomb: really detailed prompt in there? Is that really me creating 93 00:05:43,650 --> 00:05:46,890 Simon Newcomb: the work rather than the AI? And look, that's arguable 94 00:05:46,890 --> 00:05:51,330 Simon Newcomb: for highly detailed prompts, but in most cases it's really 95 00:05:51,330 --> 00:05:53,759 Simon Newcomb: just giving the AI an idea. And that's not enough 96 00:05:53,759 --> 00:05:57,299 Simon Newcomb: from a copyright perspective. Copyright protects the expression rather than 97 00:05:57,299 --> 00:05:57,990 Simon Newcomb: the idea. 98 00:05:58,410 --> 00:06:02,970 Sean Aylmer: So can we breach copyright if we're using AI? If 99 00:06:02,970 --> 00:06:05,070 Sean Aylmer: I'm using AI, can I be breaching copyright? 100 00:06:05,490 --> 00:06:09,570 Simon Newcomb: Well, that is possible. And the IP infringement issues are 101 00:06:09,570 --> 00:06:12,449 Simon Newcomb: really a bit of an existential risk I suppose for the 102 00:06:12,450 --> 00:06:16,589 Simon Newcomb: whole industry here, because training and AI involves scraping a 103 00:06:16,589 --> 00:06:20,910 Simon Newcomb: huge amount of data from existing sources. So terabytes of 104 00:06:20,910 --> 00:06:23,880 Simon Newcomb: data from the internet, from books and journals and so 105 00:06:23,880 --> 00:06:28,890 Simon Newcomb: on. And these large models then use that to create 106 00:06:29,010 --> 00:06:32,580 Simon Newcomb: the technology, the models, the neural networks that create the 107 00:06:32,580 --> 00:06:36,119 Simon Newcomb: content. And the thing is that some content creators are 108 00:06:36,119 --> 00:06:39,090 Simon Newcomb: not very happy with that because they're not being compensated 109 00:06:39,210 --> 00:06:43,140 Simon Newcomb: or acknowledged. And so there are some cases underway, and 110 00:06:43,140 --> 00:06:46,200 Simon Newcomb: these cases at the moment are against the AI creators, 111 00:06:46,469 --> 00:06:50,550 Simon Newcomb: the technology companies. So there's one against Microsoft and OpenAI over GitHub, 112 00:06:52,080 --> 00:06:55,620 Simon Newcomb: which is the software source code. There's one over artworks 113 00:06:55,680 --> 00:06:59,010 Simon Newcomb: against Stability AI, and some other image generators. And there's 114 00:06:59,010 --> 00:07:02,219 Simon Newcomb: another case by Getty Images, the stock image company over 115 00:07:02,250 --> 00:07:05,820 Simon Newcomb: photographs. And they're all alleging that their content's been taken 116 00:07:05,820 --> 00:07:09,419 Simon Newcomb: without permission and used to train the models. Now, you 117 00:07:09,420 --> 00:07:12,540 Simon Newcomb: can sort of by extension say, well, what if I 118 00:07:12,540 --> 00:07:16,590 Simon Newcomb: then generate content that comes out of an AI that's reproducing 119 00:07:16,590 --> 00:07:18,900 Simon Newcomb: something that came from the training data? And that's where 120 00:07:19,530 --> 00:07:22,530 Simon Newcomb: I think that it is possible that you could be 121 00:07:22,530 --> 00:07:27,119 Simon Newcomb: infringing content by doing that. It's complex though, because the 122 00:07:27,870 --> 00:07:30,540 Simon Newcomb: training data is not at least directly stored in the 123 00:07:30,540 --> 00:07:34,739 Simon Newcomb: model. It's being stored in our brain with synapses and 124 00:07:35,550 --> 00:07:39,840 Simon Newcomb: neurons with millions of interconnections. And these cases are really 125 00:07:39,840 --> 00:07:43,319 Simon Newcomb: going to turn on some of the big exceptions in 126 00:07:43,320 --> 00:07:46,859 Simon Newcomb: the copyright legislation. So in the US they've got this concept 127 00:07:46,859 --> 00:07:50,160 Simon Newcomb: of fair use where you can create transformed works that 128 00:07:50,160 --> 00:07:53,370 Simon Newcomb: don't overly harm the original work. And so that's going 129 00:07:53,370 --> 00:07:57,150 Simon Newcomb: to be a big question. Does training an AI creating 130 00:07:57,450 --> 00:08:02,130 Simon Newcomb: a new technology like this fall within those copyright exceptions? 131 00:08:02,700 --> 00:08:06,359 Sean Aylmer: Wow. Okay. So there's a massive area. What about privacy? 132 00:08:07,650 --> 00:08:11,520 Simon Newcomb: Yeah, so look, similar concerns in a way in that, 133 00:08:11,640 --> 00:08:16,259 Simon Newcomb: again, training these models up involves a large scale collection 134 00:08:16,260 --> 00:08:19,620 Simon Newcomb: and processing of personal information. And many people are concerned 135 00:08:19,620 --> 00:08:22,980 Simon Newcomb: about privacy. And indeed so are regulators. You might have 136 00:08:22,980 --> 00:08:29,250 Simon Newcomb: seen recently that Italy temporarily banned ChatGPT over privacy concerns. And 137 00:08:29,580 --> 00:08:34,289 Simon Newcomb: in Australia we had a high profile privacy incident with 138 00:08:34,620 --> 00:08:38,040 Simon Newcomb: that company, Clearview AI, which you might remember, scraped up 139 00:08:38,040 --> 00:08:41,369 Simon Newcomb: a whole lot of people's images and created biometrics of 140 00:08:41,369 --> 00:08:44,820 Simon Newcomb: them. And that was found to be in violation of 141 00:08:44,820 --> 00:08:48,150 Simon Newcomb: our privacy act. And some of the issues for these 142 00:08:48,150 --> 00:08:52,200 Simon Newcomb: types of technologies are that people aren't notified, they put 143 00:08:52,200 --> 00:08:55,049 Simon Newcomb: their information up, say on the web, and they expect 144 00:08:55,049 --> 00:08:57,449 Simon Newcomb: it to be used in that context, but maybe not 145 00:08:57,450 --> 00:09:00,809 Simon Newcomb: in other contexts to turn up in ChatGPT. The models 146 00:09:00,809 --> 00:09:04,949 Simon Newcomb: can either collect or infer sensitive information about people. And 147 00:09:04,950 --> 00:09:08,280 Simon Newcomb: also we have some changes coming through in our privacy 148 00:09:08,280 --> 00:09:11,309 Simon Newcomb: laws that are going to have some bearing on the 149 00:09:11,400 --> 00:09:15,479 Simon Newcomb: AI technologies like new laws to do with automated decision 150 00:09:15,480 --> 00:09:19,530 Simon Newcomb: making and a new right to have your personal information erased. 151 00:09:22,050 --> 00:09:24,480 Sean Aylmer: That's almost as complex as IP privacy. What about we 152 00:09:24,480 --> 00:09:28,530 Sean Aylmer: bring it back to the business world a bit? I mean, so a transaction, an 153 00:09:28,530 --> 00:09:32,429 Sean Aylmer: M&A deal, for example. Are there implications from AI for 154 00:09:32,429 --> 00:09:36,059 Sean Aylmer: that sort of thing? If someone is undertaking some sort 155 00:09:36,059 --> 00:09:39,960 Sean Aylmer: of business deal or M&A, should they even be worried 156 00:09:39,960 --> 00:09:40,588 Sean Aylmer: about AI? 157 00:09:41,429 --> 00:09:44,968 Simon Newcomb: Yeah, it's a good question. Look, we are changing some 158 00:09:44,970 --> 00:09:48,300 Simon Newcomb: of our approach in the way that we handle those transactions. 159 00:09:48,300 --> 00:09:53,279 Simon Newcomb: So in M&A transactions to acquire AI businesses, we've been 160 00:09:53,279 --> 00:09:57,809 Simon Newcomb: asking some additional questions in the due diligence that are 161 00:09:57,809 --> 00:10:02,190 Simon Newcomb: about AI and we've added some warranties to our share 162 00:10:02,190 --> 00:10:05,309 Simon Newcomb: sale agreement to deal with some AI specific issues because 163 00:10:05,550 --> 00:10:08,700 Simon Newcomb: some of these issues could affect the valuation of the investment. 164 00:10:09,059 --> 00:10:13,650 Simon Newcomb: And similarly in procurement transactions, there are some unique considerations 165 00:10:13,650 --> 00:10:18,059 Simon Newcomb: when procuring AI systems because they work differently to traditional rule- 166 00:10:18,059 --> 00:10:19,170 Simon Newcomb: based systems. 167 00:10:20,250 --> 00:10:21,990 Sean Aylmer: What about things like employment? 168 00:10:22,380 --> 00:10:24,719 Simon Newcomb: Yeah, well look, lots of people are concerned, I suppose, 169 00:10:24,719 --> 00:10:29,189 Simon Newcomb: about changes in their roles or potentially their job even 170 00:10:29,250 --> 00:10:33,060 Simon Newcomb: not existing anymore. And so employers really need to manage 171 00:10:33,450 --> 00:10:37,620 Simon Newcomb: the way that they deal with their workforce and they 172 00:10:37,620 --> 00:10:41,250 Simon Newcomb: communicate with people. And in some cases, they have obligations to 173 00:10:41,550 --> 00:10:45,779 Simon Newcomb: talk to people before doing these types of changes in 174 00:10:45,780 --> 00:10:49,439 Simon Newcomb: their business processes. And our workplace experts are advising that 175 00:10:49,800 --> 00:10:52,980 Simon Newcomb: employers should be having conversations with employees as early as 176 00:10:52,980 --> 00:10:57,300 Simon Newcomb: possible. And there are also issues of discrimination and bias 177 00:10:57,300 --> 00:11:01,949 Simon Newcomb: in using AI in performance management systems where, for example, 178 00:11:02,280 --> 00:11:06,958 Simon Newcomb: in using it for recruitment or in managing under- performance. 179 00:11:07,830 --> 00:11:10,230 Sean Aylmer: It just sounds to me, Simon, that AI is going 180 00:11:10,230 --> 00:11:12,809 Sean Aylmer: to eventually or inevitably is probably a better way of 181 00:11:13,170 --> 00:11:17,370 Sean Aylmer: putting it, touch all sorts of aspects of our business life 182 00:11:17,370 --> 00:11:18,929 Sean Aylmer: and our home life too. 183 00:11:19,080 --> 00:11:20,190 Simon Newcomb: I think that's right, Sean. 184 00:11:20,759 --> 00:11:24,690 Sean Aylmer: And so if you had advice to a business now, 185 00:11:25,380 --> 00:11:27,420 Sean Aylmer: how do you get ahead of the curve in AI? 186 00:11:28,110 --> 00:11:31,440 Simon Newcomb: Yeah, so I think, look, I would be giving, certainly 187 00:11:31,440 --> 00:11:35,219 Simon Newcomb: we are in our business, giving early guidance to people and encouraging them 188 00:11:35,219 --> 00:11:39,089 Simon Newcomb: to explore, but doing that safely. So making sure that 189 00:11:39,089 --> 00:11:44,700 Simon Newcomb: they don't put confidential information into ChatGPT, and very importantly, ensuring 190 00:11:44,700 --> 00:11:48,988 Simon Newcomb: that a human reviews everything that is produced by it 191 00:11:49,590 --> 00:11:52,858 Simon Newcomb: because it is prone to getting things wrong. Then sort 192 00:11:52,860 --> 00:11:56,339 Simon Newcomb: of going on from there, I think businesses should understand 193 00:11:56,400 --> 00:11:59,280 Simon Newcomb: the types of issues we've been talking about and develop 194 00:11:59,429 --> 00:12:02,880 Simon Newcomb: appropriate frameworks to manage them or modify their existing frameworks 195 00:12:02,880 --> 00:12:05,490 Simon Newcomb: to build these types of issues in, adding it to 196 00:12:05,490 --> 00:12:09,358 Simon Newcomb: your cybersecurity program, for example. And then I think in 197 00:12:09,360 --> 00:12:15,000 Simon Newcomb: the medium term, there's potential for much more tailored projects 198 00:12:15,000 --> 00:12:18,179 Simon Newcomb: where organizations start to use their own data to produce 199 00:12:18,630 --> 00:12:24,450 Simon Newcomb: more relevant and targeted and accurate services by fine- tuning 200 00:12:24,450 --> 00:12:28,500 Simon Newcomb: models. And in those sorts of projects where there's big business 201 00:12:28,500 --> 00:12:32,159 Simon Newcomb: process changes or much more targeted solutions, I think that 202 00:12:32,219 --> 00:12:35,400 Simon Newcomb: they really need to incorporate legal compliance by design in 203 00:12:35,400 --> 00:12:36,780 Simon Newcomb: those projects from the outset. 204 00:12:37,350 --> 00:12:39,240 Sean Aylmer: Simon, thank you for talking to Fear & Greed. 205 00:12:39,750 --> 00:12:41,220 Simon Newcomb: It's been great to be here. Thanks, Sean. 206 00:12:42,000 --> 00:12:45,540 Sean Aylmer: That was Simon Newcomb, Technology and Intellectual Property Partner at 207 00:12:45,540 --> 00:12:48,599 Sean Aylmer: Clayton Utz. This is the Fear & Greed Daily interview. Join 208 00:12:48,599 --> 00:12:50,880 Sean Aylmer: us every morning for the full episode of Fear & Greed, 209 00:12:50,880 --> 00:12:54,840 Sean Aylmer: Australia's most popular business podcast. I'm Sean Aylmer. Enjoy your day.