Latest / Views and News with Clarence Ford / SABS publishes draft AI standards for public comment
Transcript
- 0:00Speaker 1: But it is 10.40. It is time for us to
- 0:03Speaker 1: talk about the South African Bureau of Standards. They've published
- 0:07Speaker 1: a suite, essentially, of draft AI standards. And we said
- 0:11Speaker 1: this is necessary. It spans, essentially, transparency. There's data equality.
- 0:17Speaker 1: There's controllability. There's environmental impact. And they've published this for
- 0:21Speaker 1: public comment. And Wendy Rosenberg is the director and the
- 0:25Speaker 1: head of digital media and electronic communications at Workman's Attorneys. Well,
- 0:30Speaker 1: Wendy's noted that these draft standards adopt international ISO and
- 0:34Speaker 1: IEC texts verbatim and are strictly voluntary. They are poised
- 0:41Speaker 1: to become South Africa's default legal and commercial benchmark for
- 0:44Speaker 1: responsible AI. Welcome, Wendy. It's great to have you with us.
- 0:49Speaker 2: Hi, Clarence. G'day to your listeners. Thanks for having me.
- 0:52Speaker 1: I'm just like worried about what you noted there. Do
- 0:55Speaker 1: you think that they used AI to draft this discussion document?
- 1:01Speaker 2: No, no, not this discussion document. That I think you
- 1:04Speaker 2: might be thinking of the draft AI policy, which was
- 1:07Speaker 2: published earlier this year by the department, which was withdrawn
- 1:11Speaker 2: because there were some AI hallucinations there. But to the
- 1:15Speaker 2: minister's credit, he did withdraw that policy and he took
- 1:19Speaker 2: ownership of it and they they've gone back to the
- 1:22Speaker 2: drawing board to make sure that everything stacks up. So that's, yeah,
- 1:27Speaker 2: that's still happening. This, this, these standards have been published
- 1:31Speaker 2: by the South African Bureau of Standards and they, they
- 1:34Speaker 2: separate to the government policy. These are, these are effectively
- 1:37Speaker 2: voluntary standards for AI. Okay.
- 1:40Speaker 1: And I do take in, into consideration, I think, um,
- 1:45Speaker 1: I think the, the local realities, um, I think that's
- 1:48Speaker 1: very necessary that while it's great for us to have
- 1:51Speaker 1: some kind of global standard or benchmark, there are South
- 1:54Speaker 1: African realities that defines us as different, isn't there?
- 1:58Speaker 2: Yes, I think so. I think we must always be
- 2:02Speaker 2: careful about not just copying and pasting from international sources.
- 2:07Speaker 2: In this case, though, I think there are quite a
- 2:09Speaker 2: number of advantages. It's actually a worldwide norm to have
- 2:13Speaker 2: these global experts develop standards that are then internationally recognized standards,
- 2:19Speaker 2: and it's quite the norm for countries to adopt these standards.
- 2:23Speaker 2: And what it then gives us is the advantage of
- 2:28Speaker 2: world standards without reinventing the wheel. And, you know, if
- 2:32Speaker 2: you're trading with Europe, for example, and your products stack
- 2:36Speaker 2: up to international standards, I think that will stand you
- 2:40Speaker 2: in good stead. But, of course, the point of consulting
- 2:44Speaker 2: by the SABS is specifically to make sure that these
- 2:49Speaker 2: are fit for South Africa. So the intention or proposal
- 2:52Speaker 2: from the SAVS is to replicate the international standards, but
- 2:57Speaker 2: these are actually draft standards open for comments and anyone
- 3:00Speaker 2: can comment on them. And to the extent that people
- 3:03Speaker 2: think that we should be taking local considerations into account
- 3:07Speaker 2: in any particular way, then it's up for people to
- 3:10Speaker 2: make those comments.
- 3:13Speaker 1: I was talking to a colleague a little earlier from
- 3:15Speaker 1: Deutsche Welle in Berlin, and the conversation was around data sovereignty.
- 3:23Speaker 1: Maybe you can help us to understand what exactly that
- 3:25Speaker 1: means and essentially what considerations must be part of this
- 3:30Speaker 1: entire understanding of where we're going with AI.
- 3:34Speaker 2: So I think the key thing with AI is that
- 3:38Speaker 2: these large LLM models, they are trained on data sets.
- 3:42Speaker 2: And so they need huge volumes of data on which
- 3:45Speaker 2: to train the AI models. And so whatever your AI is,
- 3:49Speaker 2: it's got to be learning from something. So one use case,
- 3:53Speaker 2: for example, that we see coming out of the use
- 3:56Speaker 2: cases published by the SABS, an example would be you
- 4:00Speaker 2: can apply for credit, say, and it's a fully AI
- 4:04Speaker 2: system that is... automated and you can apply and receive
- 4:09Speaker 2: credit in seven minutes this is this is one of
- 4:11Speaker 2: the examples actually that's just been published and so it's
- 4:14Speaker 2: top of my mind now what would happen there is
- 4:17Speaker 2: that there's data that needs to feed into that system
- 4:20Speaker 2: to train the model and i think that's why data
- 4:23Speaker 2: is so critical to anything ai related because of the
- 4:28Speaker 2: importance of transparency and the data quality and as you
- 4:32Speaker 2: mentioned data quality is a key pillar of these standards
- 4:35Speaker 2: and uh As the saying goes, rubbish in, rubbish out.
- 4:39Speaker 2: If you're putting in flawed data or biased data, for example,
- 4:45Speaker 2: then you can and should expect a flawed output. So
- 4:49Speaker 2: data is absolutely critical. Okay.
- 4:52Speaker 1: Let us just also understand this SABS document because– South
- 4:57Speaker 1: Africa doesn't have a standalone AI law yet. So this
- 5:01Speaker 1: is just kind of benchmarks for business out there at
- 5:04Speaker 1: this moment in time. Or will this translate or comprise,
- 5:08Speaker 1: you know, laws that I believe are in the making
- 5:11Speaker 1: as well?
- 5:13Speaker 2: So, correct. So we don't have a standalone law on
- 5:17Speaker 2: AI yet. We do have the draft policy, which was
- 5:20Speaker 2: withdrawn and which will likely be before cabinet early next year.
- 5:25Speaker 2: But we do have certain laws that do, of course,
- 5:28Speaker 2: impact on AI. So we've got data protection laws, consumer
- 5:31Speaker 2: protection laws, cyber laws, et cetera. So we're not completely
- 5:35Speaker 2: out to sea. But what these standards that are published
- 5:38Speaker 2: by the SAVS do, what they do is they provide
- 5:42Speaker 2: voluntary standards. They're not law unless a law comes into
- 5:45Speaker 2: force that says you have to comply with these standards
- 5:48Speaker 2: if you do X. So, for example, there are laws.
- 5:51Speaker 2: that say you must comply with SAP standard, whatever it
- 5:55Speaker 2: might be for your washing machine. It might be for,
- 5:57Speaker 2: this is a norm where you, you know, for your
- 6:00Speaker 2: mobile device, whatever it is, there's certain standards which are
- 6:04Speaker 2: legally required. This is not yet a legal requirement. They're
- 6:08Speaker 2: still a draft standard. But what they do is they
- 6:11Speaker 2: put in place practical steps that any business can then
- 6:14Speaker 2: have verified and say, I've done X, Y, Z. And
- 6:18Speaker 2: that's why, My AI system is transparent, for example. You
- 6:22Speaker 2: know that you're dealing with an AI bot. You know
- 6:26Speaker 2: what's gone into this. You know what data I'm collecting
- 6:30Speaker 2: from you, for example. So transparency is one of the
- 6:35Speaker 2: key pillars. So what it does is it creates a how.
- 6:38Speaker 2: It creates a mechanism for businesses to say, I'm going
- 6:41Speaker 2: to implement these standards within my business, and then I
- 6:45Speaker 2: can have that accredited, like a BEE certificate if you
- 6:48Speaker 2: think about it. Just as an example. I get that.
- 6:57Speaker 1: Pardon my ignorance. And I think I want to take
- 6:59Speaker 1: as many people on this journey because this is a
- 7:02Speaker 1: journey of discovery and I think a necessary one as
- 7:06Speaker 1: well for businesses using technology. machine learning tools today, essentially.
- 7:14Speaker 1: Give us a sense of what a mini audit kind
- 7:16Speaker 1: of looks like before they make their submissions by the
- 7:21Speaker 1: 10th of November. What should they be doing right now?
- 7:25Speaker 2: So you mean in terms of the standards, what should
- 7:28Speaker 2: businesses be doing? Or generally, what should businesses be doing
- 7:31Speaker 2: for AI?
- 7:32Speaker 1: Well, comments are due by the 10th of November, and
- 7:34Speaker 1: it must be an informed comment, obviously.
- 7:36Speaker 2: Okay, right.
- 7:37Speaker 1: Yeah.
- 7:38Speaker 2: So these draft standards are publicly available. And I think
- 7:42Speaker 2: anyone who is either developing AI or deploying AI, so
- 7:47Speaker 2: if you're a business commissioning AI, for example, and AI
- 7:50Speaker 2: is becoming a big part of your offering and AI
- 7:54Speaker 2: is particularly integral to you, then you want to have
- 7:57Speaker 2: a look at these standards and say, okay, if I
- 7:59Speaker 2: want to be ticking this box and getting my gold stamp,
- 8:04Speaker 2: what kind of processes should I be doing and do
- 8:07Speaker 2: I think that these standards that have been proposed are
- 8:11Speaker 2: an appropriate way to go? I generally think they are
- 8:14Speaker 2: an appropriate way to go. I think they've taken world-class
- 8:17Speaker 2: standards and suggesting making them a South African standard. But
- 8:23Speaker 2: if you think there's a different standard or a better
- 8:25Speaker 2: standard or there are ways that this could be localised
- 8:27Speaker 2: for South Africa better or whatever your comments are, this
- 8:31Speaker 2: is a public participation process. It's open to anybody to
- 8:34Speaker 2: comment on. And anyone can do so.
- 8:38Speaker 1: Okay. So this is no different to what? To that
- 8:43Speaker 1: SABS product that is approved on a shelf. It signals
- 8:48Speaker 1: to the world that this product is being tested in
- 8:51Speaker 1: very many different ways. And it's a benchmark, an international benchmark, essentially.
- 8:57Speaker 1: And you want to have that, obviously.
- 8:59Speaker 2: Exactly. Okay.
- 9:00Speaker 1: Now– Who should ultimately be accountable then? Let's move down
- 9:05Speaker 1: the line and this SABS benchmark is achieved. Maybe parts
- 9:12Speaker 1: of it finds form in a law. Let's understand accountability there.
- 9:20Speaker 1: When AI system produces harmful or incorrect outcome, the developer,
- 9:25Speaker 1: the organization deploying it, what the data provider, the human
- 9:28Speaker 1: decision maker, Do the proposed standards offer sufficient clarity on
- 9:32Speaker 1: this yet, or is that under production?
- 9:36Speaker 2: So accountability is one key principle of AI governance. It's
- 9:41Speaker 2: not a focal point of these particular standards, but accountability
- 9:45Speaker 2: is critical. So if I go back to the earlier
- 9:49Speaker 2: example where I said you could apply for credit within
- 9:52Speaker 2: seven minutes and that would be an entirely AI automated system,
- 9:57Speaker 2: if you don't know that you're engaging with AI, if
- 10:00Speaker 2: you don't have that information disclosure, it's very hard to
- 10:03Speaker 2: challenge it, for example. So transparency would feed into accountability.
- 10:08Speaker 2: Another example would be that if you have an AI
- 10:12Speaker 2: system deciding on your credit and not giving you a
- 10:16Speaker 2: chance to appeal that, And it might make biased decisions,
- 10:20Speaker 2: for example, based on your gender or your address or
- 10:24Speaker 2: your race or whatever it is. That's one particular example
- 10:27Speaker 2: of where we see AI can sometimes perpetuate human biases.
- 10:33Speaker 2: So that would be an example of accountability means I
- 10:37Speaker 2: can then go back to you and say, hey, hold on,
- 10:40Speaker 2: I don't agree with you or I want to be
- 10:42Speaker 2: able to challenge this and you've got to give me
- 10:45Speaker 2: some right of recourse. as opposed to a world where, well,
- 10:48Speaker 2: the AI system has made its mind, there are no
- 10:51Speaker 2: humans in the loop, and sorry for you. So another
- 10:55Speaker 2: pillar of these standards would, for example, be ensuring some
- 10:57Speaker 2: level of human involvement, which again goes to accountability, to
- 11:02Speaker 2: ensure that you don't just have a computer making decisions
- 11:05Speaker 2: where no human is involved.
- 11:08Speaker 1: Okay, now... What happens if you say, no, I'm not
- 11:10Speaker 1: interested in these SABS benchmarks, I'm not interested in the
- 11:14Speaker 1: global benchmarks, and it's going to give me, well, a USP.
- 11:19Speaker 1: I'll be able to underprice my product.
- 11:21Speaker 2: Look, I think every business must make its own decisions.
- 11:28Speaker 2: These are voluntary, and I think anyone who puts any
- 11:31Speaker 2: system out has to make, you know, obviously have to
- 11:34Speaker 2: operate lawfully. But beyond that, you know, I can decide
- 11:37Speaker 2: on the quality of my product and my customer service,
- 11:40Speaker 2: for example. So as long as people are operating lawfully,
- 11:44Speaker 2: we can't dictate to people what to do. But then
- 11:46Speaker 2: people need to be informed and people need to make
- 11:49Speaker 2: informed choices on which products they use. About a half
- 11:53Speaker 2: an hour ago, I came out of a forum where
- 11:55Speaker 2: someone from a commercial entity was saying it was negotiating
- 11:59Speaker 2: an agreement with someone who is not prepared to make
- 12:02Speaker 2: any warranties on what AI systems they use and whether
- 12:05Speaker 2: they're reliable, then your decision as a business is, well,
- 12:08Speaker 2: I either use you or I don't. I either negotiate
- 12:11Speaker 2: some kind of indemnity or I go elsewhere to people
- 12:15Speaker 2: who are willing to give me certain assurances.
- 12:19Speaker 1: Yeah, okay. So transparency is one of the key areas
- 12:22Speaker 1: I would imagine covered by by the drafts, but I
- 12:25Speaker 1: would imagine there's some maybe proprietary information there. Just how
- 12:30Speaker 1: much can businesses be expected to disclose about their AI systems?
- 12:36Speaker 2: So think about it a bit like an ingredient label
- 12:39Speaker 2: on a food product, right? So there's an ingredients label,
- 12:44Speaker 2: you're telling me the nutritional makeup, you're telling me that
- 12:47Speaker 2: it's got a common allergens in it. You're telling me
- 12:50Speaker 2: that it's got wheat, sugar, eggs, whatever it is. You're
- 12:52Speaker 2: not giving me your trade secrets. You're not telling me
- 12:55Speaker 2: how you made the product or what your secret recipe is.
- 12:59Speaker 2: And I think transparency is much the same. So it's
- 13:02Speaker 2: about giving people the information that they need to understand
- 13:07Speaker 2: what goes into this, where my risks are, but without
- 13:10Speaker 2: requiring them to disclose the crown jewels and trade secrets.
- 13:14Speaker 1: Or third-party kind of service providers in respect. No, absolutely.
- 13:19Speaker 2: We've got to respect confidentiality and private information.
- 13:26Speaker 1: But we've learnt of later third-party service providers could be
- 13:28Speaker 1: weaknesses in the ecosystem as well.
- 13:33Speaker 2: I think that's a part of life. I think you
- 13:35Speaker 2: see that with data breaches and cybercrime. Unfortunately, that's part
- 13:39Speaker 2: of the reality we live in.
- 13:40Speaker 1: Okay, so again, I hear where it's going, getting the
- 13:43Speaker 1: standards or wrapping our minds around related standards. But I
- 13:48Speaker 1: again see small businesses, I see medium businesses who can
- 13:52Speaker 1: probably benefit most from AI at this particular moment in time,
- 13:55Speaker 1: maybe lagging behind, maybe cost factor setting them back. Where's
- 13:59Speaker 1: your mind on that, Wendy?
- 14:03Speaker 2: You mean small businesses that are developing AI or actually
- 14:06Speaker 2: using AI?
- 14:06Speaker 1: Or using AI, both, yeah.
- 14:09Speaker 2: Listen, I think if you're using AI, you've got to
- 14:12Speaker 2: make sure you're doing it responsibly. And I think no
- 14:15Speaker 2: matter how small you are, you have a responsibility to
- 14:20Speaker 2: use AI in the same way as if you use email,
- 14:23Speaker 2: you've got to use it responsibly. Or if you use
- 14:27Speaker 2: social media, you've got to use it responsibly. If you're
- 14:30Speaker 2: driving on the road, you've got to drive carefully. So
- 14:33Speaker 2: I do see what you're saying about the cost of
- 14:35Speaker 2: doing business. I don't think there are any insurmountable obstacles
- 14:40Speaker 2: to doing business. And I think every person has to
- 14:42Speaker 2: operate at the level that's appropriate for them. So depending
- 14:46Speaker 2: on your sophistication and your size, et cetera, I think
- 14:48Speaker 2: you've got to work in a way that's appropriate for
- 14:52Speaker 2: your business.
- 14:53Speaker 1: Okay. Then we also talk about AI as, as an
- 14:57Speaker 1: homogenous entity, nearly while we seeing maybe two different values
- 15:02Speaker 1: underpinning its generation, the innovation and, that is coming from
- 15:06Speaker 1: the East and the innovation that is coming from the West.
- 15:09Speaker 1: Do we evaluate all AI in the same way?
- 15:15Speaker 2: So look, each country has developed its own AI policies,
- 15:20Speaker 2: taking into account its own trade objectives, its social political objectives, etc.,
- 15:26Speaker 2: And we see, for example, out of the EU, the
- 15:29Speaker 2: EU AI Act is actually based on product safety legislation.
- 15:33Speaker 2: They're trying to push an EU, what we call the
- 15:35Speaker 2: Brussels effect, similar to what they've done with data protection legislation.
- 15:39Speaker 2: And they want that to become the global standard. It
- 15:44Speaker 2: is a heavy-handed approach. And similar, if you went to
- 15:47Speaker 2: the States, you would receive a lighter touch, more trade-friendly
- 15:52Speaker 2: kind of approach. regulations on AI. So I think depending
- 15:57Speaker 2: on the policy, you might have different regulation levels and different,
- 16:03Speaker 2: you know, more or less onerous.
- 16:06Speaker 1: Would that be from brand to brand or is it
- 16:09Speaker 1: East-West divide?
- 16:10Speaker 2: I think it's almost... Country by country, so for example,
- 16:15Speaker 2: the USA having adopted a particular approach, UK a different
- 16:19Speaker 2: approach to the rest of the EU, well, post-Brexit. You've
- 16:23Speaker 2: got China, for example, adopting a very particular approach to AI.
- 16:28Speaker 2: And each one is driven by their own objectives.
- 16:32Speaker 1: Yeah, commercial interests, unfortunately. Instead of human interests, but I
- 16:36Speaker 1: guess you can't have human outcomes without commercial outcomes. Okay,
- 16:40Speaker 1: the deadline for submission, for comment, is the 10th of November.
- 16:44Speaker 1: Maybe you can drive us into practical direction. What concerns
- 16:49Speaker 1: should those interested in this conversation be exercising? What kind
- 16:53Speaker 1: of concerns should they be voicing by or on the 10th?
- 16:56Speaker 2: So I actually don't have too many concerns about these
- 17:02Speaker 2: draft standards. I think they're a positive proposal and I
- 17:07Speaker 2: think that the benefit of adopting international standards is that
- 17:10Speaker 2: we do then have a gold mark stamp that we
- 17:13Speaker 2: can apply. But if people think there are other standards
- 17:17Speaker 2: maybe that should be incorporated as well or certain localization
- 17:22Speaker 2: requirements or those kind of things, there is a template
- 17:26Speaker 2: for comments to the SAVS. You can make granular comments
- 17:29Speaker 2: if you wish. Yeah, I think it's just a public
- 17:34Speaker 2: participation process that's there for everybody to know about and
- 17:38Speaker 2: participate in.
- 17:40Speaker 1: And an important heads up as well. And thank you
- 17:43Speaker 1: for helping us through the responsibilities of the day in
- 17:46Speaker 1: respect of business and AI and South African Bureau of
- 17:49Speaker 1: Standards may be showing the way. Wendy Rosenberg, a director
- 17:52Speaker 1: at Workman's Attorneys, bringing us up to speed. I think
- 17:55Speaker 1: there's a lot more conversation that we'll have and must
- 17:57Speaker 1: have on this topic. Let that serve as an introduction
- 18:00Speaker 1: at three minutes to 11 o'clock.