Assuring data and AI

I'd love to you and Steve Garvey Garvey. Off, He is the senior policy adviser at the center on Innovation, and he is going to be a few nights before I sign three, but it's highly. You can send questions and read it before 10. Thank you. Thank you very much. And thank you. All welcome to this conversation on ensuring. They turn a. I l said I'm a, I'm a, I'm a senior policy adviser at the CDI where I laid out, work on a Asher. And I'm was also part of the team that also reviewing Tobias and algorithmic

decision-making last year. I've been hearing throughout the many conversations and Cognac and also elsewhere while becoming increasingly aware that they are. And all the jobs are driven technologies have the potential to unlock really significant benefits to Society of many sectors from the convenience. We take for granted in our smartphones and web applications the fundamental breakthroughs in medicine and science. These Technologies. Also bring significant risk. Most obviously around issues like privacy in our personal data, but also the potential to buy some discrimination or when

you risk to our Safety and Security won big on the line challenge. Is that the birthday Toya and I are to be trusted it requires that those who develop use and affected by these technologies have meaningful and reliable information about how these Technologies are made. And used we call this the car insurance. Insurance cover the number of governance mechanisms that allow third-parties develop trust in the compliance and risk of a system organization originally developing and developed from the accounting profession. There's already a range of assurance techniques that have been legend

of the contacts and then, things like order that application or impact assessment would also seen a number of efforts to develop Insurance methods such as orders in, but they are, and I and II, however, these efforts have been quite fragmented and it had mixed success. Today to discuss these challenges. We are joined by Milly's on the top and Marie accidente. Really, the Messiah is head of public policy at the open daughter Institute. She previously LED work on data and policy at the Royal Society and the medical research Council. She holds degrees in Philosophy from Oxford, Cambridge.

Maria can't I at least a double you say is what I'm responsible a I and I feel good when she advises industry Academia government's anmol and how to address the power of AI in Mexico and responsible. Male is also very active in a life policy and I are standard with her at the Antelope Valley. Poppy parliamentary group, when I text you guys do, I join AIT or stand up comedies of the British and the ultra Play. Welcome to the party. On the ground, we could come out in in this conversation with you Maria. Why do you think the AI insurance is important in the board of context of a

high government governance and responsible AI? Hello. Hello, everybody. Thank you for having me and Eliza to say hello from Kings Cross. And although, this is a virtual event. I paid my way on site. Because how else would you celebrate call? Makes the biggest Festival of a i and the question. Why do we need quality assurance? Wouldn't we want to make sure that the Arts in the computer science and social development have the same standards of quality as we do with everything else, right? We call it. You sure our cars are trains are buildings

to make sure that they comply with certain standards. They are Feats the purpose. We prevent accident and they are dead to be used for long. Of time. So we if we have developed this discipline of a showing the quality, for a lot of the artifact would build around us. We need to apply the same thoroughness in the software development. Well, I just Call assurances of the development has been there for a long time but that the increased adoption of development and use of AI points, few challenges and I

want to talk a little bit about those those challenges and why quality assurance in the world of AI is is is is Paramount. First of all, we talkin about a stalker artifact that we're building, whereas the processes that strong, and Gavin the building. Use a vir deterministic in the second big challenge that we have when it comes to a r, is that a i display, its own agency interacts with its external environment adapt to it? And has a certain degree of autonomy, and

a new way of thinking, of, how do we use this? This, this artifact is needed and lost. Probably the most important while wet when we need to to accelerate. The the the quality assurance of AI is a fact that we are dealing with technology that has a profound impact. Not only at the individual Level Society but a planetary level and while we still scrambling to governess with local and international and potentially supranational tools. We need to pay more attention know

just how well we build it. But keep an eye on when we use it how well we use it. Thank you, and just picking up on that. I think there's a really powerful potential for Assurance to support. So the regulatory efforts by the government as well as support industry to understand what they buy and what they using. It's not necessarily that these are alternatives to regulation or self-regulation. Thank you for that. I'm really coming to you while many of the conversations of this week or obviously about a i a really

important characteristic of a particular. The machine learning is that it relies on data and trustworthy date or is obviously an increasingly important will generally in society. What do you think? It means to build trust and trustworthiness in darkness, you Thank you very much. It is to be a assurance and it's a process that provides confidence that they can. I say she's collecting using and sharing sites of the day. I'm going to school. I'm thinking about what you want to know about the school before you decide to send your child death.

I'm thinking about 4 today to reuse. It look like a choir entice, her face Insurance provide transparency directions. I suppose an internal aspect which is kind of, you know, I'm in which facing the relational aspect that's very contact sensitive depends on. It depends on, it depends on what else is happening. It depends on your history. Be quite Dynamic and complex sentences. The other aspects of trust his trustworthiness, which is internal, which is about you are holding yourself accountable. Taking measures to ensure Integrity of

insanity and governance framework checks and balances skills and training competition. So, this is so different. I'm thinking about it not just in time to face her, for our Christmas use, but they took for other uses to the wedding is. So I found out on the old liking, the site to be really, really influential and interesting the way that you distinguish trustworthiness about is about yourself about the organization itself and that sort of an important starting point, but it's not enough. It's not

just enough that what you're doing is trustworthy. Need to have good reason to trust you. And these Insurance tools are really important to the communication device. One challenge that raises his translation because what information you might need a developer to to know whether you can trust the system, might be way more complex, or in a totally different caring about totally different things than what the end-user might care about you, knowing whether they should trust. So thank you for bringing this really useful pretty much for thinking about the

problem. Where do you think about, where do we think? The problem of insurance is most urgent? And where do we think? There's a most compelling made a way. We should be focusing on the bus with you again. What do you think? I insurance is most urgent. That's an excellent question. And I would say, every where we need to make quality assurance of a iPods, the sound, a part of the BAU the business-as-usual. The standard process is we go about build a, I envy you say. I don't do this at this point in time, when my focus too much on high-risk

application. For example, where we would will receive in the European commission with their AI, a new Xbox costing. We need the phone quality assurance across the life cycle, for all the application with building with a, with a certain degree of thoroughness and answer to the gross of his vacations of the two-week. We might be using. And, you know, that you do this. You need to be having some prerequisites building already like impact assessment. And, and, and the fact that we need to consider the How old is stakeholder? Not just thought that those directly impacted by, by the application

will with will will, will will receive the application on long-term, carry one human, human rights assessment, but also equally risk assessment. Very importantly is that we need to be still the discipline of documentation among data scientists and Engineers to be able to, to hop that transparency in the process and understand how sitting having made his responsible. And also very importantly, how to be able to mitigate when things can go wrong and ultimately, another prerequisite, as

important quality assurance is the governor's. Know, we have talked so much about Gavin is focusing. Probably a little bit too much on operation rising at 6. I'll smack the governance is about having Clarity of who is responsible of what and where do we have? Where? Set the boundaries of responsibility the ending and you wants to know, how do we knocked governance models? That will recognize those key characteristic of, a are the fact that we can see a profound impact. And, and we all we require to rethink how we stay in

control of of the highway built. So, yes, I will take everywhere. Let's stop consider this. An important part of the way with Bill deploy and use AI, you know, that than to be deciding. Where do we want to focus a bit more attention. For example, for the high, I high risk application, but do not lose the eyes from the bowl quality. Assurance is is essential if we want a artwork for us. Thank you, Mary. I might just throw a couple of friends from that. Where you say, for example, that you focus is on the high-risk uses of Technology from the perspective of a

mandatory can pull. That is an assurance mechanism Assurance methods such as impact assessments, going to be really helpful, really important for making sure that the people who are responsible for using days lower risk, but still in some cases quite significant Technologies, but they may not be providing that insurance to governments. They might be providing insurance through there and uses the other distinction. I wanted to throw it out is between risk assessment an impact assessment. There's quite a lot of discussion

about this. The way I see, this is a risk assessment is quite subjective or internal its risk to you will refer to you as the organization. Let's make it a point to Technology's impact assessment. Thanks. Why. And I think it's important to do both because otherwise, if you just take a risk, I can patch a price. You might only care about impact on society. And so far as I create reputational, well, compliance damage. When is quite a going to think about that at the

right. Anyway, it's a very helpful clarification and you are absolutely right to say that. Then application-level. We see the emergence of stakeholder impact assessment of risk. Assessment that more and more. A part of the development process is to see that the government's talk about those impact assessment for the application to being using the public sector. So you have the impact Assessment bill by the Canadian government is a very good example about how to increase agree, that was trailing government has been talkin an ender pearl. Noting that the human

rights impact assessments and it sends a very clear message is you just said that he's important to go beyond the boundaries of responsibilities of an additional level and consider, what can be cold as extended responsibility. Where do we start? Looking at the size of problem, that where we have Collective responsibility for all those issues, and being able to solve it as a collective rather than individual. Write, thank you, ma'am. I'm coming to you now. Wandering Way use your, saying the issues of Gator, Insurance, being most urgent. Or whether

the most clear case that we need to be on the same thing as a Christian. Some people argue that faces Jewish, maybe it's kind of protection because I lost so many other things, but otherwise we can pick up this. This increasing our double date of stewardship, as opposed to just stay to protect and that and that's what a date of stewardship. All governments also means a responsibility to think about unlocking the benefits and not just focusing on the downside risks. I think that is true, but I am tied up, but it's interesting to say that. Now, to pick up in the

day, the language. I just wanted to highlight there with you. Both been talkin about is this is not just about Assurance of the Assurance of the artifact and not just talking about Assurance of the algorithm. Or surance. All the data. I'm quite often. We're talkin about Assurance around the development process or Assurance about how these things are on. And I think I'm taking that attitude is is a really important distinction as well, because they may be limits to how much we can assure

these artifacts when they're constantly evolving. Like I said, may it may be that we put that we put the wagon to getting getting the assurance that the standard price standard process has been responsible. Governance level ideally both but they may be some places we have tonight. What to turn to living from the sudden urgency to maybe the the recent developments? And what's up with you? What have you been saying is the main developments in bed or insurance?

I kind of want something that wasn't a forever, you know, the Racine Public depending on on on patients. Acknowledged its limitations and fight through it and discuss the risks around that stuff. I think of standard making in the sense of common language. We all know that that the daughter is always in perfect. But how do we have a Common Language about how good it is in what ways is Florida? So that we understand what might be appropriate and inappropriate uses? And one challenge was saying today, but it is starting to get that. Is that open the

people who are involved in the collection of his daughter model? My speak a different language in terms of like technical specialty to the people who end up using it and making sure that as daughter is increasing part of the economy to reuse that, we are picking sign language about these limitations. Do you think about some of these topics and times that I order staying in the like at the law school gets conference and I'm wondering in the past year. What have you been saying that the developments in AI

assurance and as a result, do you think you're more concerned or more hopeful about the potential of becoming II? As it's true people who don't be that. I know that I'm incurable Optimist. I will start by saying that I'm definitely not the most things are going in the right direction. But also what I seen in the last year since we had the session a call back switch to channel to read about the case of quality assurance and I like how much things have Evolved. First of all, the work you guys are doing to the Center for date. Isaac's of innovation.

Hoo-ha. Let's, let's celebrate. Because you're doing exactly what was needed to correct. Framing of what are surance. He is, and what's the need for it? But other than that, the stakeholders and who should be involved in it, that did not exist a year ago. We had initial and conceptual definition of what it is, but not not something that brings it all together. So I'm very excited to see this world that you guys are doing and I'll encourage everyone who's interested in the topic to contribute to the war bad guys, on the team are doing the second. I think that we have In the last

year is an increased awareness of the topic among those who are not necessary in the public policy space or in the external engagement. We've been running the research last year and responsibilities. I can understand where our clients Minds when he comes off a responsible and ethical, you know, how well-prepared they offer. The option of a I was sort of measures. They taking her own governance and risk and embedding ethics and we were pleased to hear that more and more Executives. Do you know, y'all are considering the

ethics of AI by developing various initiatives from principles Frameworks. The internal policies imposed, but I'll still having a high-risk part of a standalone part of the, a strategy that tells quite a lot about that. Mine shift change that has happened in, in the last year alone in between positioning something is Spell beneficial vs considering of the that the last positive implications and in parts of a i l s to meet. Probably my main reason for Optimus to see that the industry is picking up on it and

hand in hand with preparing this ground for a more robust Governor, so that we could see the interest of quality assurance bubbling on the side so much so that we can before. The last year. We seen a lot of interest from companies building plans. So I can play Solutions in testing by 6 or basses in security is pretty much the responsibility to keep that we have developed as well. Only to realize that we know quite there yet then. And I hope I recently with Cass Camille and she said that's probably where somewhere five five years before we're able to Plug and Play this type of of

ready-made solution. The first thing we need to be doing is processing the contacts and how those problems have been formulated. And if you are a fan of Hawaii, What time you've read her Booka, you know, weapons of mass destruction, you have seen how applying the principles of quality assurance and she has discover, you know, that the surgeon need to reconsider, how we develop algorithms in a certain context, who is involved, what sort of objectives were sort of policy. We have attached to it and how to go about it in a way that reflects of this profound impact

on people's lives. So yes, I'm, I'm often mistaken remain to be see how quickly the quality assurance will be fixed up as a strategic proxies in the development of the eye and make it, you know, demonstrate that the benefit of Karen quality assurance. Go beyond the increasing bureaucracy, and I really love what what Billy was saying. It's like we need to start thinking that this is not just yet another bureaucratic exercise. Just another combines exercise helps his insurer achieving That was ethically responsible outcomes.

Today. We all want their match to if we if we want to get that lets that incorporate a little bit more hassle than General complexity, you know what the cheat sheet to get that results that we all want so much. Thank you, Maria. Oh, I agree with that. In terms of the rising awareness. I'm where I may be a bit more skeptical. Kind of in principle is how and and definitely it's an open debate is how far would these things really be industrialized on standardized? I'm in the reason I say that is a is so

sexy. So general purpose, that it might be really difficult to come to the same. Come to a general assessment about what something is fast. But if something is safe, we might be able to develop and I think we could be able to develop that in particular context, but I think that's still always going to be around all the judgment about what level level of how fair is fair. And how safe is safe? Where I think we are making progress is coming to sort of command language increasingly common. A technique,

I agree with you that's good cause for optimism standardization, but I think the signals we pick up from from from the industry's to say that we've have her. All the debates that few have had you we have consulted the policy walk. So we are making all those contextual consideration about how should we go about about measuring quality? I think it is one thing that remains to be debated in the next year is like, when and where we started moving toward the full melodious, but that's probably

for a, for a n. N. Another another at least, an hour of conversation. Absolutely. Thank you been telling key Milian daughter Assurance. I think that's been less. So two of the same sort of like I guess the public outrage around for the rest of the eye that way saying, I think it is, but it's sort of a little bit less visible. Some of these issues in the data in the data. Weld, I placed it from an industry perspective. I'm came to your thoughts on where where we think things are currently developing and where we

think there are maybe some weed help with unity. State Insurance. Then it would be easier organizations to show themselves. It's just that there was a baby shark and Analysis. Write, thank you. And I think the interesting message thing is Assurance is about information. And therefore, Gator, insurance is dator about, they do often involves stator and about that are inside thinking about not just the coordinator that go, that way using, but what sort of information do we need about the date of that? That is that it used

and so and so there is also a big enabler of assurance about the thoughts but such as of yesterday, the analogy, I like to use about the ability of standardization is we have, I drawn accounting because it's kind of way a lot of this stuff was developed centuries ago. I bought we have Accounting Standards, but we don't have standards about how profitable is okay or not. That is your afford the market. And also up to decision-makers. We do have some minimum standards. What does insolvency the very, very basics of

it, but you're hoping that no one's really worrying about that. Like, you should be waking, well, well, away from that, and therefore, how much is gold? An ounce is not really a question of standing, but, you know, you doing it. You do need something. Okay, I might just turn to the theme of of this year cogats, which is how do we get the next 10 years, right? So, we've been so focused on the recent Austin and the car, instead button. So the time Horizon thinking about the a decade from, now, what do you think, will be the main challenges? And the man of

trinities in date restaurants in the washer and Amelia my talk with you and and decoration. Thanks, Kathy, and I hope I'm as much as much of an optimist is Maria is my oscopy. Suppose. One is around to finding ways to provide additional is around the base home device. Why do help organisations? Thank you, Billy. I think I would be a very, very try to get you to be a bottle to you about the future of AI Insurance. Then and what do you think will be the challenges and opportunities at? You have two missing? Me will spoil would like to start with the opportunities of

his pay and when, when we will have the, the new domain, knowledge of a ashore, in a bit more mature, than what we have it at the moment. I think we will finally found the one tool to help us validate. How well would you a? I r i will be talking so much, you know, Allison so many wonderful people need to feel the way I attics of being able to add the principles in the world of a high. But we don't have a tool to be able to tell us how well we have been doing. Well quality. Assurance of a high will be one of those saving grace. Is that the only that will keep us honest about how?

Well, we invited our morality, the right Morality In the small machines, but will be able to give us the chance to correct the wrong and to address the unintended consequences. But with that comes out of the challenge, the fact that no matter how good those tools are if you know. The right Behavior, where people? Are incentivizing the right way to use a quality assurance tools or any other tools to propagandhi. I will you will. You will end up having yet another set of tools that could be play. Could be play to serve any political organization

purpose and not achieve the, The Higher Goals of achieving beneficial, and responsible, and ethical. And until we figure it out, how to perform a full my audit, it will be very difficult to hold companies accountable. Especially that was to develop an application in the public domain that will impact millions of lies. And, and it's very important to make a, to push for Progress, not only on the policy, feel bad. So in the regulations, so that we can have a bit more teeth form of most forms of

audit of AI. Passionate about the topic is that finally we can close the circle, right? We started so well by operationalizing ethics of AI in different ways from principles to board prophecies and end trainings and experts will allow us to close that Loop and invalidate how well we're doing. And if we are honest and using those tools, it's only going to mean that. The next thing he has a I will be building a right way and we'll have more control over it and its outcomes. Thank you Maria. I just wanted to pick up a little bit on the language of tools there because I think people this is

one of those words that I feel like I'm constantly translating because buy tools, I think ml Engineers made out of a package that they made whereas other people like Executives who are accountable for the optical peppers for probably told you about tools, that are quite a different sort. And I'm wondering, like, maybe like standard list of questions, and definitions with off of success. And I'm curious as to where you think we're going to get to, in terms of is this something that we going to be able to order made a way and have a single to Laura or a family of tools that I'll package is a

10. L engine is put, in is Old Vine, or it sounds like you're more talking about a tool kit. Your new profession of people, who are you have the right Frameworks and and questions to be able to exercise some judgment about whether they think of it, look up look up. Obviously, we would want you all to make it of the stars on on medium size, but we know they are 10 and we're not going to be that at least for the next 10 years. That's not just jumping, Cholla to Mason before. We actually a hundred percent all we have a high degree of confidence of what

we're trying to go to make when I say tools. I mean m, a n n. A different box that you use to achieve certain objectives. I'm not necessary talking about software or called bass to that have been plenty, develop the dimensional to tell biosphere examples. I think those will be useful. And I think now let's allow ourselves to experiment with that. But if our to pick up with Kathy on hillsides, let's, let's Play cheek to assess and and and a short for the context of developing and issuing. A i and see

whatever tools they have at the moment and how to best use it to make sure that we stopped exercise, which is being delivered deployed and end and the quality of the product as as it evolves over time automation. Yes, it's it's something on the horizon, but I will stay clear of address for for for the next few years of development, new new tools for machine learning Engineers. I think of the love word for the policymakers and Regulators before before we can start thinking about standardization in this field, but

I know that I'm not that social. I'm not a computer scientist procedure that they might be disagreeing with me, but from from a organizational perspective. I think there's quite a bit of Quite a few things that we need to be thing before we start with amazing and the time. But I really wanted to say thank you and look forward to continuing the conversation.