Episode 79: Ensuring Equal Opportunities On The Business Playing Field

Addressing Unconscious Bias In AI

In AI, there's a general distinction between general artificial intelligence and narrow artificial intelligence. General is consciousness, the Terminator. Narrow is a lot narrower; it’s AI built for specific tasks accomplishment. The vast majority of narrow AI that's out there right now is machine learning. Machine learning is essentially software that learns by example.
 

Reid Blackman, Founder and CEO of Virtue Consultants and Advisory Board Member of Bizconnect joins Nick White on the podcast to discuss the biggest challenges in ethical AI and addressing bias in algorithms. Together they also explore the Bizconnect and D&B Accelerate partnership and the exciting work Bizconnect is doing to build an ethical B2B search engine businesses can trust, and ensure every business gets an equal opportunity. Reid also helps demystify AI and Machine Learning and the importance of deep learning and thinking.

Listen or click on 'subscribe' below for links to subscribe directly via your favourite podcast platform, including Apple, Google and Spotify.

(Please note that this podcast was recorded remotely.)

 

 

If you’d like to join us on the Power of Data podcast or have suggestions for potential guests, please get in touch from the button below.

Get in Touch

Read full transcript

The Power of Data Podcast

Episode 79: Ensuring Equal Opportunities On The Business Playing Field

Guest: Reid Blackman, Founder and CEO of Virtue Consultants and Advisory Board Member of Bizconnect
Interviewer: Nick White, Head of D&B Accelerate, Dun & Bradstreet

Nick 00:00
Hello, and welcome to The Power of Data Podcast. I'm Nick White and today I'm pleased to be joined on the podcast by Reid Blackman, founder and CEO of Virtue Consultants, advisory board member of Bizconnect. Hi Reid, how are you today?

Reid 00:12
Good, how is it going?

Nick 00:13
Really well, thank you, really well. By way of introduction, would you mind sharing a little bit with the audience about your background, your career to date, and how you managed to get into this industry?

Reid 00:22
Sure, so I sort of pose that that my career sort of has two halves. Although I'm currently building the second half. The first half, I was a Philosophy Professor. So I spent 20 years in academia, roughly the first 10 as an undergraduate and graduate student getting a PhD and then as a Professor of Philosophy for a decade. I primarily focused on ethics, so I've been researching, teaching, publishing on ethics for about 20 years, I left academia about three and a half years ago, and I started an ethics consultancy, essentially, I saw that businesses were getting beat up in the news media and social media for ethical misconduct, ethical problems, I saw the ethical risks that are also reputational and legal risks for companies around emerging technologies, especially AI. And I thought companies need to get their ethical houses in order, but they don't really understand ethics. I do, so I can help there. And so I left and I started that consultancy, and that's what I help senior leaders do figure out how to systematically identify and mitigate ethical risks, especially those pertaining to emerging technologies.

Nick 01:16
Great, thanks, Reid. And how long have you had your consulting business for?

Reid 01:19
About three and a half years.

Nick 01:20
Worked with any brands you'd like to share or know?

Reid 01:22
Yeah, sure. So I sat on Ernst and Young’s (EY) AI advisory board, I've worked with the Motley Fool. OpenText, I think it's Canada's largest software company, although nobody knows the name, but they're quite large. They grow by acquisition, which is why OpenText isn't a very well-known name. And a host of other companies of all shapes and sizes, from mid-size, product design, and IT companies, to startups, to fortune 500’s.

Nick 01:47
Okay, that brings me nicely onto Bizconnect themselves, a fairly new player in the market.

Reid 01:51
Yeah.

Nick 01:51
And what made it the right time for you to get involved?

Reid 01:53
So they take ethics really seriously. They are a startup, but as it was described to me, we want to have the kind of company that we're proud to tell our children about, that our children don't say, you know, 10-15 years down the line, provided that we're successful. How could you build such an ethically questionable organization? It's really important to them, and so they reached out to me and said, this is really important to us and we want to help integrating and implementing high ethical standards into everything that we do and produce. And of course, that sounded appealing to me. It's what I specialize in, generally, and so I thought, Oh, yeah, if you want to help with that on an ongoing basis, from really the very foundation of your business, that sounds great. I also thought that the business model was really interesting. What they're doing is really cool, so I hopped on board.

Nick 02:35
I guess that's one of the factors that that makes Bizconnect different. But what else kind of makes Bizconnect different to the other standard search engines and maybe saying search engines is even incorrect.

Reid 02:46
Search engine is fair enough at least at a high level. The basic way to get into it is by taking it in contrast to Google. So if you're a business and you're looking for another business to buy from, then you go into Google, and you're primarily going to get consumer based results, because Google is a search engine that's built for consumers. If you Google, you know, solar panels, for instance, you're going to get things around solar panels for your home, you're not going to get solar panels for your business, something like that. And Bizconnect is a B2B search engine, so it's built for businesses to find other businesses. So you don't get consumer facing businesses, or at least those with only consumer facing parts of their business in your search results you get businesses. And they're also businesses that are verified by Dun and Bradstreet. We have a partnership with D&B so that we have the data to know, is this a legit business or is this just some fly by night pop up? Google doesn't have that, you can get some fly by night pop up on Google, especially if that fly by night pop up is good at SEO, they can capture you. And you don't know that this company was born yesterday. Whereas on the Bizconnect site, thanks to D&B’s data, we know the business has been in existence for at least two years, we know roughly what their revenue is, we know roughly what their employee count is. So they're verified. And that way, they're more likely to be trustworthy than what shows up on Google. So that's on the if you like the searcher side, that the buyer side, the person who's looking to source a new vendor to diversify their supply chain. On the other side of things, if you're a seller, and you know, you want to show up in Google's organic search results, but you don't want to show up on page 35. Right, no one's gonna find you. So what do you do? You sponsor some keywords, you know, and you pay per click. How much do you pay per click? It varies, it depends upon what other people are willing to pay, because it's a bidding environment. So you have to bid against everyone out there for those same keywords. So again, let's go back to solar panels, you're a small to medium sized solar panel company, you've got to bid against some of the biggest companies in the world. And if you can't meet that, then you're going to go lower in the search results, which means the probability of you getting a click from a lead is smaller and smaller. So Bizconnect does this really interesting thing where everyone pays the same amount per click, there's no bidding, and it's a rotating carousel for who shows up in the first result, second result, third result, fourth result, and so on. So everyone gets an equal opportunity at being at the top. And the small and medium sized businesses aren’t squeezed out by the larger companies with massive marketing budgets that they can't compete with.

Nick 05:13
That's fascinating. I guess the only factor there that determines the inquiry is for a larger business versus a smaller one is if you happen to be at the top of the list at that time or not. And that's just the luck of the draw.

Reid 05:24
Yes, so whether you're at the top of the list depends upon, one, if you paid anything, you know, we have something like 20 million businesses already loaded, that can show up in the organic search results. When I say organic search results, I mean, the results that you don't have to pay anything in order to show up, you just show up just like in Google. If you're a small business and you paid for a package of ads, you know, sponsored keywords, then you'll show up at the first slot with the same frequency as whatever Walmart or Coca Cola or Tesla or whatever, you might not be able to compete with them in terms of your marketing budget, but you will have equal visibility to a Tesla on our search engine.

Nick 06:01
Oh that's great, so I can see exactly where your role is coming into play there Reid. I'm going to get into the industry in a little bit more detail now. And there are an increasing level of concerns about ethics and policies established around the use of AI, data and internet. What in your opinion, are the biggest challenges when it comes to ethical AI?

Reid 06:18
There's a lot I mean. If you say the biggest challenges in ethical AI, there's three main ones that everyone talks about. The first one on the top of everyone's list is bias or discriminatory algorithms. So we're talking about an AI that, quote unquote, makes decisions about who gets what. And in many cases, the who is protected subclass, or some people might put it the historically marginalized. And what they get is less than other people, less than other groups, in many cases less than what they deserve. And so there's a concern that there's going to be discriminatory algorithms. So to give you just a quick example, let's say you're distributing risk ratings for criminal defendants, this is a famous case made famous by pro publica, in 2016 publication. The courts were using risk recidivism rating scores that were generated by AI software. And what pro publica discovered is that it gives much higher risk ratings to black people than to white people, even if you hold fixed various factors like criminal history, education, class, income, etc. That's problematic. So we can talk about if you want to how AI gets to be that way, but that's one of the big issues. The second big issue is explainability, or you might have heard of various black box algorithms. Yeah. So look, AI is a piece of software, you give some inputs that pops out some outputs, what happens between the inputs and the outputs? Usually, we'd like to know roughly how it's arriving at those outputs, in a number of cases with AI. And incidentally, by AI, I mean, machine learning here, we could talk about what that means that if you like, with explainability issues, we don't know why we're getting the output that we're getting. Why are we getting these risk ratings? Why was this person denied a loan? Why were they denied a mortgage? Why were they denied a job interview. The AI just said, so seems to be a particularly unsatisfying explanation and we want to know what's going on between the inputs and the outputs. But if it's a black box, and we can talk about why it's a black box, if you like, when it's a black box, we don't know why those decisions, quote unquote, decisions are arrived at and that looks problematic. Third thing, privacy violations, lots to say here. In short, a lot of AI is powered by people's data, which organizations are highly incentivized to collect as much personal data as they can to make their AI more and more accurate that just to say that it incentivizes privacy violations. And the fourth thing that nobody really talks about is all the other ways that you can screw up using AI ethically. It's a problem because if we just focus on price, explainability and privacy violations, people tend to ignore all the various other ethical risks that can be realized, depending upon the use case for your AI. And that's really the source of the greatest ethical risk is your use case, are using it to power or to make possible self-driving cars? Well, then you might want to watch out for killing and maiming that seems to be a pretty significant ethical risk. Anyway, I can go on, but I don't want to ramble too much.

Nick 08:59
No that's just great Reid. I wanted to come back to your comment there at the beginning about the difference between AI and machine learning. Can you just explain for our listeners your definition of machine learning in the context of Bizconnect?

Reid 09:11
Yeah, sure. So one thing just to say is that people will talk about AI a lot and most AI around now was machine learning. So in AI, there's a general distinction between general artificial intelligence and narrow artificial intelligence. General is consciousness, the Terminator, etc. Narrow, is a lot narrower, it’s AI is built for specific tasks accomplishment, particular tasks, the vast majority of narrow AI that's out there right now is machine learning. Machine learning is essentially software that learns by example. So how do I know this is a picture of Nick? Well, you know, I've been given a bunch of pictures of Nick, I've analyzed those pictures and now I sort of know what Nick looks like. So if you give me a new picture, Nick, I'll label it Nick. That's a case of machine learning, I've learned by example. In the case of Bizconnect, we are not at this moment deploying a tremendous amount of machine learning, but we will increasingly use it to get the kind of data that we need about businesses to give the best results that we can to our searchers. We might give recommendations, you know, down the line, we might give recommendations to people like this company has these special offers that you might find of interest. And that will be because we have some level of knowledge about the searcher that we've been transparent about what data we're collecting about them. And then we might make recommendations about what they should check out such and such’s special offer. At the moment, we're not doing a tremendous amount of machine learning. And moreover, we are not, in contrast to say Google collecting a tremendous amount of data about what the searchers are searching for, how they're searching about what their phone number is, what their email is, you know, where else they go on the internet, yada, yada, yada, we don't do that stuff. And because we don't collect all that personal data, we're at a much lower risk of violating anybody's privacy.

Nick 10:47
Thank you, Reid. Thanks for explaining that for me and for the listeners. My next question kind of goes back to the biggest challenges faced by the industry in ethics and policies around AI, data and internet. So if we looked at those three big topics, algorithm discrimination, explain a bit of privacy violation? How do you think the industry overall is going to try and tackle these and solve for these concerns?

Reid 11:10
Well, one hopes that eventually, they will, they'll put it in something like an AI ethical risk program. These things need to be dealt with at an organizational level. It's not just something that a data scientist can fix, or an engineer can fix. It needs a robust program, processes, practices, etc. So yes, you do need to say data scientists, engineers, product developers, to be involved in various ethical risk mitigation procedures. But they also need things like the ability to elevate certain kinds of concerns and decisions to another body and another entity within the organization, someone who has a kind of expertise that say data scientists don't have. So for instance, let's say that you are trying to figure out or vet whether your AI model is biased against a certain subpopulation some way, you might ask, okay, is the way that it's distributing these goods or services across these subpopulations, is this fair? Is this an equitable way of distributing things? And that's a substantive ethical decision to make. Now there are metrics, quantitative metrics that the computer science literature has that says, look, here's how you measure fairness. And there's something like two dozen plus so-called metrics for fairness. But those metrics are actually incompatible with each other, you can't score well on all of them at the same time, by doing well, according to one metric, you're failing, according to another metric. And so then the question becomes what's the appropriate metric for this particular use case. And that's not a decision that a data scientist is well suited to make, they standardardly lack the kind of experience education, training, background, etc, to make a substantive, ethical and business decision about what's the appropriate metric for fairness for this particular use case. And so they need the means by which they can elevate those kinds of questions to some other kind of entity, body, individual within the organization. In short, I mean, I think that ship has sailed, but in short, what companies will need to do is create a means by which there's some frontline of defense, but also a means by which they can elevate that the higher levels of the organization.

Nick 13:05
Right, thank you Reid, I appreciate it. I'm interested in your perspective on those unconscious bias decisions that we make on a day to day basis.

Reid 13:13
Okay, let's separate out two questions. One is, what role does unconscious bias play in the development of AI? And the second question is, is it the case that ai standardly mitigates bias in comparison to humans?

Nick 13:26
Perfect.

Reid 13:26
So to answer the first question, what role does unconscious bias play in the development of AI? I mean, we're going to need to learn a little bit more about what you mean by unconscious bias. The term is using a variety of ways. For instance, you might be thinking of something like implicit bias as research by psychologists at Harvard, although that research has been called the question of whether there actually is implicit bias. And even if there is implicit bias, what the causal connection is between implicit bias and action, planning to say you're implicitly biased, it's another thing to say that that bias actually manifests inaction in a systematic way. Those things are being researched and to say the least, it's controversial. There's evidence that there's no implicit bias, or that the test for implicit bias is flawed. And there's evidence to think that the causal connection between implicit bias and action is questionable. Now, that said, all that stuff is, if you like, on the on the psychological side, and it's not clear at all, that implicit bias, unconscious bias, etc., plays a deep role in how AI is developed. That's not the same thing as saying people's assumptions don't play some role, but I'm hesitant to call all assumptions biases. I think that those are two different things. Is that a too overly convoluted answer to the first question?

Nick 14:37
Not at all Reid, thank you.

Reid 14:38
Okay. The second question is, in a way more vexed. And that's because there are times when AI actually exacerbates biases instead of mitigating or replacing them. So for instance, one standard way that people like to talk about is that you can have certain kinds of training data for your AI the examples by which your AI learns that when being trained on that data can lead to discriminatory outputs. So for instance, I referenced earlier, the software that discriminates against black people in risk ratings for recidivism in the criminal justice system, it was trained on the historical data of the US’ criminal justice system. But as anyone basically familiar with our criminal justice history knows that history is riddled with racist problems, racist policies, racist behaviors, etc. So instead of making your recidivism risk rating software less biased, because it's AI, you're actually training it on the historical biases, the discriminatory policies, actors, etc., that are embedded or reflected in the data that you're using to train your AI. So you're actually scaling the bias rather than mitigating the bias that instance. And there are actually loads of examples of that. So here's a pretty well-known one, you know, Amazon receives 10s, of 1000s of applications for jobs. And it's hard to get an HR department to look at every single one of those. So they had a reasonable idea. Let's train some AI software to read the resumes and figure out which ones are interview worthy and which ones are not. So what did it do? It took, let's say, the past 10 years of hiring data, it took the last 10 years of resumes and said this resume led to an interview was judged worthy of being deserving of an interview by some human and this one was not, this one was this one wasn't. So you do that over 10,000 resumes or so or actually hundreds of 1000s of resumes, because we're talking about many years of Amazon applications resumes, and they get, you know, 10s of 1000s a day. So the software learns by example. So what did it learn, what was the lesson, the software learned about the kinds of resumes that lead to interviews around here? And what it learned is, we don't hire women around here. So what it started doing is if you'd like, red lighting, as opposed to greenlighting the woman's resumes, so if it said, women's NCAA basketball, red light. So you might say, Oh, this is easy. Just tell the AI tell the software ignore the word woman in a person's resume. The problem is that there are other words in a resume that highly correlate with that variable. So there are words that highly correlate with being a woman are words that correlate with being a man. And so then what the AI learns is Oh, we don't hire people who use that kind of word, or we hire people who use this kind of word. So for instance, execute, men tend to use the word I executed on some strategy, that phrase is used more by men than by women. And so this is called proxy bias, it still takes women's resumes out still right, let them the Amazon team worked on this for two years. And to their credit, they actually throughout the project, they scrapped the project because they couldn't figure out how to sufficiently de-bias or mitigate the bias in their software. And that's, of course, because of the historical data. Now we can have a discussion about how did the training data get to be that way. And there's a variety of explanations, but the short of it is, the pattern that it saw in the past 10 years is we don't hire women around here, and then started applying it to all the new resumes. And so they had to scrap it.

Nick 17:55
Unbelievable, Reid, we can have a series of podcasts, I think, based on your expertise in this area, I thank you so much for indulging my personal questions there, really appreciated. I'm going to move back onto topic which is you know at D&B, we pride ourselves at partnering with the right organizations, Bizconnect is one of those and you're now at D&B Accelerate partner. For those who missed earlier on, can you just explain to our listeners how you're using the D&B Data Cloud in Bizconnect?

Reid 18:21
Yeah, so one of our main goals is that we be trustworthy, and that we give results that are trustworthy. That the people who are looking to diversify their supply chain, looking for a new vendor, etc., that they can trust the results that they get in a way that they can trust the results from other options that are on the market. So the problem for us was how do we make it the case that we have information about businesses that people can trust. And so that's why this connection, turned to D&B and said, Hey, listen, let's have some kind of partnership here. Let's figure something out. You have data about companies from how long they've been in existence, to whether you know, whether they're minority owned, to what their revenue looks like, to the quantity of employee looks like their headcount looks like. And that's useful information when you're trying to make a decision about whether you should purchase from a vendor. And so it made sense for us to partner with D&B, because you've got that kind of information that lends itself to trusting or deciding not to trust some organization.

Nick 19:17
Thank you Reid, really, really helpful. I guess the longer-term vision will be that this will expand globally. I think, as I understand it, at the moment it’s a North American facing proposition. But you're looking to take this globally?

Reid 19:29
Actually, you said what I should have said before, which is that it is a global search engine. So it already is global, and its global by virtue of D&B's data. So let's say you want to find a new supplier. Let's say that you're going to look around in Asia, you might look at China, you might think, Oh, you know, the supply chain out of China is problematic. Freight has massively increased out of China recently. And so you want to look elsewhere, maybe you want to look to Vietnam, maybe you want to look to Korea, maybe want to look to Japan. But do you know what is showing up? Let's say you go into a Google you search for I don't know, manufacturers in Vietnam. Do you know what you're landing on? Do you know anything about those companies? Are they legit? Do they have a history? What's their background like? Can we trust them? It's hard. But because you've got D&B’s data, the 20 million companies that we have in our search engine are verified by Dun & Bradstreet. So we have a sense of who they are, we know how long they've existed roughly what their revenue is, what their employee headcount is, we know what kinds of orders they might be capable or incapable of handling. And, look, there are a lot more than 20 million global businesses. But we've whittled that down to 20 million based on criteria that we think are conducive to justify judgments that the companies are trustworthy, at a bare minimum, they're worth looking into. And so we do have a global search engine where the search engine turns out results from around the world where those companies have been verified by D&B. And I'll just add that we have a filter that you can say, look, I just want to look at India, I just want to look in Vietnam, or I want to look in Vietnam, and India, etc. It's not that we aim or that we aspire one day to be a global B2B search engine we already are. And that is in no small part due to our partnership with D&B.

Nick 21:12
Brilliant, thank you. Well, this is where D&B does their marketing piece. And I just wanted to add that the D&B Data Cloud currently holds over 420 million entities.

Reid 21:21
D&B does have data on more companies that we include in our search engine and that's only because we wanted to include companies of a certain stripe. So for instance, we wanted them to be in existence for at least two years, you have data on companies that have been in existence for less than two years, we wanted companies to have a certain minimum level of revenue so that one could reasonably believe they could handle larger orders from American buyers, you have data about people who have revenue of something like less than 250,000 a year, we included only those that have 250 K or more per year. You have data about companies with you know, 1-2-3 employees, we have a certain minimum requirement for how many employees the company has to have. So for now, we've limited it, because we want to make it as verified and as trustworthy as we can, given all the information we have now. But lucky if it turns out that buyers are you know, they want more companies are willing to take more risks, or they think it's not more of a risk, then we'll expand the results appropriately. But we feel that the 20 million we have now is robust and it's a good place to start, at the very least.

Nick 22:20
Absolutely, thanks Reid. We're almost to the end of the podcast. And this is the part where I like to get a bit of a view on the people that have helped to shape you. So are there any particular mentors in your career, and any pieces of wisdom or lessons that you've learned that you'd like to pass on to our listeners before we wrap up this afternoon?

Reid 22:38
So I'm not one that big on mentorship, I can't say that I've had any great mentors. I will say that there's a moment that always sticks with me, which is when I was an undergraduate, I was a sophomore at the time, a second year student and a senior level class, it was a seminar on the philosopher Friedrich Nietzsche. And the Professor Maude Murray Clark was talking and lecturing and fielding people's questions and cetera, et cetera. And she was amazing and I thought, I want to be like that, you know, she has all of this knowledge, she has all this deep understanding of Nietzsche, of ethics, etc. And I wanted to have it. Not in a sort of jealous way, just I want to be like that, I want to be able to have that kind of depth of knowledge at my disposal. She actually was a mentor to me, but she was also just inspiring by virtue of her deep expertise. So that's what I would say in terms of something like mentorship or inspiration, I wanted to have that level of understanding, especially because I was intrigued by the topic of ethics. And I knew that I didn't know so much and I was sort of lost, like most any undergraduate is on that topic and I wanted to understand it more. That's the first thing. The second thing to say about my piece of advice is, do the research, do the reading. I specialize in ethics and AI ethics. There are a lot of people out there who are talking about ethics of AI ethics, frankly, I think a lot of the stuff out there is not very good. And that's because people don't, they don't do the reading, you know, they were aware of the issues of things like discriminatory algorithms or black box algorithms. And then it just sort of stays there very surface level. I think that there's a tremendous amount of value to you know, doing the reading and not just reading the news, or reading medium posts, which can be useful, but actually read the research, go get some academic articles and read that research, even read the research that's coming out of say, Microsoft or Google, let's say you're into tech ethics. There's really smart people doing really deep work and you don't get to be a deep thinker about these things, not just deep thinker, but an effective thinker, that is to say, a thinker who can translate those thoughts into action without really understanding of the subject matter. And you can't do that without taking the time to do the research.

Nick 24:42
Great, I think that's wise words I'd say Reid. There is so much information available today that we often kind of just scratch the surface we get enough to be dangerous, not to be competent.

Reid 24:53
Yeah, people don't seek out experts.

Nick 24:55
Absolutely not Reid, and I can tell from our conversation how your deep subject matter expertise in all things, ethics here. It's been really enjoyable for me personally. You've given me food for thought in a number of areas. Thank you very much for your time this afternoon. And thank you for joining The Power of Data Podcast.

Reid 25:10
No, my pleasure. Thanks for having me.

Nick 25:12
Thank you. Goodbye.