Latest / The Prompt / Ep. 491 | Balancing AI and Human Judgement in Decision Making
Transcript
- 0:01This is on base your one stop shop for solving your most
- 0:04pressing B2B go to market challenges.
- 0:07Each week, host Chris Moody and Paul Gibson talk to sales and
- 0:09marketing leaders to get you in depth insights and creative
- 0:12strategies to set your revenue teams up for success.
- 0:18Hey, everyone. Welcome back to Own Base.
- 0:20I'm your host Chris Moody, and today I'm really excited to be
- 0:23joined by Mana Yonescu on marketing with Mindfulness.
- 0:27Mana, I know you have an incredible background for our
- 0:30listeners. If you could tell everyone a
- 0:31little bit about yourself. Yeah, hi.
- 0:33I'm so happy to be here. I lead the digital marketing
- 0:37practice at Exos Bank and prior to that, I LED my own agency for
- 0:42almost 13 years. I am very, very fascinated with
- 0:46how we use data to make decisions and with today's
- 0:49topic, AI and marketing. Awesome.
- 0:52We talked about this extensively when we met before and it was
- 0:56really educational for me. I took great notes and one of
- 0:58the things that we talked about was balancing the speed and
- 1:02accuracy benefits of AI, but making sure that you have
- 1:05ethical and unbiased decision making.
- 1:07So if that's something that we should all be focused on, could
- 1:11you tell us a little bit about how we do that?
- 1:13It's important to implement a few key and fundamental
- 1:17practices. Obviously, everybody knows about
- 1:20the diverse data sets, and you'll hear data scientists talk
- 1:24a lot about how the outcome is the output is only as good as
- 1:28the data, how clean the data is, how good the data is, but also
- 1:32how wide and diverse and representative it is of all
- 1:36demographics so that we minimize bias.
- 1:39The other important practice is regular audits.
- 1:43So we don't, we don't want to do that right.
- 1:45We we want to set it and forgive it and then move very fast
- 1:50because we want to make those dollars that ideally these
- 1:53predictions or the output of our algorithms or our models can
- 1:58help us with. We want to move very fast to
- 2:00making money, but I believe that there's a set of governance
- 2:06models or frameworks that we need to put in place so that we
- 2:09can do this well. And regular audits is one of
- 2:12those. And don't worry markers, You
- 2:14don't have to look at the code, there's yours.
- 2:18Marketing scientists should be able to help you with that.
- 2:20But it's, it's really important to do it.
- 2:22Transparency is another one, is another important principle or
- 2:26practice to maintain transparency and AI, You have to
- 2:30again, team up with your marketing scientists, but also
- 2:34know that they, a lot of marketing scientists are very
- 2:38good at illustrating concepts. And they should be able to
- 2:42explain to you how it works, what went into the model, how
- 2:45it's making decisions, how they're ensuring that decisions
- 2:50are as accurate as possible. And then all of this involves
- 2:54human, human oversight. A lot of companies have
- 2:57governance teams in these frameworks that I'm talking
- 3:00about to make sure that your models are aligned with ethical
- 3:04standards, that those audits will happen, that all
- 3:08stakeholders are clear and understand how it all works and
- 3:13what risks you may have built into the system.
- 3:16And so that's not the fun part, That's the procedural part, the
- 3:20structural part. But it's important that happens
- 3:24because there's a lot. There are a lot of risks and
- 3:26there's a lot, a lot of risks to our reputation.
- 3:29There's a lot of risks to making predictions that do not fit a
- 3:34new data set that we bring into the picture or, or just help us
- 3:38waste money. And we don't want to do that.
- 3:41So those are the key practices I would recommend for ensuring
- 3:45that our models make ethical and unbiased decisions.
- 3:49I love that. I mean, we, we definitely don't
- 3:51want to waste money that comes up in almost every conversation.
- 3:54But the notion of transparency, I, I think this is one that not
- 3:59everyone talks about as much as they should, but there are times
- 4:02where we're blindly trusting the tools that we use and the
- 4:06recommendations and some of the black box type terminology that
- 4:09gets thrown out a lot. What are some of the fundamental
- 4:13questions that us as marketers should ask to validate outputs
- 4:16and assumptions of AI or predictive models?
- 4:20It's interesting that you did call out transparency because
- 4:22I'm a huge believer in it. And there are a few questions
- 4:25that you can ask. And I'm going to teach you one
- 4:27question to ask marketers, pay attention.
- 4:29There's one question you can ask, and one, you're going to
- 4:33sound really smart, I promise. Two, it's actually it's a
- 4:38question that could summarize this entire conversation we're
- 4:41having. And the question is whether
- 4:43whether somebody's trying to sell you the new shiny tool of
- 4:47AI or whether your own data scientists or internal teams are
- 4:51building models. The question is, how do you
- 4:55watch out for overfitting? How do you address overfitting
- 5:00or just say, hey, what's overfitting?
- 5:03Explain to me overfitting and thus how can we avoid making
- 5:08that mistake? And that could sum everything.
- 5:11What we just talked about the audits, the transparency, the
- 5:15governance and it's going to make you sound smart.
- 5:19And what overfitting is basically when we typically when
- 5:23the model we trained had some noise in the data set and it's
- 5:28learning based on the noise, not just on the actual data.
- 5:32And so we may try to apply it to a real case scenario and it's,
- 5:38it's not working so or it's working accurately.
- 5:41And if we don't have the right checks and balances in place, we
- 5:44may not even know. We may not even know that we're
- 5:46over fitting so that the models are over fitting.
- 5:49So very, I think if you ask that one question, if they know how
- 5:54to answer the question, now you're having a conversation.
- 5:56If they don't know how to answer the question, we have a problem.
- 6:00So things to look at are your data sources, Where does the
- 6:04data come from? Do you have enough data?
- 6:06Is it reliable? Is a representative?
- 6:08The data, you know what's implied in there, right?
- 6:12The data diversity you want to look at performance metrics and
- 6:17basically what how are we going to measure the models
- 6:20performance bias fairness, we talked about that.
- 6:24And then there's scenario analysis.
- 6:27And again, if you ask about over fitting, you're going to get
- 6:30into a conversation about about scenario analysis.
- 6:33How does your model perform under different scenarios?
- 6:37And can we stress the model? Can we introduce edge scenarios
- 6:41and you know, how's the model going to perform?
- 6:44And then validating and testing, how do we validate, how we how
- 6:46do we test? What validations validation
- 6:49methods are we going to use? The interesting thing is, you
- 6:52know, a couple years ago I decided that I wanted to learn
- 6:55about this and I joined the MIT six month program in machine
- 6:59learning. And what was very interesting
- 7:03was that the professors were very good illustrators.
- 7:07They were very good storytellers.
- 7:10And I had forgotten. I'd study statistics, but I've
- 7:13forgotten how visual it can be that a lot of the concept, a lot
- 7:17of the math, so to speak, is spatial.
- 7:20It's geometry. And so this is not just about
- 7:25the numbers. Good storytellers should be able
- 7:28to help you visualize what they're talking about and
- 7:33explain to you how they're going to do this scenario analysis,
- 7:37how they're going to. They'll talk about things like
- 7:40regularization or early stopping of the model to avoid
- 7:45overfitting. And you can ask them what does
- 7:47that mean? They typically should be able to
- 7:50to explain to you. So it's not all math.
- 7:53Anybody should learn how to explain it.
- 7:55It's going to be very important to learn how to explain it to
- 7:57stakeholders because they're going to be the ones paying for
- 8:00these resources. They need to understand, they
- 8:03need to understand the risks, need to understand the benefits.
- 8:06And I'm a huge proponent of marketers learning become
- 8:11marketing scientists. I'm a believer in the future of
- 8:14marketing, of future of the profession being more
- 8:18scientifical than creative. And so these concepts are very
- 8:22important to understand and they're not hard, they're not
- 8:25complicated. And if all those fails, just ask
- 8:28the question, tell me about how we're going to avoid
- 8:31overfitting. I think that's a great question.
- 8:34I'm going to write that one down.
- 8:35I mean, there are some things that I think many would be
- 8:38surprised about with any predictive model if they start
- 8:43to ask questions because there are times where you assume that
- 8:47the data is updating all the time and that it's watching
- 8:49what's happening. And you know, we've seen huge
- 8:52differences in some tools where it's updated every six months
- 8:54versus nightly, where to your point, I mean, it could be based
- 8:58on incorrect assumptions. And you also hit on the notion
- 9:01of reliability. I'm curious when you think about
- 9:04data size and diversity, how does that impact the reliability
- 9:09of AI model outputs? I.
- 9:11Mean, it's a great question. So one of the things that that
- 9:15data scientists do when they build these models, we think of
- 9:17them as a model. But in fact, you build models
- 9:21upon models, you use models to test models, you use models
- 9:25stacked upon models and different statistical methods to
- 9:30check the reliability of the outcome of the outputs.
- 9:34It's not just one, we think of it as a model, but it's
- 9:37algorithms upon algorithms. And data scientists are going to
- 9:40like and hate mail right now, but I'm trying to really
- 9:43simplify this for for marketers. And so again, if you open the
- 9:48conversation with the data scientists, they should be able
- 9:52to explain to you how I'll explain it like this because
- 9:58it's an interesting visualization.
- 10:00Those those of us who have started using ChatGPT, we know
- 10:03that we give inputs and restack the inputs.
- 10:08So we'll say, you know, please analyze this piece of content.
- 10:12Now give me the three top themes, whatever it may be,
- 10:16right? So one by 1:00 we, we give it
- 10:19inputs to get to the outcome that we're trying to get to.
- 10:23So a lot of times the models are built like that with the series
- 10:26of inputs that we need to create a system that with predictions
- 10:32that are as reliable as possible.
- 10:35Yeah, that's great. I would love to come back to
- 10:37transparency. It's one of my favorite words.
- 10:39And you know, we talk about transparency, believability,
- 10:43explain ability, all, all very similar concepts that we're
- 10:47trying to head out here. But I'd love for you to talk
- 10:51about in what situations is it necessary for marketers to have
- 10:55visibility into the why behind AI recommendations.
- 10:58And you know, the caveat to that, which I would flag for
- 11:02many folks listening, is there are plenty of AI recommendations
- 11:07that they're probably acting upon where they don't know the
- 11:10why. But I'd love for you to hit on
- 11:13why that's so important and how we should start to think about
- 11:16that visibility. I think fundamentally these
- 11:20models are are tools and they cannot operate without human
- 11:26oversight. And so the human being in the
- 11:30loop is very important, and that's part of transparency.
- 11:35And in that it's really important because ultimately the
- 11:39outcomes of that model affect humans, can affect humans.
- 11:44The classical example you will hear in school is the prediction
- 11:49of an illness and how you'd rather predict an illness and
- 11:55then you find out that no, I don't have that illness, then
- 11:59miss it and risk that that person's well-being.
- 12:04So high stakes decisions is one of the reasons why it's very
- 12:09important that we understand what's behind these
- 12:12recommendations. And that's important one and
- 12:15other high stakes decisions are really important business
- 12:18decisions or things like loan approvals in our universe and
- 12:23financial services. And so you can't have a black
- 12:27box. You mentioned the the word black
- 12:28box. Now black boxes right now are
- 12:31sexy because it feels futuristic.
- 12:34I don't know what's behind this. So it must be of the future.
- 12:38I must be in the future type of thing but I find knowing sexier
- 12:45than not knowing so and it could help us avoid mistakes when it
- 12:49comes to high stake decisions. Obviously there's compliance and
- 12:53regulatory considerations, then bias detection.
- 12:58You know, I think in most cases we shouldn't avoid bias just
- 13:01because just for regulatory reasons.
- 13:04It's just good practice and typically we can get better
- 13:08results, better make more money if we actually avoid bias.
- 13:14Then there's trust this trust of the stakeholders into what we're
- 13:18doing, the trust of our customers.
- 13:21But you can go back to the example of of a medical
- 13:23diagnosis, right? You're patients need to be able
- 13:27to trust, so you need to be able to explain to them, you need to
- 13:30know the why behind that recommendation for them to be
- 13:33able to trust and believe that, Oh yes, I actually there's no
- 13:38indicator of high likelihood indicators of cancer wherever
- 13:43maybe. And then probably one of the
- 13:46most important ones is improving the model.
- 13:49So you have to be able to understand why a model is giving
- 13:52you the outputs given to be able to refine it, to understand if
- 13:56you're missing an expansion of that data set, if you're missing
- 13:59scenarios that the model is not learning from, because as you
- 14:03said, it has to evolve. You can't just update it once
- 14:06every two years. You definitely can't set it and
- 14:09forget it. Yeah, that was perfect.
- 14:11I think you nailed AT shirt design there too.
- 14:14Knowing is sexier than not knowing.
- 14:16We're gonna have that as a poll quote, guaranteed.
- 14:18That will be the social clip. I mean, because it is so true.
- 14:23I mean, if we think about a lot of folks who are listening to
- 14:26this podcast, many of them are making decisions about which
- 14:30accounts sales should be targeting and they may not
- 14:33really understand why those decisions are being made.
- 14:36So to your point, we need to understand that so we can have
- 14:39that conversation. And I know you had on loan
- 14:42approval and some of the the intricacies of financial
- 14:46services. What are some of the practices
- 14:49or regulations that exist in different industries like
- 14:53financial services to make sure that we have fair non
- 14:56discriminatory use of AI? Yeah.
- 14:58So in financial services there's a number of regulations to being
- 15:03fair Credit Reporting Act and Equal Credit Opportunity Act.
- 15:07And both of those require that AI models used in lending and
- 15:11and credit decisions are fair and non discriminatory.
- 15:16These acts have always required that.
- 15:18But basically what this is telling us is you can't have a
- 15:21black box. And for example, I'll give you a
- 15:24real life example where you could step into a black box and
- 15:27that's Facebook advertising where the algorithm may make
- 15:31decisions for you and based on your limited, limited data set
- 15:36and show ads to a limited demographic.
- 15:41Now Facebook has introduced these special categories, so
- 15:45that doesn't happen. And in those special categories,
- 15:48you can't target based on certain criteria.
- 15:51But from a regulatory perspective, we need to ensure
- 15:54that we understand how our models work and make sure that
- 15:59the decisions are fair and non discriminatory.
- 16:02In healthcare, you have HIPAA, you know, with that mandates
- 16:05also fair use of AI and medical decisions.
- 16:08And then in terms of general practices, I mean, I talked a
- 16:11little bit about the key practices earlier on the diverse
- 16:15data sets, the ethical guidelines, the governance.
- 16:18There's this concept of so, and by the way, I thought that the
- 16:21T-shirt that was going to be, you know, tell me about
- 16:24overfitting. I thought that was going to be
- 16:26the T-shirt. But here's another catch phrase.
- 16:29Algorithmic accountability. That's an important general
- 16:33practice. So we talked about transparency,
- 16:36algorithmic accountability, bias mitigation.
- 16:40We talked about that earlier. And then, you know, ethics and
- 16:43guidelines. So a lot of that is governance
- 16:47is structural. It's not the sexy stuff, but we
- 16:51got to do the not sexy stuff to get to the sexy stuff.
- 16:54And so my mind is blown all the time by how little attention we
- 16:58pay to that. And it's really tricky when we
- 17:02are being sold so many tools and systems under the label of AI.
- 17:06One, we haven't gotten to true AI yet.
- 17:09And two, most of the times we have no idea what's behind it,
- 17:13if it's even a solid machine learning algorithm or set of
- 17:17models. So it's very important to ask
- 17:20those questions and put the controls in place and then you
- 17:24can move faster. Yeah, I love that.
- 17:26And I mean such an important point.
- 17:29As we start to wrap here, I just want to stress the point you
- 17:32just made. There are folks posting on
- 17:34LinkedIn every single day about various tools that are powered
- 17:38by AI. That may be eliminating job
- 17:42functions. I mean that it's that level of
- 17:45what people are saying when that's usually not the case and
- 17:48we may not even understand everything that's happening.
- 17:51When we think about the the modern evolution of technology
- 17:55and go back to the early 1900s, where people were afraid that
- 18:00the new machines, the new equipment in factories were
- 18:03going to replace people, and when in fact it created jobs, it
- 18:08led to an industrial revolution. We're looking at something
- 18:12similar here. I don't think AI is going to
- 18:15replace people. I think it's keep creating the
- 18:17opportunity for more of us to specialize in new things.
- 18:21There's going to be more marketers who are going to learn
- 18:24about machine learning. And one thing I wanted earlier
- 18:28when we talked, when we were talking about transparency and
- 18:32even overfitting, I'm going to keep, I'm going to keep bringing
- 18:36that up. I'm going to be a very happy
- 18:38person today when every marketer asks the overfitting question.
- 18:43But you know, I always think of this, what we're doing here or
- 18:47how to explain to marketers how these models and algorithms work
- 18:52is that imagine if, if aliens tried to learn how humans speak
- 18:58and not just how we speak, but be able to predict how our
- 19:02language is going to evolve. And for that, they have to go
- 19:06back historically, they have to talk to a bunch of us, different
- 19:09languages, different accents, what you may have it and, and
- 19:14then it also. And so as part of that, I was
- 19:17also reminded how we all know those people who tried to teach
- 19:22us a curse word in the language without telling us that it's a
- 19:26curse word. So the noise in the system that
- 19:30we need to avoid by using a good data set, a wide enough data
- 19:36set, a diverse enough data set is to if we were to build this
- 19:41to teach aliens or the aliens were using this to teach how to
- 19:44speak and how to speak in the future, that noise can really
- 19:49affect how that model works. And so that's my default when I
- 19:54try to explain to people the importance of of the data that
- 19:59we fit in the system, the controls that we need in place
- 20:01of this transparency we need. It's like, you know, learning a
- 20:06brand new language. Trying to go through the
- 20:09vocabulary, understand the grammar structure, all of that.
- 20:12And there's so much potential noise in there, and it takes so
- 20:15long to learn. And you meet new people who have
- 20:18different accents and who have their own ways of saying things.
- 20:22And you're learning again from them.
- 20:25And you're expanding that data set and it's this evolution of
- 20:29your knowledge of that language. And then some people go for PhDs
- 20:33and then use historical data and try and predict how language
- 20:36will change in the future. But that wraps in a lot of the
- 20:40concepts that we talked about today, the key practices, the
- 20:42data diversity, you know, the regular audits.
- 20:45When you learn the language, you may have to take some tests and
- 20:48show what you've learned, that transparency and the ability to,
- 20:53you know, somebody should call you out if you try to teach me a
- 20:56curse word and so on. So it's truly not as much of A
- 21:02black box as many are trying to make us think.
- 21:05It's very comprehensible. We just have to go through the
- 21:10practice to be able to have that access for visibility knowledge
- 21:17and then we can the the models, the algorithms to work to speed
- 21:22up things for us. That is a great point and great
- 21:25example. And for those of us who have
- 21:28children, I mean the language changes every few years.
- 21:30My 12 year old says things that I have no idea what they mean.
- 21:33So to your point, I would not want that in an e-mail.
- 21:36We need to make sure we have that reliability and that good
- 21:39data set. We're coming up on time.
- 21:41There's two very quick questions we'd love to nail in a really
- 21:46short amount of time. The 1st is who are the folks
- 21:49that are inspiring you and the B to B space that you recommend we
- 21:53bring to the show? Yeah, Jill Rowley Think she was
- 21:57the chief something officer at Marketo at one point.
- 22:01John Miller Sangram. Yep, Sangram it go to market
- 22:05partners, yeah. Yeah, in the content marketing
- 22:09EBM world, I would say Andrew, Crystal, Dina, I had to throw a
- 22:14in a Chicago shout out there. Those are some of my favorites.
- 22:18Yeah, I love those. I've worked with a few of them,
- 22:20and Andy is one of the smartest folks I've ever met.
- 22:23So yes, great people there. And then the most important
- 22:26question, how can everyone get in touch with you after this
- 22:30episode? Well I'm on LinkedIn and pretty
- 22:32much on every social network is Mana Mika MAN AM ICA.
- 22:39Every social network except TikTok you can find me by that
- 22:43handle, including LinkedIn. Awesome.
- 22:45Well, thank you so much for your time today.
- 22:48It was a great episode. Really appreciate you joining.
- 22:50Thank you. Thanks for joining us for On
- 22:54Base. If you enjoyed this episode,
- 22:56please be sure to share, rate, review and subscribe on Spotify,
- 22:59Apple, YouTube, or wherever you get your podcasts.
- 23:03This show is brought to you by Demand Base, the only account
- 23:05based experience platform built for sales and marketing
- 23:08alignment.