Building Trusted AI Systems: A Fireside Chat with IBM

I recently had the opportunity to sit down with Lauren Frazier from IBM to discuss how we go about building trusted AI systems in a fireside chat livestream. We covered a ton of ground.

Can’t see anything? Watch it on YouTube here.

What is Fair?

Fairness is a difficult subject to tackle, because people have many different ideas of what constitutes fair treatment. In the context of things like bank loans, citizens’ rights, being hired for a job, etc. what is fair?

The dictionary definition is both straightforward and unhelpful:

impartial and just treatment or behavior without favoritism or discrimination“”

What constitutes fairness? This is where things get really messy. Broadly, there are four different kinds of fairness, and each has its own implementation, advantages, and pitfalls:

  • Blinded: all potential biased information is removed, eliminating the ability to be biased based on provided data
  • Representative parity: samples are built to reflect demographics of the population
  • Equal opportunity: everyone who is eligible gets a shot
  • Equal outcome: everyone who is eligible gets the same outcome

For example, let’s say we’re hiring for a data scientist, and we want to hire in a fair way based on gender. We have a population breakdown where 45% identifies as male, 45% identifies as female, and 10% identifies as something else or chooses not to identify. With each of these types of fairness, how would we make the first step of hiring, interviewing, fair?

  • Blinded: gender and gender-adjacent data (like first names) are removed from applications.
  • Representative parity: our interview pool reflects the population. If we’re in China or India, there are 115 males for every 100 females, so our interview pool should look like that if we’re using representative parity.
  • Equal opportunity: we interview everyone who meets the hiring criteria until we reach 45% male, 45% female, 10% other.
  • Equal outcome: we interview everyone until we have second-round candidates in the proportions of 45% male, 45% female, 10% other.

Each of these scenarios has its drawbacks as well, either on excluding qualified candidates or including unqualified candidates.

  • Blinded fairness doesn’t address underlying structural fairness problems. For example, if women feel excluded from data science jobs, then the pool of applicants would still reflect an overall bias, blinded or not.
  • Representative parity doesn’t address the structural fairness problem as well, though it does do slightly better than purely blinding data.
  • Equal opportunity may exclude qualified candidates in the majority, especially if there’s a substantial imbalance in the population, and potentially could include lower quality candidates in the minority.
  • Equal outcome may achieve the overall intended quality benchmarks but could take substantially longer to achieve the result – and depending on the imbalance, might not achieve a result in an acceptable timeframe.

Why does any of this matter? These decisions already mattered when it was humans like you and me making decisions, but they matter much more when machines are making those decisions based on algorithms in their code, because the type of fairness – and its drawbacks – can have massive, even society-level impacts. From everything like determining what the minimum wage should be to who gets hired for a job to even who a company markets to, fairness algorithms can either reduce biases or magnify them.

How should we be thinking about these kinds of algorithms? We have to approach them from a balance of what our ethics and values are, balanced with our business objectives. Our ethics and values will dictate which fairness approach we take.

Many different simulation tools exist that can evaluate a dataset and provide projections about likely outcomes based on a variety of fairness metrics, like IBM’s AI Fairness 360 Toolkit and Google’s What If Toolkit. But the onus to think about and incorporate fairness techniques is on us, the humans, at every stage of decision-making.

What Constitutes Trusted AI?

What is trusted AI? It’s AI software that achieves four key traits:

  1. Fair
  2. Accountable
  3. Values Driven
  4. Explainable

Let’s dig into each of these.

AI should be fair. Since AI systems make so many decisions on our behalf, we need to know that the decisions it’s making are fundamentally fair. Fairness, as we discussed in previous issues of the newsletter, can be tricky to navigate in terms of outcomes, but the bare minimum standard of fairness is that AI does not discriminate on protected classes (age, gender, race, religion, disabiity, etc.) or inferred variables that correlate to protected classes. Every decision AI makes should at a minimum be blind to those considerations, except where permitted by law and ethics.

AI should be accountable. When we build systems to make decisions, whether it’s who to show our ads to, what constitutes a valuable customer, our systems must inform the users – us, and our customers – how it made those decisions so that we can hold the system accountable. If an AI systemm declines your loan, it should explain what factors led to that decline. It’s not enough for the system to say a loan application was declined, but also spit out things like insufficient household income, or credit score below the required threshold – whatever variables used to make its decision should be communicated to the user.

AI should be values-driven. This is a BIG one. Our AI systems – and their outcomes – have to match our values. If we claim we support, for example, non-discrimination based on age, and our AI models discriminate based on age, we have a system that’s out of alignment with our values.

As an interesting side note, we often say that Facebook has built a system that fundamentally makes the world a worse place by amplifying negative emotions and promoting rampant misinformation. Interestingly, this doesn’t conflict with their core values: Be bold. Focus on impact. Move fast. Be open. Build social value. Nowhere in their statement of values do things like “engender happiness” or “make the world a better place” exist, so it should be no surprise to us that they build AI which is aligned with their values – even if it doesn’t align with our values.

AI should be explainable. Ultimately, any AI model – which is nothing more than a piece of software – should be interpretable and explainable. How did a system make its decisions? What data did it learn from? What algorithms did it incorporate? When we know what’s in the engine, it’s much easier to fix it when it goes wrong. When we know what the ingredients are in our cooking, it’s much easier to correct our dishes.

All this sounds great as abstract theory. This is what we want in systems that make decisions on our behalf, every day. The question is, how do we practically implement some of this?

Building Trusted AI with IBM Cloud Pak for Data

The system we discussed using to make trusted AI happen is IBM’s Cloud Pak for Data, which includes the service I use extensively, IBM Watson Studio. Why does a system matter? Aren’t there tons of best-of-breed tools out there?

There are, but the integration is what matters when it comes to trusted AI, because you need common rules, common governance, common access controls, and common monitoring to ensure that your AI is doing what it’s supposed to be doing at every step of the process. Only integrated systems deliver that on the enterprise level, and IBM’s Cloud Pak for Data is one of the best solutions on the market.

For production AI, nothing beats Watson Studio’s ability to monitor your machine learning models and alert you about models drifting away from the rules you’ve set – before you incur liability for them doing things they’re not supposed to do.

Learn More at IBM THINK 2021

Want to learn more about trusted AI? Join me and tens of thousands of folks at IBM THINK 2021 on May 11, 2021. Virtual, from your desktop, and totally free of cost, IBM THINK 2021 will let you talk to the folks making the goods and ask questions about how you’d approach your toughest AI software problems.

Disclosures

I am an IBM Champion, and my company, Trust Insights, is a registered IBM Business Partner. As such, anything you purchase from IBM through Trust Insights indirectly benefits me financially.

Machine-Generated Transcript

This transcript is generated by speech recognition; it contains flaws and is not a substitute for watching the video.

Christopher Penn 0:12

All right, everyone.

Welcome This is implementing responsible trusted AI systems a fireside chat.

I’m Chris here with Lauren Frazier from IBM.

Today we want to talk about exactly what says in the box building trusted artificial intelligence.

Before we begin just a bit of housekeeping, wherever it is you’re tuning in, please go ahead.

And you know, leave a comment, let us know we’re in from if you are on, if you’re watching us on Facebook, you will need to go to streaming our.com slash Facebook, if you want us to know your names, if you just want to be, you know, anonymous cheater, or whatever in the comments, that is fine as well.

But if you’re on Facebook, go ahead and leave your comments there after authenticating.

So, Lauren, why don’t you introduce yourself real quick? Yeah,

Lauren Frazier 0:54

thanks for hosting us today, Chris.

I’m Lauren Frazier.

I’m an IBM Content Marketing Manager with IBM Cloud Pak for data.

So that’s our leading data and AI platform runs on any cloud.

And hey, we’re focused really on trustworthy AI right now.

So the timing couldn’t be any better.

So we can go ahead and kick it off.

And you know, discuss the responsible AI, especially, you know, now the stakes are higher, right, AI can be used for good or if you use it wrong, it’ll have negative consequences, whether that means in money, financials, or just trust with your customers.

So businesses that handle data, they can no longer just ignore the societal responsibilities, we really need to put that at the forefront of operationalizing.

AI, how do we make a trustworthy? So Chris, my first question for you is, why is it important? And what implications are there in deploying AI, and while especially ensuring that responsible AI is infused within, you know,

Christopher Penn 1:49

it comes down to, if we want to trust something, we need to know that it’s going to operate, you know, with fairness and stuff, this there’s a lot that goes into trust.

But fundamentally, we’re trying to roll out this technology as a society as a civilization to as many all these different applications, right mortgage and loan applications, criminal recidivism, more mundane stuff, like marketing effectiveness, which is sort of the area that I study.

And we need to know that the machines are doing what we want them to do, and not exposing us to unnecessary risk.

You know, there are no shortage of examples where AI hasn’t been used responsibly, right, it hasn’t been built to be trustworthy.

And I think that we should probably, like, define what trustworthy means.

If you go to research.ibm.com, there’s actually a really good whole section on trusted AI.

But there’s four fundamental things that make AI trustworthy, fair, is it accountable? Is it values driven? And then is it explainable? Real quick, Lauren, when you think about fairness, what does what does that word mean to you, for fairness,

Lauren Frazier 3:02

for me means equality, it means, you know, people are being treated all the same, no matter what that data is used fairly.

So that means data is used properly is used for the good of people, the good of the world, the good of making decisions and better business decisions, which ultimately brings in the money but also changes and impacts the world.

And it doesn’t matter who and what that person does, but fairness is giving everybody that equal slate.

Christopher Penn 3:31

Yeah, it’s it’s challenging, because there’s different definitions of fairness, right.

You know, some real simple examples.

There’s, there’s what’s called blinded fairness, we say anything that is protected, your age, your race, your gender, that data is removed, it can’t be used for decision making.

It’s collect the bare bones.

But one of the things that AI is really good at is, is doing what’s called correlates, where you say, Okay, I may not know your age, but if you like, you know, Goonies.

And you like, you know, I’m trying to go way back, like MC Hammer in the early days and stuff, we can infer your age, right, because you miss things that you’d like, I’ll have a certain time ministers.

So that’s one aspect.

A second would be what’s called representative parity, where if I’m trying to sample some data, I try to make the data represent the population.

I used to work at a company in Atlanta, and on staff at 100.

person company, there wasn’t a single black person.

Yeah.

Atlanta’s 54% black

Lauren Frazier 4:42

and pretty good community.

Yeah,

Christopher Penn 4:45

exactly.

So there was that’s a case where there is not representative parity.

And then there’s a two where we have real significant philosophical debates.

equality of opportunity and equality of outcome.

equality of opportunity means we Get the same chance.

Yeah, at that success, but success is left up to our individual merits.

And then equality of outcome is no matter who we are, we all get the same thing.

And there are definitely cases where, like COVID vaccines, we want equality of outcome, everybody gets it.

Right.

Lauren Frazier 5:17

Everybody gets it.

But you know how hard it was to get it AI could have been used more to drive who needs to get that first instead of us, for instance, me fighting over event bright in my mom, also in a whole other state trying to get my Nana has 90 a vaccine, AI could have helped us just improve that.

And hopefully, we don’t have to see that going forward.

But we will be ready.

If something, you know, a health crisis does come up again.

Christopher Penn 5:42

Exactly.

So fairness is part one of trusted AI to was accountability, where the machine tells us how it made us decisions.

So I go to apply for a loan, and it says, Hey, Chris, your loan was denied for you know, because your credit score was below 670, or your household income was insufficient.

But it should also tell us what wasn’t involved.

This isn’t like, Hey, Chris, the fact that you’re a guy wasn’t a factor in the decision, the fact that you’re old.

And we need our machines to tell us like this is how I made the decisions and a lot, again, a lot of machines, they’re very, very opaque they, they won’t tell us what’s going on.

Number three is our AI should be values driven.

And this is where I’m just gonna get myself into a whole bunch of trouble here are companies and the values that we have, as human beings informed the data that we put into these machines, AI is nothing more than a bunch of math, right? It’s not magic, it’s math.

And it’s math, it’s trained on data.

So the data we put in means that that’s what the machine learns to write its own code from, we have to have values that are aligned with the outcomes we want.

There’s a power packet, you know, if you look at some of the things that like Facebook does, and they have been rightly criticized in the public press, for making some questionable decisions.

And if you look at their core values, be bold, focus on impact, move fast, be open, build social value, at no point in there, does it say make the world a better place, make people healthier, promote truth, you know, these are other values that other companies might have.

And so they’re AI reflects their values.

So as part of trusted AI, you almost have to trust the company making the AI Yeah,

Lauren Frazier 7:34

and especially as a customer, we don’t, as a consumer of Facebook, or anything, or even just applying for an HR loan, or you know, even behind the scenes in HR applying for a mortgage loan, you don’t understand the big technology around it.

So, you know, companies have to make sure that there’s a way that they can explain it, because I think you know, you don’t want to be in the court of law, you don’t want to be the front page on the news.

And then that’s when your customer realizes, oh, wait, my data hasn’t been being used properly.

So I mean, I guess with the rise of all of these events last year to including how we work in the pandemic, and some of these societal and political events, I think we all know of, I think data reflects a drastic changes in human behaviors itself.

So as we kind of talked about already, the pandemic, but what else do you see that is different from last from this year from last? And why does this matter today and in the scope of AI?

Christopher Penn 8:29

Um, I want to go back real quick.

There’s one more piece of of trust AI that I think matters to me answers this question really well is that is the difference between explainability and interpretability.

AI has to be both of these things.

The analogy that I love to use is, imagine you’re a famous restaurant, right? And you’re trying to figure out like, what would they use to make that chocolate cake? Just by tasting it, you can tell like, Okay, this is good, then you go home and try it yourself.

like okay, I can’t I it’s not quite the same.

That’s explainable.

AI, right? Well, you see the outcome of the model and you go, Okay, this looks okay.

It seems okay.

interpretable AI is when you are in the kitchen with the chef watching them like, oh, you’re not using almond extract using cherry extract.

That’s, that’s it this Oh, you you let that rise for a little bit longer than normal.

Why do you do that? Like it helped develop the flavor of something.

When we’re talking about interpretable versus explainable AI? A lot of companies are really pushing for explainable because it’s cost effective, to just say, Oh, yeah, the model did the right thing.

See, the results are like, Yeah, but I don’t know that.

If I am concerned about things like bias in my AI.

That’s not good enough.

One of the things I love about tools like IBM Watson Studio, is that it will in the AutoAI module, it’ll build a model, but then you push a button says turn this back into code, and then it turns it back into code.

Now I can step through line by line and say, what decisions did you make How did you build this code? And I see some like, Ah, you did you something.

I told you Not to do I can take that out.

Because you’re right.

in a court of law, I’m going to need to produce the code.

Yeah.

I honestly think that’s okay.

For humans, we have a presumption in law that you’re innocent until proven guilty.

I almost feel like with machines, it should be the reverse, like the machine should be suspected of bias until we can prove that it isn’t.

And we prove it isn’t by producing the code.

The challenge is, and the reason why so many tech companies don’t want to go that route, is it’s expensive.

It’s cost cost inefficient.

And it’s controversial.

But going to your question about what’s different this year than last year.

The pandemic has been essentially a World War, right? It’s a World War of the entirety of the human race against a very, very, very tiny enemy that can replicate like crazy.

And it’s a crisis.

And the funny thing about crisis is that it brings out it amplifies things, the good gets better, the bad gets worse.

And all the things the inequalities, the inequities in our healthcare system versus, you know, income gaps, pay gaps, get worse, in in a crisis, just like you see.

What was the stat GoFundMe is the as America’s third largest health insurance plan? Right? people asking for help begging for help is the third largest health insurance plan.

And we know, to your point, AI could help solve a lot of these things.

It was deployed responsibly, and in a trustworthy way.

But the challenge is, things like training data that goes into it, we have to help people build better systems that say, that look for for bias at all times in the system.

So we have to say, you know, is the data going unbiased? Is the model biased? And does the model drift? You know, again, one of the things I like in Watson Studio is that it tells you in the model monitoring, hey, this thing’s going off the rails, you want to do something?

Lauren Frazier 11:59

No, that’s true.

And I think it’s important that we just also with Watson Studio, you are able to model or monitor that model, but also interpret and explain.

And that’s the key things you were saying it’s not just about explaining, but proving it to anybody, and also making it easy for every maybe that court of law or the external folks to understand, Okay, I see how my data was used if I ever did need to ask that.

So you know, I know that for us, IBM has always strived with him and innovation and bringing benefits to everyone and not just a few.

And I think even in hiring, you know, I might, my own team is pretty diverse.

So I have enjoyed being an IBM for the past seven years.

But this fopp philosophy is also applied to AI and we aim to create an offer reliable, and understanding technology.

We understand that AI is embedded in everyday life, right, which we’re talking about whether it’s business, government, medicine, health care, all of that.

But our goal is to help people and organizations adopt it responsibly.

So I know we kind of defined trustworthy AI, would you just find responsible AI is the same thing as that.

And what are the opportunities and challenges that might come with the use of ethical AI?

Christopher Penn 13:09

Trust is what you build up front? responsibles What happens after, right? So you, it’s kind of like, it’s like any relationship, right? You you build a relationship, you build trust, upfront, and then on the back end, you prove that that trust was well founded or not, right? Depending on the technology, when you look at, at the ethical use of AI, it’s funny, ethics is a tricky word.

Because ethics, you know, sort of in the classical Greek sense, means you do what you say, right? If you look at, you know, the way Facebook and once it’s AI, they do what they say.

And at no point did they say they’re gonna make the world a better place.

A whole world on fire.

But it’s ethical.

Yeah, they’re doing what they said.

The question that we have to ask ourselves as, as the people who maintain AI, is are we doing the things that we want to accomplish? are we creating the outcomes that we think are fair, and that are equitable? And for a more practical way of things looking at? Are we doing things that are gonna get sued, right? It’s like, Oh, yeah, sorry about that.

We accidentally when you have those, those data sources, you know, inside the machines there’s so many ways can go wrong.

I was at a I was at a conference a couple years ago, the mahr tech conference, which and of course every vendor on the floor had you know, we have AI and our product like yeah, it’s not it’s not the tele guys you don’t need doesn’t need to go on everything.

This one vendor had his map of Boston.

Yeah.

And they were trying to predict ideal customers and it’s for Dunkin Donuts right now.

For those of you who are actually We’ll see in the comments you have people listening from all around the world, Dunkin Donuts sort of a mass market.

Coffee and donut shop, right? And pretty much everybody in New England, the New England region of the United States consumes Dunkin Donuts in some form, the only people who don’t are dead.

And this company tried to predict these ideal customers has a map of Boston, there were red dots in the areas that were, you know, ideal.

And then there were no black dots neighbors weren’t.

I looked at this map.

And I said, so you think the ideal customer is all in the financial district, downtown Cambridge, and Roxbury, Dorchester mattapan, which are predominantly lower income, predominantly black areas, say there’s no ideal customers.

I’m like, I’m sorry, you’re full of shit.

Because there everybody in Boston, regardless of race, gender back on, get you consumed dunks in some fashion.

And I said, what you really did is you invented redlining? Yeah.

Which is, again, yeah.

So again, for those folks who are not familiar with American history, in the 1930s, insurance companies would take maps of the world of the cities and draw red lines around, predominantly, you know, minority areas of cities say we don’t want to give loans in these areas.

And it that’s not an equitable outcome, particularly for something like coffee like, Okay, if you’re selling airplanes, and yes, there’s an argument to be made that some sections the city by income level, you might be justified that but you’re selling coffee, you’re selling $1 coffee, everybody can get that.

And so with that, you know, ethical, responsible use of AI, we have to think about what kind of risk are we opening ourselves up to if we implement it badly?

Lauren Frazier 16:38

And I think it’s important to also say, I think it’s something you mentioned before, it’s who’s in the boardroom, who’s behind there making these decisions.

So I think someone in the chat brought up a good question is, where do you get training data when you know, the data itself does not represent the overall pool accurately? You know, if folks aren’t behind the scenes and can say, you know, wait, this is redlining? Again, this is because clearly someone didn’t look at that it’s quite obvious that we’re just doing something that we’re trying to still kind of going on sometimes, and we’re trying to pivot and, and change the world, right? So how do people get that correct data? How do we cleanse it? How do we even get there?

Christopher Penn 17:18

data data is too far gone.

At that point, you actually raised a really good point.

This, you can get bias in AI and all kinds, including allowable bias to creep in at six different spots in the process.

But the number one place which starts is in the people you hire, right? Yeah.

If the people you hire, and I’m not saying that you’re hiring bias people, but if you’re if you hire people who don’t think to ask the question, hey, is there a bias in this data, then you will never get to the point where the systems can can detect it.

Now, if you get somebody to say, Hey, I think this, this data might have a problem.

I don’t know what it is.

But there might be a problem in here.

And as built into your strategy, which is the second place, it can creep in, then there are tools that you can use to assess your data.

IBM has a great toolkit called the AI fairness 360 toolkit, it’s free, it’s open source, you can use it in R and Python, I use the R version, and you feed it data and it says, Hey, what are the protected classes? Right? What are the things that you that cannot be discriminatory? What kind of fairness are you looking for? We talked about the different kinds of fairness.

And then what do you want to do about it? We’ll say like, yes, there’s a skew of plus or minus this percentage, or there are issues here.

And then it’s up to us to say, how do we want to deal with that? In Watson Studio, you can handle this with a couple different points in the in the data, the model building site up front, you can actually with the toolkits help flip bits.

So if I have a, say a data data set, that 60% male and 40% female Watson Studio with, with our guidance can say, and we you have to tell it to do this, I want you to flip the bit, randomly sample the data set and flip the bit on 10% of the males to turn them female so that it balances the days out.

The model monitoring does the same thing as well, I will say okay, I can flip bits around or change data around to try and remix the sample, to keep it fair to keep it on the rails.

The other option is, you, you you filter the data upfront and say, Okay, I’m going to do say propensity score matching, and I’m only going to allow an even gender split are only going to allow a representative population split in the data.

So that what goes into the training for the model construction is fair to begin with.

That’s a really good question.

It’s a It’s a challenging question, because you have to be aware of how to do these things.

Lauren Frazier 19:42

Yeah.

And aware of what bias is

Christopher Penn 19:45

exactly how to

Lauren Frazier 19:47

spot it.

Right.

So I guess that goes into kind of the automation of AI you know, more company needs are used to using AI operationalizing, but only by embedding ethical principles into these Applications and processes, can they probably be built on trust? Right? So what do you see as a creek key criteria for bringing models to production and driving value from the deployments? And what do you see in like trends in the architecture that folks are adopting or should adopt? You know,

Christopher Penn 20:16

there’s there’s a few things here that I think are important.

One is automated machine learning has really come a long way.

Lucas was asking in the comments of the link to the IBM Phil, if you go to AI, f 360, dot my bloomix dotnet, I put a link in the comments.

That’s the AI fairness 360 toolkit.

So there’s, there’s a few different components that you need to have in the system.

And here’s the challenge that, again, a system like cloud pack for data will address that, you know, sort of mixed bag of individual solutions will not necessarily do because they’re not connected to each other.

So you really want the integration, you need to be able to get at the data, where it where it lives, right.

So being able to use something like Red Hat openshift to virtualize the data out of where it is, and make it into a common layer.

You need a system like AI fairness 360, to look at the data and say, Okay, is there bias going into it is, what kinds of issues are there, you need, I like tools like Watson Studio AutoAI AI, because in some ways, it takes some of the decision making and the potential biases I have as a data scientist out because it’ll feed it a dataset and say, here’s the 44 things I tried, here’s the best result, here’s the seven different measures of accuracy, which you know, I think this is the best one, but then I can always go back and I can push the button, say, okay, generate the code, I always go back and say, I really actually want to use gradient boosting for this.

So you need to have that sort of that in the model construction phase, then you have deployment, you got to get that model into production.

And then you have to monitor the model as well.

And this needs to be an ecosystem that where the pieces talk to each other, as opposed to being you know, individual point solutions, because what tends to happen with point solutions is they break really easily.

I can pass a model from say, you know, our studio into a standalone platform, but that standalone platform can’t monitor drift, and then can’t pass back into my original code and say, this is this is a problem, I have to do that manually.

And if I’m, you know, working on five or six projects for different clients, whatever.

I may not remember to do that.

If I’ve got a system like cloud pack for data and Watson Studio.

It does it for me, right? So I in a lot of ways, it takes my vices out of the equation.

And it also automates a lot of the maintenance the the operation of AI and that that part is it’s something people don’t think about, when when people think about AI, they think like this magical unicorn that you know, you strap your date who flies off into.

Unknown Speaker 22:52
Here it goes No,

Christopher Penn 22:56

exactly.

And it’s not it’s the Oh, it’s almost like AI really is nothing more than really fancy spreadsheets, right? You don’t expect Microsoft Excel to run itself, you have to do stuff with it.

And in the same way, AI is just software, except it’s software that a machine wrote from your data.

So you want that ecosystem so that it’s running your your data, your models, your and monitoring, all in one place.

And that way, it can tell you proactively, I think something’s wrong here.

Lauren Frazier 23:30

And your whole team’s gets the visibility of it as well, not just you, you can see where the issue happened, how can we go back with wherever can we you know, mitigate that risk or mitigate that bias? And, you know, I know you already brought up HR and I know one of IBM biggest clients is using AI to ensure hiring and other HR practices are fair, and especially with corporate policies and the social responsibilities of today.

But what kind of client questions are you getting when it comes to operationalizing AI, or the use of AI?

Christopher Penn 24:02

You know, it’s funny, um, our clients in a lot of ways don’t care about AI.

What they care about is better, faster, cheaper results.

We want things to work better.

We want more accurate models we want.

Not even the the models, you know, one of our larger clients and automotive client, they just want to know what to do help me make better decisions faster.

But going into that, there’s a lot of challenges.

The biggest challenge that a lot of people face is, you know, it mirrors the AI lifecycle.

Do you have the right people? Do you have the right strategy? Do you have the right data? Do you have the right algorithm choices? Do you have like models? And do you have the right monitoring it to keep it all intact? That hierarchy that bet process, in a lot of cases is really broken.

A lot of companies they don’t have the right people, which is why they need, you know, firms like Trust Insights and companies like IBM, they have a strategy, but the strategy may not be optimized for AI, because AI is all about getting machines to do things that humans, you know, tasks that humans do.

And if you’re not thinking about being process oriented, and you think about how do I be efficient, then AI is not really going to work for you.

And then the big one, by far is the data is a hot mess.

It’s everywhere,

Unknown Speaker 25:33
here.

Yeah,

Christopher Penn 25:34

exactly.

Yep, it’s everywhere.

It’s in the wrong format.

It’s not structured, well, it’s corrupted.

Even something as simple like, well, things we see a lot when we’re doing marketing analytics is, hey, hey, you launched a new website.

And you forgot to put your Google Analytics tracking code for three weeks.

So you go back in the data, this is a big three week gap here.

Like what what happened guys like, Oh, they they something as simple as that can really hose even basic analysis.

And then there’s all this stuff around the humans.

So how do we communicate? What AI is doing to, you know, to the people? Who are the stakeholders? How do we help them understand that you will get better outcomes? How do we show them some early easy wins? So one of the things that we do a lot of is attribution analysis, take all the data that you have, and say, these are the things that work in your marketing.

That’s a pretty easy win, because it helps people understand, oh, I’m investing 44% of my budget in Facebook ads, but it’s delivering 2% of my leads, I get it.

Lauren Frazier 26:43

Where else can we put this? Yeah,

Christopher Penn 26:45

exactly.

So those are a couple of the examples where we use it extensively, we’re actually working on a couple other projects, where we’re trying to look at buildings, you know, ongoing running models that help do some predictions and forecasting.

We just did one recently doing predictive analytics, just helping a client understand, hey, here’s what’s likely to happen in the next three months for this particular type of content.

You should time your promotions, to when interest by the eyes is going to be the highest, because that’s what people are paying attention.

People are like, like spotlights, right? And you know, the spotlight moves around.

It’s watching, you know, the Falcon and Winter Soldier.

It’s watching.

Unknown Speaker 27:34
It’s such a good show.

Christopher Penn 27:36

But you have you earn very small amounts of time as people’s attention.

But when they’re thinking about something that they care about, that is something you solve the spotlights on you.

So what can you do take advantage of that time, if you’re not prepared? the spotlight hits you then moves on, right, like, well, they go.

But if you use predictive analytics, if you use AI intelligently, when the spotlight hits us, you’re like, Hey, here’s the thing you can buy, and then they buy it, and then the spotlight moves on.

Yeah.

And that’s the, the real operationalizing of AI is not just getting the system running, but getting the benefits from it.

Lauren Frazier 28:13

And it’s not just the benefits, it’s being or its benefits, but for predicting those outcomes in intelligently through automated processes, I think are key.

And it’s also I think it goes back to what we were saying at the beginning.

It’s not just about the business impact, it’s about the impact to the world you’re making into your customers, and how you’re improving lives by these decisions, whether that’s loans, whether that’s, you know, even data for up for universities and students, there’s so many ways that data can be used.

So before,

Christopher Penn 28:45

we have to be very careful about to when we should not be using AI.

I think there are cases where it is clear that AI is the wrong choice.

Yeah, I’ll give you two examples.

Okay, example one.

A pro publica did an investigation in 2016 of police departments creating an algorithm attempting to predict recidivism.

recidivism, for those who don’t know, is the likelihood that someone will re offend or commit another crime.

The algorithm that this company came up with, predicted that black Americans would reoffend five times more than they actually did.

But even worse, the algorithm itself was 20% accurate, right? You can flip a coin and do better than this album by a substantial margin, not just a little bit.

That was a case where it’s still not clear whether somebody had their thumb on the scale, it was intentionally biased.

Or if they just fed it such bad data that it came up with it.

So that’s a case where the data was bad, and the people putting it together probably didn’t know what they were doing or they did and that was malicious.

Second example.

This is a big one.

AI is not a good choice in a lot of things like health care.

in finance for specific populations, not because AI the technology is bad, right? But because the whole data set is corrupted example, black American healthcare outcomes, you can there is zero good data about zero in the entire country.

Why? Because systemic racism has created such a mass of negative outcomes that it does not matter where you sample your data from, it’s going to be bad.

What you have to do is kind of like what you’re talking about with Watson Studio, where you have to almost change other people’s races in a system to introduce known good data to say like, yeah, the ideal health expectancy outcome should be like 70 years old, because of problems that are outside the data, macro systemic problems, you can’t trust that data.

So one of the things that you have to ask yourself when you’re deploying AI is, is the data itself.

So corrupted, that it cannot be usable, that you can’t recover it? And there’s no good data to be found? If that’s the case in AI is not the right choice, you will have to rely on boring old natural intelligence, do you have better data,

Lauren Frazier 31:16

and I’d rather us rely on that.

But I do have hope for code for the future that, you know, hopefully, these companies, I know that IBM is striving for it.

But hopefully we continue to see just from the past year, and all of these items being brought to the forefront, right? I think there’s been a lot more visibility on how just much systemic racism has affected all of us, and outcomes.

And I just hope that all organizations start to think how can we really start to go behind the scenes look at our data from the beginning, is this half what we should even be using? And hopefully in the future, it could be used for good in those areas as well.

Always improvement, right? In all technologies, especially with AI cuz, you know, the folks always think it’s going to take over their jobs, too.

But hopefully, it can just be used for good.

And that’s the key thing is in what we’re trying to drive here as well.

Christopher Penn 32:07

Yeah, I think the whole thing like is, am I going to take my job is it’s a nuanced conversation, because a job is a series of tasks, right? Yeah, yeah, you know, you don’t just do one thing anymore.

At the very least, you at least have to attend meetings about the one thing that you do.

So AI is really good at tasks, it’s still good at being very narrow, at least until IBM perfects quantum computing.

And they may become a little ways away.

But right now, it’s it is very much good at taking tasks.

And the more repetitive a task is, the easier it is to automate.

The good news is that a lot of those tasks that are so easy to automate, you don’t really want to be doing anyway, I used to work at a PR firm.

And this one task a junior person had they were copying and pasting results from Google to a spreadsheet eight hours a day, like, how have you not clawed your eyeballs out by?

Lauren Frazier 33:08

Now, that’s my role there.

So

Christopher Penn 33:12

it’s like, why are you doing this this Sunday, the machine should be doing it, it’s so repetitive.

And the human there adds so little value, that the goal then is say, okay, you, we’re gonna save you seven hours and 45 minutes your day.

But now, we want you to use what you’re good at.

Thinking across domains and stuff to add value to this instead of just copying and pasting spreadsheets.

So this is concept Oh, gosh, from the 1930s.

Joseph Schumpeter the idea of creative destruction that yes, things like AI absolutely will destroy certain tasks.

But doing so they will create new opportunities that will ideally be better nobody enjoy getting a saw going until they can sign up ice in the wintertime, right? Nobody enjoys that.

Lauren Frazier 34:01

Exactly.

And no one predicted AI back and back then, right.

So that leaves as I might take over some things, at least folks for innovation and other things that we might not even know could be possible in the future.

So with that, I mean, Chris, it’s been a great conversation.

I mean, I’m thank you for hosting us.

Thank you.

Are there any last words you want to say before I share some of the resources in the description?

Christopher Penn 34:27

I would encourage people to, to at least start if you haven’t already started thinking about the applications of automation for even just, you know, not AI, but just automation.

We are entering a period of time now where productivity is sort of the golden calf that everybody is looking for in business.

Even if your company doesn’t have a enterprise wide strategy for AI, you as a business person should be thinking about how do we implement even on a smallscale piloting it, you know, you can sign up for an IBM Cloud account and try out Watson Studio, I believe you get 50 CPU hours per month, which is enough to test some stuff out.

It’s not enough to, you know, to run an enterprise wide thing, but you can start testing it out.

There are so much good open source, in our in Python learning those languages.

If you go to a cognitive class.ai, that’s IBM is free, online university to learn Big Data, Data Science, machine learning AI, it’s an incredible resource completely 100% for you, you can get cool little badges, you can go on your LinkedIn profile.

I think getting your your toes wet in this stuff is so important, just so you know, what’s possible.

And the more if people try it out.

I think that the better is for everyone, because it helps to demystify it.

It’s not magic, it’s just a bunch of math.

So Lauren, what are some of the resources you want to talk about? Yeah,

Lauren Frazier 35:58

so you know, think 2021 is coming up.

Of course, it’s not our great event that used to be in person that I love, but it is virtual.

So it’s coming up may 11, in the Americas, in May 12, in APAC, Japan, and EMI, you can also sign up for part two of our what’s next and AI webinar series.

That’s something we’ve been working through IBM, this is on May 19.

And that will dive in a little bit more with some of our experts on the IBM side some of the product marketers on just the importance, what’s next, how you can improve your own AI strategy.

And then finally, and not last, the gardener report.

This one goes into our how Watson Studio is one of the leaders in machine learning and data science capabilities.

So just so you can see it, you know, what the analysts have to say, but, you know, it’s been a pleasure.

And I think just, I think what you said we need to just get, companies just need to use AI responsibly.

And I think consumers need to also ensure that these companies do get held to those standards as well.

So it’ll be cool to see how it goes going forward, especially improvements in AI as well.

Christopher Penn 37:09

Exactly.

And if you have any interest in it, we have a course as well is not free, go to TrustInsights.ai slash data science one one if you’re so interested.

Thanks for watching.

If you’d like to share this episode, go to LinkedIn that you probably founded by TrustInsights.ai dot AI slash IBM trusted AI.

That link will take you back to the recording of this show.

You can share it with your friends, please do we’d love to have more people get a sense of what’s important about AI.

And making sure that is fair making sure that we’re all working towards outcomes that are equitable for everybody.

Thank you, Lauren and the IBM team for being with us today and we’ll talk to you soon.

Take care everyone.

Lauren Frazier 37:47

Thanks, Happy Friday.


You might also enjoy:


Want to read more like this from Christopher Penn? Get updates here:

subscribe to my newsletter here


AI for Marketers Book
Get your copy of AI For Marketers

Analytics for Marketers Discussion Group
Join my Analytics for Marketers Slack Group!


Pin It on Pinterest

Shares
Share This