Sep 25, 2019

Ray Fleming and Dan Bowen, who both work in the Microsoft Australia Education team, talk about Artificial Intelligence in Education - what it is, how it works, and the different ways it is being used. It's not too serious, or too technical, and is intended to be a good conversation of background information. Please note that we're planning to have a chat each week about the story of AI in education, and are definitely not going to be describing the offical Microsoft position on anything!

We'll start by discussing what Artificial Intelligence means to us, and then through future weeks we'll talk about specific technologies, scenarios, and bring along some guests to talk about their work.

TRANSCRIPT FOR The AI in Education Podcast
Series: 1
Episode: 1 

This transcript and summary are auto-generated. If you spot any important errors, do feel free to email the podcast hosts for corrections.

This podcast excerpt features a conversation between Microsoft colleagues Dan Bowen and Ray Fleming, who discuss the nature, applications, and ethical considerations of Artificial Intelligence (AI), particularly within the education sector. Fleming, the Higher Education Lead for Microsoft in Australia, defines AI through two main lenses: machine learning, which uses data patterns to make predictions, and conversational interfaces like chatbots, referencing the Turing test ideal. The speakers agree that AI is becoming commonplace, citing everyday examples such as personalised online advertisements and predictive navigation systems, but they also emphasise the ongoing "arms race" in technology that pushes AI to outperform humans in areas like speech recognition and image analysis. Crucially, the discussion pivots to the ethics of AI, highlighting the risks of amplifying human bias in recruitment systems and the danger of using AI for life-altering decisions, such as predicting reoffenders, without a "human in the loop." Ultimately, they conclude that education is a data-rich environment ripe for AI-driven personalisation, noting that the easier accessibility of AI tools presents a timely opportunity to leverage this data more effectively.

Okay. So, welcome to our podcast. I'm Dan Bowen and
I'm Ray Fleming.
We work for Microsoft and today's podcast is going to be about artificial intelligence. Ray?
Yeah. AI
it is. So, tell me about yourself, Ray. And also tell me Who is your favorite AI robot?
Oh gosh. So the the first bit's easy. So I'm Ray Fleming. I'm the higher education lead for Microsoft in Australia. But probably what's much more important is my background. So I've had 30 something years of uh education technology. I've worked for a bunch of companies um in the UK and here in Australia. I work for Microsoft in the education team, but my background is always education technology and never education. So I'm not an ex educationalist.
And what was your favorite AI robot? Oh, it's probably some it's probably it's probably Marvin. Marvin the paranoid the paranoid android from the Hitchhiker's Galaxy.
Um, so ah yeah, it was just like seeing a robot that had personality.
Yeah.
You know, cuz you always see robots as being robotic, you know, that that and so seeing, you know, Marvin, I remember when it was I read the book, it was interesting, but then when I saw it featured on TV, It was like, "Oh my word. Robots aren't just aren't just things devoid of emotion that they could actually be almost almost human."
Yeah. It's really interesting, isn't it?
And you, Dan, your background.
I'm Dan. I use as an ex-teer and a school inspector in the UK.
Boohist. And um yeah. Uh and and I work in the education field in Australia for Microsoft at the minute looking at schools specifically around that quick K12 area. So,
and I know you Dan, you will have her favorite. What?
Absolutely. It's got to be HAL, hasn't it? From 2001. I think, you know, it it's always interested me and interestes probably everybody who's been looking in the in the sphere over the last couple of years around and and beyond, you know, when you think back to HAL, you know, the way that actually the worm turns, the way that actually technology
bites back and I think in today's podcast, we can kind of discuss that more, but HAL was that interesting AI for me that really kind of started to say, well, what happens when that AI becomes all-encompassing and actually able to control things? What can it be? What's that kind of end state? And I suppose that really fascinated me in a massive way. So, when we're looking at AI then uh Rey, from your point of view and and from your kind of lens across the education sphere, what what is it? What is AI generally?
Well, let let's take it out the education sphere first of all because you know, we we we talk about AI and we talk about artificial intelligence, but we don't define it really well. And there so many different aspects to it that we're probably, you know, a six-hour podcast and we wouldn't got to the end of of defining what it means. So maybe better if I start with a what do I see it as? What what are the examples I talk about? And I I really think there's a couple of things. One is thinking about all the data and the way that we can use data to predict the future. So that's often referred to as machine learning where you're really using the patterns of the past to predict the future. And we've got so much data that you know it's easy to do it. So that's one kind of way that AI is used. And the other which comes up pretty regularly in conversations is things like chat bots, the interfaces, you know, that's often where AI started from. It was thinking about could you create a conversational interface that was indistinguishable from a human.
Yeah. With the Turing test.
Yeah. Exactly. So could could you build something that somebody on the other end didn't know wasn't a human? But you know, there's a lot of conversation around that because we're all so busy in our lives that we need answers to questions all of the time and there's many organizations where they just don't have the capacity to be able to answer all of our queries and so people are thinking about bots as a way of handling that so you know when I think about artificial intelligence it's those two areas about making predictions of the future with data and it's about conversational interfaces but there are so many other examples as well but let's just stick with those two as we talk for a bit
so so it's interesting you're kind of bringing that you know at a high level there because it it does seem to be in the news more and more. Why why are we talking about it now? What's really driving the current conversation?
I I think it's because it's becoming so common place in our world and it's becoming so commonplace because actually the technology is so much easier now. But the the commonplace thing is, you know, we we're no longer surprised when you go to a website to place an order for an item and it tells you that you might also be interested in this other item and you think, "Yeah, actually I am." Or you go on to Facebook and the ad you see is an ad for something that is relevant to you rather than just being generic. So, it's becoming very common place in our lives. You know, when whenever I hop in the car, you know, the mapping app tries to guess where I'm going to. And so, you know, it'll every morning I hop in the car, it works out I'm going to the office. So, it's that kind of everydayness that means that it's becoming a pretty regular topic to talk about. I re I realized couple weeks ago, I was dealing with a telephone company that had made a mistake on one of my orders. And I was chatting to what I assumed was a bot. And afterwards, I was thinking, I'm not actually sure if I've talked to a human or not.
But I guess I only thought about that because I know of the technology in the background. I guess many people that we just make an assumption it doesn't really matter to them because it's becoming so automatic in our daily lives and and especially I think with younger people.
And this is a blue as well when we look at a search because you mentioned earlier on there about you know you search for something and the way the the the technologies connect in the background so you could do an internet search for one thing and then suddenly see an advert on a social media platform for the same thing I suppose it's that connection of that data as well which sometimes may
feel like artificial intelligence but it's just some clever marketing going on in the background which there's a blur between whether it is AI and whether it is just an algorithm that somebody says well race search for shoes
y Therefore, let's surface some shoes.
Yeah. And and I think a lot of the examples we talk about in AI AI aren't really artificial intelligence. It's just about putting two bits of data together and going, I can do something with it. Um, but that's not necessarily a bad thing because if I think about it from an educational perspective,
we're probably not using the data that we've got available to us to help students do, you know, the best that they can do. So, you know, it's it's it's okay for us to have the very basic scenarios as as the super complicated ones.
But what about what about generally, you know, I suppose I see some of the um announcements out there, the various companies working on cloud and AI. It seems to be a kind of arms race in in cloud at the minute uh and AI. What what what are your kind of thoughts on that?
Oh, it's definitely an arms race. Um in in the nicest possible way,
robotic arms race.
Um in the nicest possible way because you've got three or four big technology companies focusing on this area. There's a lot of um theoretical research going in as well as practical research. And so over the last few years, you've seen records being created and then records broken about the ability for computers to be able to do things that we think of as innate human tasks. So um you know there there's some standards. So for example, the ability for a computer to be able to understand speech.
We kind of take it for granted now that you can say, "Hey Siri," and ask a question and it'll come back and give you the answer. Or or um what I always do is I'm in the car and I'll say give me directions to a place and a lot of the time it understands what I'm trying to say and then gives me directions you know that that kind of arms race is sorry that kind of capability is coming around because of all the research going on trying to push push the boundaries on those things
and and interesting you thinking thinking back you know previously when when AI and uh I suppose machine learning that you alluded to earlier on was coming into um play in the market years ago. You know, they were kind of looking at things like can they play chess? Can a computer play chess? Whereas now it's more they realize actually yes that's in you know that's that's quite a straightforward problem and they thought that was a very hard problem but now it's more about those soft skills like speech which is kind of
yeah it's interesting because the chess playing it it's it's rules you know playing chess is a bunch of rules isn't it? You can only do things within certain rules but if you start to think about speech there are some rules to speech but not that many people use them these days. PH. Um, so you get into that really open-ended problem. So, you know, let let's talk about the specifics. So, speech recognition,
yeah,
there's a a standard test called um the switchboard test, which is a bank of recordings about people talking to a switchboard. And um you use that as a measure of how well the system can understand speech. And humans are about 95% accurate, which in itself is a surprise because we think we're all perfect and we can understand everything. Well, we can't. We're about % accurate and then we fill in the gaps
or we say pardon.
Um but computers didn't used to be that accurate and you know if you remember the early days of speech recognition you would speak and then you would go back and edit things or at least in my case I would um and it would understand some accents and not some others. Well, we're now in a situation where computer speech recognition is more accurate than humans. So more than 95% accurate.
Yeah. Wow.
So in all of these areas when not into perfection. We're not going to perfectly understand every word of speech. We're not going to perfectly understand um the ability to predict the future, you know, to know which kind of shoes you should be buying or what size shoes you buy all of the time.
Um but we're better than humans and
and I suppose it's speech is the application of that because I know there's been some advances in some of the inclusive technology. So you utilize it's not necessarily about the technology itself like speech recognition, it's about the application of that.
Yeah. Isn't it?
Yeah. And can we help people to do things better than they can
in a a another way. So, you know, if you think about um you upload a video, you upload a video on onto a site like YouTube and then it adds captions. Those captions may not be 100% perfect, but my goodness, for somebody with a speech uh difficulty, they are better than just watching a video with no captions.
Yeah.
And and the same with images. Um
we're now in a stage where technology is or AI is now able to look at an image and tell you what's on that image. So you you'll have seen that if you put an image into PowerPoint, it puts a little tag at the bottom and what that actually has done is take that image, put it to a cloud service that has looked at the image
and then labeled it. And the labeling that comes back is again better than humans. So and that's good from an inclusive point of view because if you've got somebody with visual impairment that's looking at a PowerPoint presentation and going through something, you've got the alternate text to be able to read the back or for a website or Yeah.
Exactly. you know, and so we've made that same leap in the ability of technology to get to the point where it is able to outperform humans and it's able to do it regularly and quickly. So, so my example of that is um I was showing somebody about the labeling of um images and we looked at the the technology in the back end and I said, "Well, okay, so here's a picture and it was one of my pictures from my
um holiday of of um some elephants and I uploaded it and it said it's an elephant standing in a body of water, which exactly was it was a group of elephants standing in a lake. Um, and then I looked a bit further and it said with 80% confidence or 90% confidence confidence, it said it was an Indian elephant.
And I hadn't really thought about where my picture had come from. It was just a collection that I'd had from the past. It's like, oh yeah, that was when I was in Sri Lanka.
And so it's like, yeah, now
I I would look at it and label it as some elephant standing in water maybe,
but I wouldn't have gone to the Indian elephant bit because For me, it reminded me that Indian elephants have ears shaped like India
and that they're smaller and African elephants have bigger ears. So, I then went to find a picture of African elephants uploaded and did the same the same interpretation and said, "This is an African elephant." But that's a really good example of the learned behavior of artificial intelligence to be able to consistently outperform humans.
Yeah. Because I suppose when we're looking at AI and accuracy and things, you know, when when you look at lots of research, you know, a lot of it is to do with but it's got to be better than a flip of a coin.
We haven't got to be 100% accurate with a lot of these things. It's just got to be better. And I think,
you know, when I was listening to something on uh on a podcast myself a couple of weeks ago, they were talking about using um AI when when it's when we talking about um criminal investigations and how actually uh judges can use AI to kind of inform them about reoffending rates based on postcode, which is quite interesting. And it it all starts to become e and you think is this is this correct? But the idea behind it is that it's better than flipping a coin. It's giving that judge more data and and more um kind of guidance the decision, not necessarily making the decision itself. And I suppose when you augment on top of that, you know, like you just mentioned about the fact that AI is now being able to see things and it's been able to listen to things in in a in a better way and almost feel as if it's thinking and it's really augmenting it further.
So hold your thought for the example of about the rear offenders because
okay
we definitely need to talk about ethics before we talk about application.
Yes.
Um but just before we go there probably another example around
that race to improve artificial intelligence is is the other area is about what what I would say is thinking you know so we can see here better than a human.
Um we're now in a situation where artificial intelligence systems can read for example documentation and answer questions on it. better than humans. So, we are in a bit of a race in China at the moment for computers to do better than humans in medical exams. Um, but you know, take a standard piece of text and get an AI to read it and then answer questions about it. If they can outperform humans,
that that goes to the root of of what's happening in classrooms, doesn't it? You know, and and so that's another reason why it's a hot topic now is because we're now just entering that phase where
all across all the domains of expertise. We're seeing this ability for AI to be able to outperform humans. And it's not just happening in, you know, massive supercomputer data centers. It's happening in your spreadsheet tool and in PowerPoint and, you know, in in your everyday systems. It's happening every time you visit Amazon and ASOS and iconic and wherever you go for your shopping where it's predicting what you're going to buy next and putting that on screen in front of you.
And and I like some of the AI the currently avail tools which which gives me an email each week about my well-being.
Um the my analytics tool really tells me, you know, how I've been utilizing my time, how I've been connecting with people. It's actually feels more of a human side of AI and the stuff that you just mentioned in PowerPoint.
Um it gives you those design options and and allows you to practice your presentations. Some fantastic stuff.
Yeah, because big data is what sits behind a lot of this and that can be quite a dehumanizing thing if you think about the early ages of you know, I don't care what your name is, just give me your student number
because that's linked to all those other things. So, it can be incredibly dehumanizing. And AI potentially has got that risk, but potentially it's got the other side to be able to put more personality into it. I was talking to a group of students a couple of days ago and said to them, "So, tell me about the organizations that you feel personalize things to you." And they immediately went to iconic and ASOS.
They said they know that I'm a different size for tops versus trousers. And so, you know, it personalizes things down. to me. And so that's the humanizing side of big data. But, you know, we're still in the early days. We're still in the experimental stage. And and probably now is the time to just take a little bit of a diversion in the conversation
cuz we should probably talk about the ethics of it before before you start solving problems with it.
True. Cuz you mentioned earlier on there about China and obviously there's been a lot going on globally around uh some of the mobile phone manufacturers and things like that and there's a lot of ethical applications of AI generically across across the technologies where where what's the current state at play
well that if I if I think about a lot of times when we're seeing data and AI being used to replace human intervention sometimes people aren't understanding um the true upside potential or the true downside risk. So you know there's a lot going on at the moment in China with facial recognition. Um but if you look at our attitude toward facial recognition as an organization, we're saying that there needs to be regulation and limits on it because, for example, we know that facial recognition isn't as good with women as it is with men. It's not as good with people of color. And so, the risk then of making decisions based on AI interpreting data in a certain way,
well, there could be incredible negative consequences.
Yeah. There's also cultural connotations as well when you talk about China there and their use of facial technology, you know, I was listening to uh an interview with with somebody inventing or developing sorry um smart cities in China and their attitude is very different to actually saying well actually the more data we've got look how efficient this could be. It's a very different mindset around use of data and privacy. Um so it's really interesting to see where the different countries would drive that.
Yeah. And so we've seen in different places different projects different things going on. So
let let me pick some examples. So we've had a a recruitment system that was using artificial intelligence where what it did was amphas um amplify the human bias. So what it ended up doing was building a recruitment system that more or less prioritized men into the recruitment process because that had been something that had happened in the original organization and so you train these systems based on the data from the past. Well, it turned out that the human bias was in there but that then accelerated through so that you could do it much faster and deploy the bias at a higher level. So you have to be really careful of that. That that example exists. The other examples are school district in the US that used a AI algorithm to work out what it thought were effective and non-effective teachers based on the results of students
and then deciding which ones to sack as a result of it.
So you know there's two things there. One is is that data good enough to do it? Well, you and I both know that the performance of a class might be down to is it a wet Friday afternoon as much as it is about any of the teaching that's going on. Um, but then the second thing is, can you possibly use AI for things that are profoundly going to impact somebody's life in a negative way without having a human in the loop looking at it? Um, and
and maybe that's also about the um the group mentality, you know, so so sometimes people feel happier when uh AI is implemented across a group of people. So saying, well, there's a demographic or a particular post code that has got a an issue with crime. or whatever and therefore we can put something we can use AI for that because it's a group it's not focusing on an individual whereas as soon as we start focusing in on
ry
y
from Australia then we then people start to get a little bit uneasy about it
but even in that group level if we if we go back to the example you talked about earlier the uh rehism example that's a word I've still not learned so that's about reaffenders yes so that the where they're using it is in a court system in the US where they have a very popular algorithm that is used to work out whether it's somebody is likely to reoffend. And the danger is that when you look at the data that's being used for that algorithm, there's some stuff in there like the postcode of the defendant. Um the they don't use racial profiling, but they use postcode. Now, postcode in the US is more or less racial profiling.
So, there's a great report out um from CSRO and data 61 about the use of AI and the examples and they talk about this one where um that's exactly the problem is that you're using a proxy for race in your data analysis. So automatically people of color get disadvantaged by the algorithm
and so as a result it was then saying no don't release them keep them locked up
because they're more likely to offend but you know you can't you were born Welsh I was born English we can't change that but it doesn't necessarily define the characteristics and the decision-m we're making at this age.
Yeah. Yeah. And I and I suppose I suppose some of the some of the um bots that are being created these days as well where you put in the knowledge base you know the bias that goes into the data sets is quite an interesting topic as well right
yeah so um you know we've done we've done some work around bots to provide chat interfaces to um well I I work in universities all the time so the the definition of a university is a is a group of academics united about problems with parking and so you build a bot to deal with parking yeah well that's pretty easy but if you start to go too far with it. So the example that we had is in China, we built a bot called Tay that we had to remove 24 hours later because it was designed to learn off the people that it was talking with. And the people that it was talking with were on social media that were semi- anonymous. So it learned their characteristics and within 24 hours the bot become racist almost.
Um fat forget almost the bottle become racist. So you have to have that eye onto the ethics of the situation about what might be the unintended consequences of implementing AI as well as the intended unintended consequences. And so, you know, uppermost in our mind as we're thinking about AI is not just the what can the technology do, but it's also what should the technology do. So, you know, we've specifically said there are cases where we're not going to allow our technology to be used for facial recognition in some law enforcement areas because we know that it is good enough or it has some blind spots in the training data or biases that will then um disadvantage a community.
And I suppose it's important for the folks listening to this podcast as well because everybody's going to be coming to this from different um learning point of view. Some people will be implementing these technologies. Some people will be making really large decisions about these and bringing that ethics and the kind of your own kind of lens on this in your institution is vital, right?
Yeah. You don't need to become a dedicated ethicist. But you do need that. In an AI term, it's called human in the loop. But you do need that somebody with that perspective when you're involved in these projects because, you know, I'm I'm a lifelong IT person. I'll admit it. We sometimes don't get humans.
Sometimes we'll we'll do stuff with technology and forget the human impact of it. And so you need a process that isn't just about the programmers doing stuff with data because that could go terribly wrong. So you do need to have that business level conversation
that is about the upsides and the potential downsides and mitigating those risks exactly as you do when you put up a fence or you build a swimming pool or you build a new building. You need exactly that same thing with IT things. And I think that's maybe a new area for it to be focusing on and it isn't something that it can do on their own. You've got to have a shared accountability across the organization.
Yeah.
With people that understand it and You know, that's part of the reason why it's important to talk about AI with everybody, not just with the IT people, because we need a better understanding of it across society and across different parts of organizations.
Yeah. Yeah. That's a good point. And the governance around it is is vital, I suppose. So, if we start to elevate this and start thinking about it from an educational point of view,
well, let me stop you because Okay, go on.
Because because you're the ex-teer and the ex inspector.
Okay. So, you tell me Why why is it an important thing in education? So bearing in mind that conversation.
Yeah. And and it's an interesting one because we we we've always talked about personalization of education uh over the years and and you know every individual matters in education and and I suppose it's always been difficult to do that. You know schools are and universities are uh campuses and systems full of data whether that's data about postcodes and gender and you know, numerousy attainment and all all, you know, a plethora of data and is very very data rich, but we've never really had the tools to bring things together. So, there's been um a lot of elements with inside the school systems where we'd use learning management systems and school information systems or are often disperate um tools that wouldn't really bring that data together and allow us to correlate data between, you know, relationships between post codes and naplan results and things like this. So, I think schools are really data rich and they always have been but we coming across a tipping point now where um we able to and the tools have been able to be dem democratized so that people can actually start to you know look at their data quite quickly um whether that being inherent tools like Excel with the AI that's built into that to look for insights and sentiment analysis and I don't know if you've done some work on that and we'll explore that in a second but then also the fact that um you know we've got this data together and some of the other tools which allow us to visualize that data and the the technologies allows us to start to warehouse that data centrally to be able to interrogate that data at quite a technical level but then users at the on the ends of that data warehousing component I suppose can actually you know a teacher can actually use natural language to inter interrogate that data and say well what are the interventions that I need to use in my year seven English class where are the gaps so I suppose it's it's the it's it's a I suppose it's a perfect storm really. We seeing more data being used more effectively in datari environments and we see the tools that we can interrogate that data being more freely available and actually easier to use.
So, so from an education point of view then we've got lots and lots of data more data probably than other organizations.
Yes.
You know more
um I was talking to some university students the amount of interactions they have with their university is way way more than they have with the iconic, but the iconic were people they thought that personalized things really well. So, so there's that data opportunity.
Um, and then there's the simplification opportunity which you've talked about which is
that more people can use this now than used to be. You know, you don't need to be a data scientist
and with that the costs have come down as well. So, so it's important that the business decision makers in in the schools and universities and and and school systems can actually now start to think well um how can I do this at a cost-effective scale to give that data to the front line as quickly as I can? Um
and and probably the other thing there is we're now becoming much more aware of the hidden patterns in the data.
True.
You know, the patterns that that we never used to see.
You did some really interesting research on that in Queensland, didn't you?
Well, so I did um so this is probably coming into the hidden patterns, but it's also coming into some of the cost and simplification bit. So um two years ago, I took an AI course And I learned how to build machine learning models. And it took me about two and a half months of a bunch of evenings building building algorithms to predict stuff.
Yeah.
Um and it was really fascinating. But the technology has moved on so fast that almost all of that has become redundant now because now I can just go and put a data set into a tool that will go and do that modeling for me.
And um one of the things I've always wanted to play with is
can you predict nap plan performance which is just one facet of academic performance but unfortunately it's the one that
nationally we see
yeah when we look at look at personalization that's one of the key
you know what data do you need to predict
I'd love to do that right we've covered some of the basics today and we've run out of time so let's pick that up next time and let's explore how AI and the concepts we've talked about today can kind of connect in with education
okay perfect you next week then
see you next week