Understanding the Realities of Artificial Intelligence. Part 2
In the second part of the interview on the realities of artificial intelligence Adi Hazan, an AI thought leader, and Mark Hillary, a British technology writer and analyst, discuss the practical applications of AI, emphasizing the need for human oversight and cautioning against overhyping capabilities. The conversation touches on various scenarios, concluding with a call for realism and common sense.
Mark:
I suppose there are use cases where it’s not as consequential. For example, we’ve talked about things like insurance claims and driving there could be life or death decisions we’re asking the AI to make but what about if I’ve just bought my new Samsung TV and I’m having trouble setting it up and I go onto the Samsung chatbot and I’m asking it questions about how to set up my television? In that case surely we’re going from the useless chatbots that we’ve all experienced when trying to interact with customers a service in the past to a generative AI system that should in theory be able to answer any question about that product.
Adi:
It will generate everything that we’ve told it about the product. This term generative AI is really misleading and it’s intended to be misleading. What I will say is this. Let’s look at narrow versus broad AI. If I have a television set with 50 settings and however many possible combinations, that’s a narrow case. And the more narrow your case, the better AI works. So top of the list RPA is. Let’s say, if you need to read a bank statement and if it says 1,000 there is no chance you ever wanted to read anything other than 1 comma zero. It does that well. If you want to know if 1,000 is good for you or bad for you, that depends on your business. It doesn’t know your business. So ask yourself how routine is this process, how much human intelligence does it need and the bigger the answer is the smaller the use of AI will be in that case. The other thing to remember is that AI doesn’t have to stand on its own.
A thing that I really like is the use of the term augmented intelligence. Your human being looking at those two cars may not know what a tail light costs the computer will so if a human being had to see that picture and say hang on it’s been awarded total loss – no no – edit that bit out. Ah that’s the price of a tail light for that car – go for it. Working together. Human beings are very good at understanding and flexibility. They’re very poor with memory. We forget a lot of things. Computers are very good with memory and very poor with other things. So let’s take diagnosis. In one of our chats it was brought up that somebody had some rare tropical disease no doctor in the city had found it, but chat GPT worked out what it is. A lovely story! Nobody tells you about the 2,000 people that go into Chat GPT and it speaks complete rubbish, and kills them because that’s not a good story. But I will say this. If you go to a good doctor who has studied for 20 years, not four minutes on Google, and they also refer to chat GPT, it might remind them of an illness that they would have overlooked – so wherever you can pair the two up. I’m a big fan of full self-driving as long as you know. In my car, if I’m driving and I start slipping lanes, my car rattles. It warns me and I like that. Sometimes it warns me when it should, but I’d rather be warned when I shouldn’t start drifting and I think it happens to everyone. We don’t realize, but it does happen to everyone. So there’s a huge case to be made for modeling, for insights.
Mark:
That means that, if we have a very narrow data set that we can train the AI on, then it should essentially be reliable like the television example, how to set up Wi-Fi on a smart TV. It shouldn’t actually be that complicated. There’s only a few variables where it may need to ask you, ‘Have you checked this or have you checked that?’ Whereas, if I was actually going to the American IRS and asking them about the deductions I can make on my tax this year, then you may find that even though there is the letter of the law and there’s probably a 2,000-page volume that describes the tax law of the United States, but there’s all kind of human variations and what different tax lawyers would consider acceptable, then you introduce the human element into the law itself. So it seems almost impossible to be able to train a bot to answer those questions because there’s a certain variability there.
Adi:
Look, first of all, what I mean just on the specific case of law. A lot of laws are deliberately phrased with loopholes. So one has to accept that. And they’re intended to benefit certain people in society they don’t really want it to affect. People who don’t know the loopholes, pay full tax. We all know that the super rich pay a fraction of the tax that normal people do, so that’s a start. You say, using a narrow AI, but you understand that to train any AI you need a lot of data. So using the standard algorithms, if you want to make it too small, you kind of run out of talent, if there isn’t enough meat there. And it’s Catch-22.
What I would say is this. Make sure it’s that’s not a great subject for a large language model because in the definition of it, it has to be large. You want you know we have large language models, we have expert systems, we have predictive systems. They’re all a little loosely defined, but check that you have a system that A: works. B: it doesn’t need $50 million to find out, if it works. C: it doesn’t only work for you. I need to go into that one for one second.
My sister’s house has no switches in it. You talk to the house now. Switches would have been a nice backup, but no – no. They’re ugly. My poor Mom – God bless her – is a very mature Bulgarian woman whose English never quite got there and she is standing in the kitchen and saying, “Alexa, de lights!” – Bing – “I didn’t get that”. “Alexa please, I’ve been here three weeks just switch off de lights!” – Bing – “I didn’t get that”. It only works for the people who made it: who are Anglo-Saxon, white, with a certain accent, etc.
Just be careful when you buy it or when you invest in it that you’re not the only people it works for. And lastly, if it doesn’t work, have the courage to put a stop to it. Don’t fall into this trap of saying “Well, it works for most of us. It works, but…” keep yourself a fail-safe, keep the switch there, so that you can always fix it, which is the Human-in-the-Loop. Is this wrong?
Mark:
No. I was just going to say this sounds very much like one of the problems of bias. It can only be aware of what it’s been trained on. For example, if we are running a human resources department and someone says, “Look, here’s a great AI system that can scan all those thousands of CVs you’ve got arriving.” Then you tell it, “Okay, I need a great software engineer.” It will essentially not find any software engineer that doesn’t look like one that’s already been in the system in the past.
Adi:
Again, like I said, it can’t originate. It also cannot understand factors that you have not given it. There’s a documented case of this one of the biggest companies in the world who are one of the biggest AI specialists in the world, Candid. They just threw it away. That’s brave and honest, and publicized the case, which is also brave and honest. But you have to understand, even if you ask ChatGPT to write you a piece of music, it’s got nowhere to get it from but the music that came before. You will never get something original because it doesn’t even know what music is. It knows music is Word #800. It goes kind of into the 800 Zone. You add a few more words, it goes into a couple of zones and couples together what it thinks you will like to hear from what it has seen. This has nothing to do with thinking certainly, not with understanding. So narrow models get the right technology. Don’t overspend on a proof of concept.
And the other thing is scaling. These things very often work well in small quantities and don’t grow nicely, so facial recognition is a great example. When they do the demo for you, it only has to decide, if it’s Mark Hillary or not, but afterwards it has to decide, if it’s Mark Hillary, Adi Hazan or 50,000 other people, and that’s when it fails. So make sure that you’ve not shown something that works only when it works and fails all the rest of the time.
Mark:
I suppose this is why it feels dangerous for police services to be using this kind of technology although it may work 99.9% of the time.
Adi:
No, Mark. I have to interrupt you. It does not work 99.9% of the time. Facial recognition on a good day works 80% of the time. Out of the James Bond movies, it doesn’t work. The system that was rolled out in China that monitors the whole population does not work. I’m sorry to say it and a system that works in China will not work in Africa, and a system that works in Africa will not work in the West. And it’s almost impossible to believe that all these people are selling stuff that doesn’t work. Your Apple works. It recognizes your face, but remember, it only has to know, if this is your face and not someone else’s. That’s one tiny decision that it’s gotten good. So sorry to interrupt, but again and again, we’ve all been trained now to say it works 98%. It really doesn’t and the reason that the EU has banned facial recognition in public is not because they want to be do gooders. It’s because countless people who had done nothing wrong were being arrested and detained incorrectly, and not one or two.
Mark:
We’ve been talking about the problems of limitations and I think that you know clearly this is what executives that are thinking about this technology need to understand. Where are the limits and what should they be wary of? What are the use cases and where it really does work effectively? I’ve seen a few examples of some very good actual projects that have rolled out using AI. Quite often, they seem to involve a human in the loop somewhere. It’s always about augmenting what a human does rather than completely handing over to an AI.
Adi:
That’s exactly it, Mark. Please don’t think I’m sitting here and saying everything everyone’s made is trash because it’s not true. Really, what they’ve done is half a miracle. First of all, you are using it all the time. Like I said, when you touch your phone screen, it tries calculating where you wanted to press. You’re not that good, no. None of us are. But really, if you’re going to sit around like War of the Worlds with your other directors and look at each other, and go, the problem is of course the humans. It’s not that way. The humans are very good and you’re not going to be getting rid of them anytime soon. And anytime somebody tells you that something is going to disrupt the whole way you do business, just look at the last 50 years of your life and think how much was disrupted really. Not a lot. Basically, the video stores and the post office are gone. Everything else is much the same. Self checkout? Catastrophic. Self-driving cars? No. The algorithm that recommends stuff on Netflix? Not a word. These people are impervious to shame. But that’s where you come in, you sit, and you say ‘yes’ that, ‘no’ that.
So human-augmented intelligence works all the time, works well, and we help each other, and we work together. You shouldn’t be doing any projects that you think are going to replace a team. I’ve never seen a project replace a team. I’ve never met anybody who says they have. Maybe it’s time to be more realistic and more human-centric. We do have our weaknesses and these machines can improve our weaknesses, which is great. But they’re not substitute and they will never be.
Mark:
I have a company that I’ve worked with in the past. They are one of the largest customer service advisers in the world and they rolled out a new system recently.
Traditionally, they would have a big contact center full of people that are receiving calls and emails. The email team would have to just literally sit there and read every email coming from a customer, formulate a response, and then reply to that customer. They put a sort of an intervention in there, so the emails come in, but then they go through the system, which then highlights to the human agent – here is the main point of this email, this is the customer problem, so that goes right to the top, even if the customer has written 500 words. Then they will have 10 standard replies and the 11th would be having to do a bespoke response. But if they just hit Reply#3, then the system will actually create the reply for them. So you’re automating the whole process, but keeping a human right there in the middle of the loop as well, and that seems to work really well. There’s a huge percentage increase in productivity.
Adi:
Wonderful! That’s exactly the kind of thing where you’re using computer not intelligence, but abilities and human decisioning and intelligence. That’s the perfect combination and that’s what we should be headed towards. But when somebody comes to you and says all you’re going to do is dump all your spreadsheets here and it’ll know what to do, it doesn’t know what a spreadsheet is, doesn’t know what you do. It’s very easy to differentiate between fiction and reality, if you know what you’re doing. Again, I agree with you, Mark, there are countless process automation, mathematical modeling, predictive modeling, and it can save people from all sorts of mediality that is beneath person. Why should a qualified doctor have to sit and read every email and decide if they’re going to go to A&E or to a doctor or to some other practitioner, and not just a doctor, but a senior doctor, when you can triage a third of them? Give them the difficult ones and free them to see people. Those are the places where AI tops: augmenting people, improving the quality of work you don’t need.
I had a client that were Telco. They used to download reports of every cell phone tower, and engineers would sit and read these reports, looking for problems. Horrific work! They had to pay them double and a half the going rate. No one lasted more than six months. So you’ve got constant churn. It’s a terrible job. What do they do? They train little bots to read these things. The bots only alert them to the problematic ones, so that they can think of the best solutions, and they have free time to make more bots and start checking better and better. The outcomes of this project were phenomenal. It didn’t have all the glamour that the driverless car had, but it works.
Mark:
Looking for anomalies is a use case that can work really well. That’s like network security whereas you used to find that a company that had been hacked would only find out about it a month or two months later once the hackers have been in there and emptying out all the sort of data that they want to get their hands on. Now you can actually just apply some bots to scan your network and highlight whenever there’s any kind of unusual activity.
Adi:
You don’t need to be a detective. If I’m your sales analyst and I’m logging on in the middle of the day from my workstation and asking to read up all the sales file and analyze them, it makes sense. If I have a brand new user logging on at two in the morning trying to upload all the sales data, if we were applying any natural intelligence, we would see it immediately. This is not a matter of some ungodly power that will look at all your numbers and find stuff doesn’t exist.
There was a huge project done in the US long ago where they looked for correlations between all the products, and they came up back in the day with a correlation between diapers and beer. Someone had to make up a story to prove that it was right and the theory became that young dads when they get sent out for diapers, grab a few beers on the way. And for ten years, diapers and beer were always put next to each other, but there was no correlation. It’s just, if you run enough correlations, one of them will be positive. So people who don’t know statistics wasted a fortune and achieved nothing. Really, if it doesn’t make sense to you as an expert. be very weary of trying it. Don’t say ‘don’t worry’. The car knows what it’s doing. It’s going to miss that truck’, unfortunately it’s not.
Mark:
Do you think that what we really need is a new type of digital literacy that encompasses some awareness of AI? Because it seems that our education at the moment is just coming from the mass media, which is almost entertainment. Every AI story in major journals is always accompanied with a picture of the Terminator or something. We used to talk about having Excel skills on our CV or email skills and now basic digital literacy like using email is just completely accepted as what any professional can do. Do you think…
Adi:
AI prompts a new skill.
Mark:
Exactly. Are we looking at another whole set of digital literacy then?
Adi:
No. We’re not. It’s going to end up badly. I think, what’s needed is a wave of honesty. One must get up and say we promised it you five years ago, three years ago, two years ago. I don’t think I’m going to pull this off. We have the finest mathematical minds of our generation sitting, working out how to make people click on internet banners. It’s time for my industry to reclaim a bit of its dignity. Galileo wasn’t killed, so that we can make people click on banners, pull yourselves right people, and on the flip side it’s going to come anywhere. If it comes against our will, if it comes and we’re not leading the story, it’s going to be much worse.
Cruise, the company the countless articles have been pulled off the road. This is a huge damage for a company it could have been avoided with a little bit more honesty and a little bit better work, and a few more delays. But unfortunately, they’re under pressure from Tesla. Everyone’s under pressure from everyone. There’s all the noise. We all are going to have and that’s why my message is what it is. Let’s just use common sense. Let’s release these products after they work, not before. This notion of fake it till you make it is actually a terrible thing.
I mean it’s all well and good, but this poor 39-year old girl Elizabeth Holmes is now sitting in jail with two kids out of jail. Henry Kissinger’s 5 billion bucks are gone. No one has benefited here. And if somebody does come up with a way to test, but better they will be dealing with skepticism. I think people need to wise up before they end up in jail and I think we do need to that companies need to litigate when they get ripped off. It’s embarrassing to say ‘look, we spent all this money and got nothing’, which is what you have to say in a lawsuit. But it’s becoming time to defend ourselves as well.
Mark:
The final point then. I think that it’s worth just summarizing that sort of key point because every business is different. You talked about Theranos there, trying to completely reshape blood testing. A bank is very different to a healthcare company and electricity utility is very different or a retailer. How do executives in a very wide variety of different types of company imagine what they can possibly do with AI then? The possibilities are endless so long as you can actually see it working first?
Adi:
First, the possibilities are not endless. Human creativity is endless. We have to go back to common sense. We’re busy rolling out into a thousand retailers in the UK and they don’t rush to do anything. They do something small. They test it. If it works, they grow it a little bit and after two years everyone’s getting impatient, except me. I’m sitting there. That’s how you should do it. And they will pay us, if we work at step A. They’ll pay us to go to step B, but never beyond that. They’re ready to pull the plug at any moment. That’s mature stable thinking.
Banks are doing an excellent job. You’re not going to roll into a bank with some weird AI and create havoc. Or if you do create havoc, very often the bank doesn’t mind. In other words, if you make a mess of the call center, they’re delighted because they want you to join, to go in online. So there’s a lot of creative destruction happening as well. But yes, I think a lot of companies are doing this maturely, slowly, carefully and watch what you buy, and make sure it makes sense and works, and works on scale, not just in small. A great advantage is to be heard.
Mark:
Common sense on making sure it works. I guess that’s a great way to wrap up.
Adi:
Humans are okay. You are not as dumb as they’re telling you are. You’re probably a good CEO and you probably do know better than a machine what needs to be done. Go humans.