LSE100: Artificial Intelligence transcript

[Music] I think we ought to start with the term artificial intelligence. It needs debunking. It's a respectable academic term which has been hijacked by marketing departments and the media. One definition by Margaret Bowden is using computers to replicate how minds work. By this definition, very few indeed are actually using AI outside the military, the intelligence and possibly in the research labs of the top technology companies.

Most are using what I call robotic process automation or cognitive automation. They are using machine learning, algorithms, statistical techniques, image processing, natural language processing backed by very powerful computers and memory. What has made the difference in recent years that causes people to call it AI, it's in fact very weak AI is the power of the computing uh the speed with which they can process and the power of the memory uh to store vast amounts of data.

Artificial intelligence it's a very catchall broad term and machine learning is part of that. Tom Mitchell who was a pioneer in machine learning said that machine learning is the study of computer algorithms that allow computer programs to automatically improve through experience. That sort of captures the the idea that they learn through experience. They learn from data. AI has advanced rapidly in recent years. AI or as it used to be called machine learning was really a sort of statistical pattern matching process that would for example extend to things like facial recognition and such like the previous head of the Turing Institute for example which is the National Center for Artificial Intelligence used to joke that once all the hype is over we'll just go back to calling AI statistics.

What's been a game changer are the large language models where you can now interact with AIS that understand natural language and have been trained on massive corpuses of data with huge computational power. The computational power and the data on which they've been trained has has led to a real sea change in the performance. But it's not just that it's the mode of interaction. Some say that it's a fourth industrial revolution after steam, electricity, the internet, now AI.

And I think there's some truth in that. As an economist, I think of AI as an example of what we call GPT or generalpurpose technology. Other example of GPT in history where things like the steam engine or electricity, this term general purpose technology is trying to target this sort of technologies that diffuse extremely widely in the economy and so they can spur innovation and productivity gains in many different sector.

But at the same time it was the idea that they can help innovations in specific sectors. When economists think about this idea, we tend to use the word spillovers. It's this idea that an innovation such as AI is going to spill over in many different sectors. I think most of the fundamental issues that people are worried about about AI and new technologies generally are fundamentally social scientific. There's already enough going on in terms of the implementation of AI across society as a whole that people are worried about certain issues.

Is AI being used in a a safe way? Is it being used in a way that people understand? Is it transparent? Are the decisions opaque? There's an inevitability of uh some level of hallucination with these systems. They will always always confidently get something wrong. It has huge environmental consequences from the mining of rare metals through to the sheer energy use and the water use of data centers and such like. We think of AI as something that we do on our shiny computer in our office.

But there's a whole ecosystem, a whole underbelly of exploited labor and exploited environment that underpins it all. Will the concentration of AI and therefore the concentration of certain types of jobs in certain regions have some influence on international migration? What can we do to use AI effectively to strengthen democracy rather than to undermine it? So there's just about no area of social science that doesn't have something useful to say about artificial intelligence.

There are all sorts of interesting questions about the effects of AI on inequality. Unless this is carefully uh shepherded, we could end up with even more concentration of wealth. We need to make sure that the developments such as they are are going to be a benefit to as many people as possible and not just concentrate power and money and so on in fewer hands because AI you know it's been described as a sort of tool that can magnify power can either use that to bolster existing inequalities and inequities or guide it in such a way that that the growth that would result from AI is is as fairly as possible spread around the population.

AI is not just a thing on a computer. It's it's a it's a physical social ecosystem. But it ranges from very high paid data scientists working in shiny offices to people in East Africa being paid slave wages to do routine mundane data labeling. Who is guiding the development of AI? Are we happy for it just to be guided by technical companies or do we need more government and importantly citizen uh steering of this? And if you're just playing on with the technical development of systems like this, that might not be your primary focus.

I do worry about the asymmetries in the marketplace with regard to automation and artificial intelligence. The top technology companies have very large R&D budgets and a huge interest in gaining the most they can from these emerging technology, but they have been voraciously eating up the skilled talent available globally, which means that that's going to create skill shortage elsewhere, which means that they're going to accelerate their ability to do automation at the expense of others ultimately.

And a lot of the studies show that those who get ahead stay ahead. We can call it war of talent but certainly you know building the skill base that others can't replicate fundamental way of of gaining competitive advantage. AI is one of these things that I think is really exciting and seems to be kind of magical and we forget that it is actually based in material places and basically computers. So everything that is happening through AI is happening through actual computer sitting somewhere.

It's not happening invisibly. This is the same thing with all of our cloud computing and cloud data. The cloud makes it seem like it's like ethereal. It's not. It's in a place. There is some sort of server farm somewhere that is doing all of this work. Data centers take up significant amounts of energy. Often times it's just extremely expensive cooling expenses. The amount of electricity they're consuming to process all of their algorithms and data is enormous.

There was a company that designed a robotic hand that was running this algorithm and learning how to solve the Rubik's cube. And while this is extremely impressive, it took more than a thousand desktop computers and a dozen other machines running specialized algorithms and graphics chips doing intensive calculations for several months. So if we're thinking about all of these interesting feats, we have to think about the materials that go on behind them and all of the actual physical energy that is consumed in this.

Some of the main actors in developing artificial intelligence are actors who also have economic interests in collecting large amounts of data. The reason to collect large amounts of data for an actor like let's say Google is in fact to train its artificial intelligence. The reason to develop an artificial intelligence that can be trained is to facilitate the rest of Google's business model which is to sell large stocks of aggregated data and also to provide personalized advertising.

That's where the power differential comes in. So these systems are designed to gather data from a set of people and then to sell it to as many buyers as possible. That doesn't mean that the result of that data collection isn't valuable for the person who's using the application. It just means there's an asymmetry there in the kind of power. We could talk about health data or care data as examples of data that would be more problematic and need different kinds of governance mechanisms.

With medical data, it is of collective value. So, can we give it away? If DeepMind's algorithms are going to become really valuable to DeepMind, but they're trained on data that was not consented to be used and in fact was part of potentially a repository of collective and public value, then we get into a really really tricky controversy. So there have been a number of proposals in this area. Some of the proposals are for things like a personal data box where you hold on to your own personal data and you share it out with permission to all of the different kinds of applications that might be asking for that.

That's one kind of solution at a personal level. You can also have a collective solution which is the solution of a kind of data trust that holds a lot of collectively valuable data together and determines how it's to be used for the benefit of the overall total beneficiaries. And this is something that the open data initiative is working on here in London. There's also a third proposal which is a sort of data commons where people who want to share data make a decision about how they're going to put that data together and determine how it's going to be sold, shared, given for again for a kind of collective benefit.

So there are these different models that you can think about that are sort of um holding uh different forms of space to protect data in relation to these automated systems in different ways. I think it's good that we have now proposals for people to kind of hold their own data, but I'm not sure that this really gets to the kind of crux of the matter. Your data is less valuable for you than you think. Our data is valuable together.