AI and Economics

Inspired from Prediction Machines by Ajay Aggarwal, and Prof. Susan Athey’s Lecture “Artificial Intelligence: The Economic and Policy Implications”.

World through an Economist's Eyes

Economists try to model the world through the lens of market forces—demand and supply, production and consumption, prices and costs. This way, even if the world around us changes, we would still be in a position to make sense out of it. We try to cut out the hype around a fad and extract the essential juice. So, if we discover that Hogwarts was real and we were living muggle lives all this while, it wouldn’t change the fact that demand is a downward sloping curve and that supply is upward sloping. In fact, we are sure the goblins at Gringotts use the same accounting principles as we muggles do. 

Thus, new inventions don’t scare us. We will never be uncertain of how markets would adapt to evolving technologies. For example, when the internet came people termed it as the “new economy”. The world shrunk in size. Money transcended boundaries. Everything was a click away. But economists weren’t as star struck as others. If we peel off the magic out of the internet, what internet essentially did was make communication and information abundant. And when a good becomes abundant, its price falls. When prices fall, we can focus on better things that weren’t possible earlier due to huge costs, which acted as deterrence for innovation. This gives rise to inventions that wouldn’t have occurred to us with the prevalent prices.

Take this example. When the early man invented the wheel, transportation became easier. But it wasn’t just transportation at which they stopped with the wheel. Construction became less tedious too. And so there were architectural marvels!

Another one. When Edison invented the light bulb, it immediately made light cheap. It got so cheap that now we don’t even give a thought before using it. There is virtually no difference between night and day now, and the possibilities it brought would have been unimaginable then. Now, there could be closed offices where artificial light would lit up the workspace even at night. Mine workers could work without the fear of mines catching fire. Light bulb doubled our day. It brought down hazards from conventional oil lamps and enhanced our productivity many folds. 

Impact of Cheap Arithmetic

Now consider the advent of computers. Personal computers were a piece of marvel to everyone living in the 90s. But not so much for economists. After cutting through the hype around them if we view computers from an economist’s eye, we could say that computers simply made arithmetic cheap. When arithmetic became cheap, we started using more of it and hence it became abundant. Calculations became faster than ever before. The possibilities it brought are right in front of us. We put man on the moon. Cars became automatic. And some things that weren’t even remotely associated with calculations, became an arithmetic problem too, simply because it was so cheap to do so!

For example, music. Everything that takes music to get to your ears—from its production and distribution, to digital signals reaching the speaker—is made possible due to the idea of making music an arithmetic problem. Where one had to buy a ticket to a music concert, now can listen for free as many times as they like. Photography, which used to be a chemistry problem with traditional cameras, also became an arithmetic problem with digital cameras. When we see it this way, things become easier to understand in the context of every technology in history. Even artificial intelligence. 

The Stage has been Set

What does AI do? AI uses the power of probability and statistics to predict outcomes—which song to play next, what TV show to recommend, which product would the user be interested in, or which word did the user mean to type. Thus, we can see AI will make predictions cheaper. The fun thing is predictions will get so cheap that we would be using it to solve non-prediction problems too, just like we saw before in arithmetic.

For instance, consider translation. Translating languages used to be a job where one made use of predefined grammar rules and vocabulary of both languages to translate one from the other. However, Google Translate and Facebook use AI algorithms to translate languages using statistical machine translation (SMT) models (this explains why they ask for feedback after translating a sentence: to train their models better). Instead of translating a sentence purely on the basis of syntax and semantics of the language, these algorithms see which sentence in the target language is more statistically likely to be the correct translation of the given sentence than other alternatives. The one with the lowest standard error is finally displayed. Thus, language translation became a prediction problem too.

Another example is of automated cars. Self-driving cars isn’t a new concept. They have been existing since decades already, albeit in controlled environments. Forklifts in warehouses became automated long back. Engineers programmed their locomotion according to the detailed floorplans of the factories using simple “if-else” logic. They worked well because the number of variables encountered by them were limited—turn right at coordinate (5,3); place retrieved items at (0,0); if there’s an obstruction, stop.

However, using such a rudimentary logic for automated cars that are meant to run on real roads is not only cumbersome but dangerous too! It would take a very large number of if-else statements to model its logic. Even after that, we can’t be certain it would work well. There are endless possibilities of mishaps on the road which couldn’t be accounted for beforehand. Hence, engineers thought of modelling it in such a way that the car operates and learns at the same time. AI was the apt technology for this problem. Thus, driving too became a prediction problem.

Still AI algorithms aren’t quite as sophisticated as they should be yet. We have heard of automated cars crashing into pedestrians, Amazon suggesting bizarre products, Google Translate failing to translate complex phrases, Siri not recognizing voice commands often, and autocorrect committing embarrassing mistakes. But let’s imagine that these algorithms actually get better. As a result, they would become cheaper and more abundant. This would give rise to new technologies that would use AI as base, and may sound extraordinary today.

Amazon's Anticipatory Shipping

Interestingly, there are already a few innovations that are set to make a mark in a few years. One such being Amazon’s anticipatory shipping. Amazon uses a recommender system to suggest items to consumers that they might want to buy. It uses a host of variables to capture user’s preferences. However, the recommender system has a very small accuracy—only 5%. This means a user would select only 1 out of 20 products displayed, which is not bad considering there are millions of items in their database.

But let’s suppose it has got reasonably accurate now. The recommender has learned enough about you through your history of shopping and several updates to its algorithm. So much that you could add, say, every second item to your cart—50% accuracy. Amazon thinks it would be tiresome for the user to purchase so many items individually. That’s why they decided to capitalise on their recommender system here and, thus, came up with the patent of anticipatory shipping. Here, Amazon would just ship items present in its recommender to your address without even you ordering them. If somehow you don’t want any of the items that appeared at your doorstep, there would be a delivery truck making rounds every week or so to collect returned/damaged items.

Sounds like an episode from Black Mirror?

Jobs and AI

The future looks promising with AI. Predictive keyboards would soon be able to churn entire articles, apart from just autocorrecting and suggesting only the next word. We wouldn’t have to bother about opening our devices anymore for shopping, let alone physically going to a store.

But what’s going to happen to our jobs?

It looks like AI can bring about massive unprecedented changes to the market that perhaps many of us aren’t prepared for. When computers and the internet appeared in the market, people were worried about their jobs going to robots. And their misgivings were right. Stenographers went out of their jobs, postmen were laid off, and telephone operators were suddenly unemployed. Almost all the jobs a computer could do were replaced by the computer. Or to put it in economic terms, the jobs that acted as a substitute to computer’s abilities were indeed substituted by it. As computers got cheaper, the demand for these newly dispensable jobs declined, and their prices (in this case wages) declined too. 

But the reverse happened with people who complemented their skills with the computer and the internet. Accountants who could work well with spreadsheet software didn’t lose their job. Traditional designers who were quick enough to learn designing software tools secured jobs as illustrators in publishing houses. Software engineers became highly sought after in the IT industry and are still in much demand nowadays. Thus, the economic phenomenon that demand for a good is directly proportional to its complements stood the test here too. 

The above example holds similarities to the present job market too. Many jobs today require human judgement. Managers hire candidates for their firm after diligently analysing their resumes and performance in interviews and tests. But this job could easily be turned into a prediction problem too.

In fact, some HR managers are using resume shortlisting software even now. They use text mining algorithms to shortlist resumes based on firm’s requirements fed to the software and they retrieve only the resumes that match these requirements. Similarly, other tasks of a manager are also in peril of being taken over by AI. With prediction software tools becoming so much better, companies are eagerly considering replacing human managers with AI tools completely in a few years. 

One of the prime reasons for doing so is the prevalence of bias among managers. Humans have bias. A manager might recruit people based on colour and other such irrelevant background information. However, with prediction machines, we can program them not to have any bias against such information. This is a win-win situation for both the firm and the candidates. The firm would cut cost by saving on wages paid to the managers, and applicants would be certain that recruits in this firm are selected only on merit. 

Another interesting and somewhat disconcerting example is that of the judicial system. A judge employs her years of law education and her good judgement skills to pass a judgement in the court. The variables that a judge analyses before giving out a judgement could be modelled in a prediction machine too.

If we place a prediction machine next to her for several hearings and feed the constitution and several law textbooks (simplistically put) in it, we could expect it to have learned something about the nuances of court hearings and decision-making. It would have formed some statistical notions of its own that we didn’t program explicitly and could simply pass the judgements having least statistical error.

Sure, algorithms aren’t so sophisticated yet that they could replace a human judge in a court. Training one such algorithm to the desired accuracy could take a lot of time and a ton of data. However, the possibility of this happening is certainly not out of the realm.

In fact, it is very likely that it will happen. Firstly, people would be assured that there are no biases involved in the judgement. Secondly, the machine can’t be manipulated if the software is made open-access. Thirdly, judicial process would take so much less time than it takes now. In short, we could have a transparent, incorruptible, and speedy justice system with AI.

However, people believe that judges are able to see into the soul of a person and have the innate ability to account for human factors. But it is observed in studies that it actually adds noise to the outcomes and that intuitions could be misleading when it comes to law. On the other hand, a machine version of the judge that mimics the judge is able to outperform her by a large margin since it can do away with such noise. And since it is only an algorithm, we can program it to accommodate racial mixes and other such equitable considerations in its decision-making process.

We don’t see it happening anytime soon though, but we certainly can’t neglect its possibility in future.

Winds of Change

We are in the midst of a paradigm shift. AI is making our lives easier and better. Its penetration in healthcare, transportation, law, and defence will benefit society immensely. But it also has the potential of making us miserable.

It is said that AI can escalate inequality. People with the know-how of this technology—or in other words, complements—would earn significantly more than the others. With most of the jobs gone, people would compete against each other for the remaining non-AI jobs which would drive down wages. Enterprises owning data would become disproportionately wealthy and eerily influence democracy by becoming lobbyists unimaginably powerful.

How societies would react to these changes is uncertain. What we can truly be sure of is the fact that everything in economics is about people. If something doesn’t add up, it shows up in the market through demand and supply. If market failures still prevail, we show dissent and policies change.

But in the end, economy runs for people and not for machines.

-Yashdeep Singh Dahiya