Elements of AI

Now that we have a better understanding of the basic concepts of AI, we are in a much better position to take part in rational discussion about the implications of already the current AI.

Implication 1: Algorithmic bias

AI, and in particular, machine learning, is being used to make important decisions in many sectors. This brings up the concept of algorithmic bias. What it means is the embedding of a tendency to discriminate according ethnicity, gender, or other factors when making decisions about job applications, bank loans, and so on.

Algorithmic bias isn’t a hypothetical threat conceived by academic researchers. It’s a real phenomenon that is already affecting people today.

Online advertising

It has been noticed that online advertisers like Google tend to display ads of lower-pay jobs to women users compared to men. Likewise, doing a search with a name that sounds African American may produce an ad for a tool for accessing criminal records, which is less likely to happen otherwise.

Social networks

Since social networks are basing their content recommendations essentially on other users’ clicks, they can easily lead to magnifying existing biases even if they are very minor to start with. For example, it was observed that when searching for professionals with female first names, LinkedIn would ask the user whether they actually meant a similar male name: searching for Andrea would result in the system asking “did you mean Andrew”? If people occasionally click Andrew’s profile, perhaps just out of curiosity, the system will boost Andrew even more in subsequent searches.

There are numerous other examples we could mention, and you have probably seen news stories about them. The main difficulty in the use of AI and machine learning instead of rule-based systems is their lack of transparency. Partially this is a consequence of the algorithms and the data being trade secrets that the companies are unlikely to open up for public scrutiny. And even if they did this, it may often be hard to identify the part of the algorithm or the elements of the data that lead to discriminating decisions.

The last point means, in other words, that companies such as Facebook and Google, at least when providing services to European users, must explain their algorithmic decision making processes. It is, however, still unclear what exactly counts as an explanation. Does for example a decision reached by using the nearest neighbor classifier (Chapter 4) count as an explainable decision, or would the coefficients of a logistic regression classifier be better? How about deep neural networks that easily involve millions of parameters trained using terabytes of data? The discussion about the technical implementation about the explainability of decisions based on machine learning is currently intensive. In any case, the GDPR has potential to improve the transparency of AI technologies.

Implication 2: Seeing is believing — or is it?

We are used to believing what we see. When we see a leader on the TV stating that their country will engage in a trade-war with another country, or when a well-known company spokesperson announces an important business decision, we tend to trust them better than just reading about the statement second-hand from the news written by someone else.

Similarly, when we see photo evidence from a crime scene or from a demonstration of a new tech gadget, we put more weight on the evidence than on written report explaining how things look.

Of course, we are aware of the possibility of fabricating fake evidence. People can be put in places they never visited, with people they never met, by photoshopping. It is also possible to change the way things look by simply adjusting lighting or pulling one’s stomach in in cheap before–after shots advertising the latest diet pill.

Implication 3: Changing notions of privacy

It has been long known that technology companies collect a lot of information about their users. Earlier it was mainly grocery stores and other retailers that collected buying data by giving their customers loyalty cards that enable the store to associate purchases to individual customers.

However, as such the above kind of data logging is not yet AI. The use of AI leads new kinds of threats to our privacy, which may be harder to avoid even if you are careful about revealing your identity.

Using data analysis to identify individuals

A good example of a hard-to-avoid issue is de-anonymization, breaking the anonymity of data that we may have thought to be safe. The basic problem is that when we report the results of an analysis, the results may be so specific that they make it possible to learn something about individual users whose data is included in the analysis. A classic example is asking for the average salary of people born in the given year and having a specific zip code. In many cases, this could be a very small group of people, often only one person, so you’d be potentially giving data about a single person’s salary.

An interesting example of a more subtle issue was pointed out by researchers at the University of Texas at Austin. They studied a public dataset made available by Netflix containing 10 million movie ratings by some 500,000 anonymous users, and showed that many of the Netflix users can actually be linked to user accounts on the Internet Movie Database because they had rated several movies on both applications. Thus the researchers were able to de-anonymize the Netflix data. While you may not think it’s big deal whether someone else knows how you rated the latest Star Wars movie, some movies may reveal aspects of our lives (such as politics or sexuality) which we should be entitled to keep private.

Other methods of identification

A similar approach could in principle be used to match user accounts in almost any service that collects detailed data about user behaviors. Another example is typing patterns. Researchers at the University of Helsinki have demonstrated that users can be identified based on their typing patterns: the short intervals between specific keystrokes when typing text. This can mean that if someone has access to data on your typing pattern (maybe you have used their website and registered by entering your name), they can identify you the next time you use their service even if you’d refuse to identify yourself explicitly. They can also sell this information to whoever wants to buy it.

While many of the above examples have come as at least in part as surprises – otherwise they could have been avoided – there is a lot of ongoing research trying to address them. In particular, an area called differential privacy aims to develop machine learning algorithms that can guarantee that the results are sufficiently coarse to prevent reverse engineering specific data points that went into them.

Implication 4: Changing work

When an early human learned to use a sharp rock to crack open bones of dead animals to access a new source of nutrition, time and energy was released for other purposes such as fighting, finding a mate, and making more inventions. The invention of the steam engine in the 1700s tapped into an easily portable form of machine power that greatly improved the efficiency of factories as well as ships and trains. Automation has always been a path to efficiency: getting more with less. Especially since the mid 20th century, technological development has led to a period of unprecedented progress in automation. AI is a continuation of this progress.

Each step towards better automation changes the working life. With a sharp rock, there was less need for hunting and gathering food; with the steam engine, there was less need for horses and horsemen; with the computer, there is less need for typists, manual accounting, and many other data processing (and apparently more need for watching cat videos). With AI and robotics, there is even less need for many kinds of dull, repetitive work.

Because we can’t predict the future of AI, predicting the rate and extent of this development is extremely hard. There have been some estimates about the extent of job automation, ranging up to 47% of US jobs being at risk reported by researchers at the University of Oxford. The exact numbers such as these – 47%, not 45% or 49% –, the complicated-sounding study designs used to get them, and the top universities that report them tend to make the estimates sound very reliable and precise (recall the point about estimating life expectancy using a linear model based on a limited amount of data). The illusion of accuracy to one percentage is a fallacy. The above number, for example, is based on looking at a large number of job descriptions – perhaps licking the tip of your finger and putting it up to feel the wind – and using subjective grounds to decide which tasks are likely to be automated. It is understandable that people don’t take the trouble to read a 79 page report that includes statements such as "the task model assumes for tractability an aggregate, constant-returns to-scale, Cobb-Douglas production function." However, if you don’t, then you should remain somewhat sceptical about the conclusions too. The real value in this kind of analysis is that it suggests which kinds of jobs are more likely to be at risk, not in the actual numbers such as 47%. The tragedy is that the headlines reporting "nearly half of US jobs at risk of computerization" are noted, and the rest is not.

So then, what actually are the tasks that are more likely to be automated? There are some clear signs concerning this that we can already observe:

• Autonomous robotics solutions such as self-driving vehicles, including cars, drones, boats or delivery robots, are just at the verge of major commercial applications. The safety of autonomous cars is hard to estimate, but the statistics suggests that it is probably not yet quite at the required level (the level of an average human driver). However, the progress has been incredibly fast and it is accelerating due to the increasing amount of available data.
  
• Customer-service applications such as helpdesks can be automated in a very cost-effective fashion. Currently the quality of service is not always to be cheered, the bottle-necks being language processing (the system not being able to recognize spoken language or to parse the grammar) and the logic and reasoning required to provide the actual service. However, working applications in constrained domains (such as making restaurant or haircut reservations) sprout up constantly.

For one thing, it is hard to tell how soon we’ll have safe and reliable self-driving cars and other solutions that can replace human work. In addition to this, we mustn’t forget that a truck or taxi driver doesn’t only turn a wheel: they are also responsible for making sure the vehicle operates correctly, they handle the goods and negotiate with customers, they guarantee the safety of their cargo and passengers, and take care of a multitude of other tasks that may be much harder to automate than the actual driving.

As with earlier technological advances, there will also be new work that is created because of AI. It is likely that in the future, a larger fraction of the workforce will focus on research and development, and tasks that require creativity and human-to-human interaction. If you’d like to read more on this topic, see for example Abhinav Suri’s nice essay on Artificial Intelligence and the Rise of Economic Inequality.