Originally published at dynomantle.com
There has been a lot of discussion over the years about the number of jobs Artificial Intelligence (AI) may "take away". Some have already lost their jobs to robots, which compounds the fear.
While the notion of losing your job to AI may be scary, it is important to consider what "AI" is. It is not the same as human intelligence. It tends to be great at things humans find hard, but find itself unable to accomplish the simplest things that children can.
This guide will talk a bit about what AI is, how similar it is to other automation events in history, and what the possible affects on jobs may be.
When most people talk about AI, they are really talking about machine learning. Let's use a recommendation algorithm as an example. We have a generic e-commerce site that sells all sorts of stuff.
The most straightforward way to provide recommendations is to just look at stuff that sold the most. Something that has 1000 sales will probably be a better recommendation than something that has 10 sales.
However, we will need to use something else if our site sells everything from toilet paper to cameras. Another factor could be how often a person is viewing cameras on the site because that probably means they're researching cameras to buy. The more they view cameras, the more likely they will purchase a camera if we show them a top selling camera.
We can combine these with other factors such as price, similar purchases with other people, or region in the world to create a score of how likely a person is to make a purchase if we recommend a particular camera.
Seems pretty smart right? Here is where we start to see some flaws though. Any human can look at a person buying a $1000 prosumer camera and know that they may be interested in buying accessories. We humans also know that they are unlikely to buy another $1000 camera. We don't even need to look at any purchasing data to make this determination. We know this because we can reason that
- a $1000 purchase is a big purchase
- not many people buy multiple cameras in a short period of time
- someone buying a $1000 camera instead of using their smartphone is taking photography as a serious hobby if not a profession. They will probably need accessories such as different lenses or tripods.
Machine learning algorithms do not reason. Algorithms do not understand things the way we do. They make "decisions" solely on correlation. To have an algorithm stop recommending cameras to people who just purchased a camera takes thousands if not millions of data points.
To use another example, a child can determine that a round peg doesn't fit into a square hole pretty easily. An algorithm will need thousands of data points. Not only that, but you would have to direct the algorithm to know that fitting pegs into appropriately shaped holes is the goal. A child can create their own goals, such as stacking all the objects and considering that a win.
NOTE: machine learning algorithms are not the only form of AI. Artificial General Intelligence (AGI) would actually be more akin to human intelligence. It is the form of AI that you will most often see in movies and tv shows. We also happen to be very far off from it. Trying to predict when computer scientists will make a breakthrough in AGI is as difficult as trying to predict when we'll have a breakthrough in curing cancer for good. It could be next year (unlikely) or it could be hundreds of years from now.
Artificial intelligence is very good at things humans would consider difficult. It can perform millions of mathematical calculations in less time for a human to do one of those calculations. This trait has a number of uses from getting useful search results to image recognition. But let's look at image recognition closer to see where AI starts to hit some limitations.
Image recognition works by looking at patterns of pixels and matching them with known objects. You can probably throw this image of a dog into any image recognition algorithm and it will come up with "dog". The algorithm may even be able to tell the breed of the dog, the fact that the dog is still a puppy, and that the puppy is chewing a stick.
However, while a child can understand what a dog looks like after a few examples, an algorithm will need thousands of photos to identify a dog. It will also need thousands more to identify breeds or age.
An algorithm is also less likely to be able to tell you that the stick was most likely brought in from the street rather than purchased from the store. Or that some poor human had to clean up little wooden shards from all over the apartment. Or that that same poor human eventually got fed up and threw out that stick. Maybe that human stubbed their foot on one of those wooden shards. The puppy was also spoiled for being allowed to bring in random things inside.
These are things humans could have easily determined from looking at this picture. It would be incredibly difficult to generate these same insights from a machine learning algorithm. Things that come naturally to humans are incredibly difficult for machines.
This is because algorithms don't actually understand the concepts in the photo. It can only identify the objects. It knows that certain patterns of pixels means a dog is in the image, but it doesn't know what a dog is. It doesn't understand how dogs behave or how humans behave, which is necessary to understand that sticks aren't bought in stores.
Pattern matching with pixels also has a lot of limitations. There are numerous instances of people manipulating images to trick algorithms.
Look at the image below. The left image will be identified as a school bus. The center is a "mask" added to the image. The right image is obviously still a school bus to any human. Many image recognition algorithms will identify it as an ostrich.
Algorithms make this mistake because they don't understand what a school bus or an ostrich are. They simply look at correlations in the pixels. Those pixels can be modified to fool an algorithm even though any human looking the image would still have the correct answer.
That being said, algorithms have some clear advantages. An algorithm can identify objects in millions of photos in seconds or minutes (depending on how much processing power you throw at it). A human will take who knows how long to go through millions of photos. Years maybe?
Algorithms are not better or worse than humans. They simply have different properties that make them better suited for certain tasks while incapable of doing others. Machine learning is similar to other forms of automation throughout history. Machines in general can perform specific tasks better and quicker than humans can, but their abilities are limited to specific tasks. Even for those specific tasks, the machine can benefit greatly from human intervention.
Deep Blue is famous for beating Garry Kasparov at chess. Kasparov has mentioned in his book Deep Thinking that today, an algorithm will almost always win at chess against a human. However, an algorithm will almost always LOSE at chess to another algorithm paired with a human. A machine can run through the permutations of different moves on a chess board to see what the board would look like many moves ahead. A human actually understands the game at a high level and can create overarching strategies. The abilities are complimentary.
Machines don't have critical thinking abilities. They only have brute force. With algorithms, this comes in the form of millions of mathematical calculations within seconds. It is still brute force nonetheless. Today's algorithms are not fundamentally much different than any other machine. This means that we can look at other forms of automation in history to understand what the impact machine learning will have on jobs.
Automation certainly has an effect on jobs. In the early 1800s, over half the U.S. population worked on farms. Less than 2% of the U.S. population does so today, and yet we have significantly more food available. Tractors, harvesters, sprinkler systems, and such make it possible for a single person to do what used to take dozens if not hundreds of people.
A single person being 100x more productive usually means there are going to be fewer people needed to do that job. Yet, that does not mean the overall effect is fewer jobs. New jobs are created to build and maintain those machines.
Machines also make new jobs possible. SEO (search engine optimization) experts did not exist in the 1990s. That job only exists because of the importance of search engines, which happen to be algorithms.
Boats, planes, trains, and cars are machines that enable the transport of people and goods over long distances. This created jobs in not only building and maintaining those machines, but also in infrastructure such as roads and airports.
The proliferation of travel across different countries has also increased the need for an industry to teach new languages. As much as algorithms have improved in language translation over the years, it is unlikely to replace the need for humans performing translation.
None of these jobs were imagined before the automation came into place. That can make things seem scary because looking forward, you can only see the jobs going away. You can not see the jobs that don't exist yet. There is just no way to make something that doesn't exist feel real. And while new jobs will be created over the long term, there will be severe short term effects with any new form of automation. Elevator operators and toll booth workers are obsolete, but those workers can't exactly get jobs fixing cars or building web pages without a significant amount of training.
The main difference between machine learning algorithms and previous forms of automation is that other forms of automation have replaced repetitive physical tasks. Algorithms are replacing repetitive mental tasks. This new trait is a little scary because it is unfamiliar, but let's look at an example to see how some jobs may change in the future rather than disappear.
One of the more commonly talked about jobs being "lost" to AI is truck driving. It makes sense in many ways. Most of the time spent driving a truck is going straight down an interstate. With over 3 million truck drivers in the U.S., the effect of this job loss is immense.
However, there are a number of factors to consider here where the net job effect of automation in truck driving could be positive instead of negative.
The first is that having self driving vehicles be available in all cases could be farther than most people think. Remember that machine learning algorithms don't actually understand things. They find correlations in data and make decisions based on those correlations. A unique or novel situation would not have much data available for it and the algorithm will not be able to make a good decision in that case.
Human drivers understand what a stop sign is. You can cover up part of the stop sign and know that it is still a stop sign. The same is not true for algorithms.
Researchers have put stickers on stop signs to trick algorithms into thinking they are something else. One could argue that the algorithm builders can simply account for that case, but that would not address the underlying problem. Humans understand concepts and can handle situations they have never encountered before. How can an algorithm builder train an algorithm to handle a situation they have not encountered before?
That rules out replacing truck drivers for the "last mile" driving within cities and local roads. However, we can still replace them on interstates... right?
That leads us to two more points. The first is that automation tends to make goods and services cheaper. Machines have a large fixed cost, but they don't need a salary or benefits. They can also work 24 hours a day without a break. But what happens when goods and services become cheaper? What happens if shipping goes from $6 to $3? Or what happens if a book costs $10 instead of $15?
Oftentimes, cheaper goods means that people purchase more of those goods. They can afford to. That increases the number of goods that get transported, which increases the number of trucks being used. While the bulk of that driving is done on highways however, it also means it increases the amount of goods being transported on local roads, which happen to be driven by human truck drivers. So there is an offset to the "job loss" there.
The other point is that machines don't run forever. They need maintenance. And with a machine driving the truck 24 hours a day, they will probably need more maintenance than usual. Many truck drivers are already capable of performing some maintenance on their trucks so it is not far fetched to say they could transition to performing more maintenance.
And having a nationwide network of maintenance workers could be expensive for a single trucking company. It could make sense to have a human truck "driver" stay on one truck in a 5 truck convoy in the event any one of those 5 trucks breaks down.
We should also mention that an increase in purchased goods being transported also means an increase in the creation of those goods. While that too is mostly automated, there are undoubtedly going to be plenty of humans in the process as well.
It is also very likely that trucking will turn out completely differently from what we just stated. It is difficult to see how technology and automation will change the job market. What we can see from history is that it does tend to create plenty of jobs that we can't even imagine.
We hope this guide has helped you understand artificial intelligence a little bit better. Knowing the capabilities and limitations of AI can help use understand what jobs are at risk, but also which jobs are more likely to change rather than go away.
If you are interested in reading a discussion of jobs other than trucking, sign up for the Dynomantle app!
If you are interested in learning more about the capabilities of artificial intelligence, here are some other great resources.
Garry Kasparov is a former world chess champion. The first time a machine has beaten a world chess champion was when IBM's Deep Blue went up against Kasparov.
This book is an excellent analysis of that match, the history of AI chess machines, and the differences between AI and human thinking.
Artificial Intelligence: A Guide for Thinking Humans by Melanie Mitchell
Melanie Mitchell is a computer scientist and AI researcher. She's written a number of books an AI.
This particular book goes pretty deep into how AI works and really digs into what AI is good at versus what it is bad it. Some of the explanations are hard to follow if you do not have a computer science or mathematics background, but you can still get a lot out of the book if you ignore those parts.
Fooling the machine by Dave Gershgorn
A fantastic article that describes numerous cases where image recognition algorithms can be fooled.