(These are excerpts from my book "Intelligence is not Artificial")
Intelligent Behavior in Structured Environments
When you need to catch a bus in some underdeveloped countries, you don't know at what time it will arrive nor how much you will be charged for the ticket. In fact you don't even know how it will look like (it could be a generic truck or a minivan) and where it will stop. Once on board, you tell the driver where you want to get off and hope that he will remember. If she is in a good mood, she might even take a little detour to drop you right in front of your hotel. On the other hand, when you take a bus in a developed country, there is an official bus stop (the bus won't stop if you are 20 meters before or after it), the bus is clearly recognizable and marked with the destination and won't take any detour for any reason, the driver is not allowed to chat with the passengers (sometimes she is physically enclosed in a glass cage), the ticket must be bought with exact change at a ticket vending machine (and sometimes validated inside at another machine). There is a door to be used to exit, and you know when to exit because the name of the bus stop is displayed on a LED screen. On many long-distance trains and buses you also get an assigned seat (you can't just sit anywhere).
It is easy to build a robot that can ride a bus in a developed country, much more difficult to build a robot that can ride a bus in an underdeveloped country. What makes it easy or difficult is the environment in which it has to operate: the more structured the environment, the easier for the robot. A structured environment requires less "thinking": just follow the rules and you'll make it. However, what really "makes it" is not you: it's you plus the structured environment. That's the key difference: operating in a chaotic, unpredictable situation is not the same thing as operating in a highly structured environment. The environment makes a huge difference. It is easy to build a machine that has to operate in a highly structured environment, just like it is easy for a bullet train to ride at 300 km/hour on rails.
We structure the chaos of nature because it makes it easier to survive and thrive in it. Humans have been spectacularly successful at structuring their environment so that it obeys simple, predictable rules. This way we don't need to "think" too much: the structured environment will take us where we want to go. We know that we can find food at the supermarket and a train at the train station. In other words, the environment makes us a little more stupid but allows anybody to achieve tasks that would otherwise be difficult and dangerous, i.e. that would require a lot of intelligence. When the system fails us, we get upset because now we have to think, we have to find a solution to an unstructured problem.
If you are in Paris and the metro is on strike and it is impossible to get a taxi, how to do you get to your appointment in time? Believe it or not, most Parisians manage. Most tourists from the USA don't. If there is no traffic light and cars don't stop for pedestrians and traffic is absolutely horrible, how do you cross a wide boulevard? Believe it or not, Iranians do it all the time. Needless to say, most Western tourists spend hours trying to figure it out.
It is certainly very impressive how well humans structure a universe that is chaotic. The more we structure it, the easier for extremely dumb people and machines to survive and thrive in it.
The claims of the robotic industry are often related to structured environments, not to their robots. It is relatively easy to build an autonomous car that rides on a highway with clearly marked lanes, clearly marked exits, ordered traffic, and maps that detail everything that is going to happen. It is much more difficult (orders of magnitude more difficult) to build an autonomous car that can drive through Tehran or Lagos (this is a compliment to Iranian and Nigerian drivers, not an insult). Whoever claims that a computer is driving a car is distorting the facts: it is not the computer that is driving the car but the environment that has been structured so that any inexperienced and not particularly intelligent driver, and even a computer, can drive a car. Today's computer cannot drive a car in the traffic of Lagos or Tehran. It will if and when the streets of Lagos and Tehran become as well structured as the streets of California, if and when Iranian and Nigerian drivers are forced to obey strict traffic rules. Saying that the on-board computer is steering the driverless car is like saying that the locomotive knows in which direction to take the train: the locomotive is simply constrained by the rails to take the correct direction.
In order for self-driving cars to use our streets, we will need to retrofit roads with devices that tell the car what to do at every point in time. It is not intelligence but old-fashioned infrastructure that will allow very dumb self-driving cars to drive safely; in other words we will need the equivalent of the highly-structured system of rails and controllers that allow the fast, safe and accurate travel of trains.
There is sometimes a bit of confusion about the expressions "driver-less" and "self-driving": they are not synonyms. A driver-less vehicle is not necessarily a self-driving vehicle. Today many trains, airliners and factory machines operate mostly without the intervention of a human being: the environment has been structured so that the machine can do its job, safely and efficiently. The extraordinary has been removed. The designers of the environment for that device made sure that only the ordinary can happen, and the ordinary is greatly simplified. The driver-less vehicle in the structured environment is NOT doing what humans do (used to do) in their unstructured environments: humans had to deal with a chaotic system in which the extraordinary was routine, where at every moment they had to make decisions. The extraordinary was ordinary. "Self-driving", instead, implies that the machine does what a human can do: make decisions in extraordinary circumstances, anytime, anywhere.
I recently had to exchange the equivalent of $3.00 in a local currency while leaving a Western country at its capital's airport. The procedure was silly beyond belief. I had to produce passport, boarding pass and receipt of previous money exchanges before getting my money, a lengthy operation for just three dollars. On the contrary at the border between Haiti and Dominican Republic, a wildly chaotic place with taxi drivers, fruit vendors and police officers yelling at each other and at everybody passing by, there was a mob of money changers chasing the travelers. I had to guess which ones were honest money changers rather than scammers, and then bargain the exchange rate, and then make sure that the money was good while all the time protecting my wallet from pickpockets. It wouldn't be difficult to build a robot that can exchange money at the airport of a Western capital, but much more difficult (orders of magnitude more difficult) to build one that can exchange money while walking from the immigration post of Haiti to the immigration post of the Dominican Republic.
The more structured the environment, the easier to build a machine that operates in it. What really "does it" is not the machine: it's the structured environment. What has made so many machines possible is not a better A.I. technology, but simply better structured environments. It's the rules and regulations that allow the machine to operate.
You can't call an automatic phone system and just explain your problem. You have to press 1 for English, 1 for customer support, 3 for your location, 2 for your kind of problem and 4 and 7 and so forth. What allows the machine to perform its job, and to replace the human operator, is that you (the human being) have removed the human aspect from the interaction and behave like a machine in a mechanical world. It is not the machine that behaves like a human being in a human world.
The fundamental thing that a self-driving car must be able to do is, of course, to stop at a gas station when it runs out of gasoline. Can these cars autonomously enter a gas station, stop in front of a pump, slide a credit card in the payment slot, pull out the hose and pour gasoline in the tank? Of course, not. What needs to be done is to create the appropriate structured environment for the driverless car (or, better, for some sensors on board the car) so that the car will NOT need to behave like an intelligent being. The gas station, the gas pump and the payment used by the driverless car will look very different from the one used so far by human drivers.
Incidentally, most of those rules and regulations that create a highly structured environment (favorable to automata) were originally introduced in order to reduce costs. Employing machines has been the next logical step in cost reduction. The machine is one step in an ongoing process of cost reduction and productivity increase. The goal was not to create superhuman intelligence, just to increase profits.
Think of your favorite sandwich chain. You know exactly what kind of questions they will ask you. There is a well-structured process by which your sandwich will be made. The moment robots become cheap enough, they will certainly take over the jobs of the kids who prepare your sandwich today. It is not a matter of "intelligence" (the intelligence of today's robots is already more than enough) but of cost: today a teenager is cheaper than a robot. The whole point of structuring the sandwich-making process was to allow inexperienced and unskilled workers (read: underpaid) to perform the task once reserved to skilled experienced chefs.
The more unstructured the environment is, the more unlikely that a machine can replace the human. Unfortunately, one very environment that is still very unstructured is health care. Medical records are kept on physical files, and doctor's notes are notoriously impossible to read. There is very little that a machine can do in that environment. The way to introduce "intelligent" machines in that environment is, first of all, to structure all that information. When it is "digitized" and stored in databases, it means that it has been structured. At that point any human being, even with little or no knowledge of medical practice, can do something intelligent in that environment. And even a machine can.
The truth is that we do not automate jobs as they are. First, we dehumanize the job, turning it into a mechanical sequence of steps. Then we use a machine to automate what is left of that job. For example, my friend Steve Kaufman, a pediatrician all his life, realized that his skills were less and less necessary: a nurse practitioner can fill all the forms and click on all the computer buttons that are required when seeing a patient; the doctor, who is increasingly required to type on a keyboard, may not even make eye contact with the patient. This has the beneficial effect of reducing the number of days that the average patient spends at a hospital, but it erases the kind of bonding between doctor and patience that was common in the "unstructured" world. When the last vestiges of humanity will have been removed from the job of the doctor, it will be relatively easy to automate the doctor's job. But that is not what Steve was doing. As Steve pointed out to me, if you don't bond with an asthmatic patient, you may never realize that he is suicidal: you will cure his asthma, but he will commit suicide; and the machine will still archive the case as a success.
Structured environments are also relying on ever stricter rules. My favorite example is the boarding procedure at an airport, where we are treated like cattle from check-in to the gate, with a brief interval during which we are treated like a walking credit card that airport shops desperately try to get. Other than the credit card thing, we are basically building the kind of hyper-bureaucratic state pioneered by Soviet Union.
There is a fundamental paradox underlying the ongoing structuring of society. What is profoundly human (and actually shared by all forms of life) is the vagueness of language and behavior. What humans (and animals) can do relatively well, and do on a daily basis, and today's machines are not good at, is to deal with ambiguity. Unfortunately, ambiguity is responsible for a lot of the miscommunication and chaos that complicate our life. Rules and regulations are useful because they remove ambiguity from society, and therefore simplify our life. As a side-effect, though, the more we structure human behavior by removing ambiguity, the more replicable it becomes. We become machines; machines that demand a high salary and all sorts of rights. It is a no-brainer for businesses to replace such expensive machines with cheaper ones that don't demand any right.
Increasingly structured environments, routines and practices will eventually enable the automation of "cognitive" skills too. I am writing while watching the indecent spectacle of the political campaigns for a presidential election in the USA. Political debates are becoming more and more structured, with a format agreed beforehand and a moderator that enforces it, and a restriction on the kind of questions that can be asked, and candidates who basically memorize press releases worded by their campaign staff. It is not difficult to imagine that sooner or later someone will build a piece of software that can credibly replace a politician in a political debate; but that feat will owe more to the lack of real debate in these political debates than to greater rhetorical skills on the part of the machine. On the other hand that software will be incapable of participating in a passionate conversation about a World Cup game with a group of rowdy and drunk soccer
fans.
It is the increasingly structured environment that is enabling and will enable the explosion of robotics and automated services. Most of the robots and
phone-based services coming to the market now rely on relatively old
technology. What has made them feasible and practical is that they can now
operate in highly structured environments.
Think of yourself. You are now identified by numbers in so many different contexts: your passport number, your social security number, your street address, your
telephone number, your insurance policy number, your bank account number, your
credit card number, your driver license number, your car's plate number, your
utility bill account number It is a rarity when someone tries to identify me
based on non-numeric features. And increasingly we depend on passwords to
access our own information. The more we reduce the individual to a digital
file, the easier it gets to build "intelligent assistants" for that file
sorry, i meant "for that person".
In a sense, humans are trying to build machines that think like humans while machines are already building humans who think like machines.
Back to the Table of Contents
Purchase "Intelligence is not Artificial"
|