Mind Matters News and Analysis on Natural and Artificial Intelligence

“Artificial” Artificial Intelligence

What happens when AI needs a human I?
Share
Facebook
Twitter
googleplus Google+
arroba Email

Most of us use artificial intelligence (AI) without really realizing it. We use Google Translate to read webpages written in other languages and Facebook to identify faces in a photo, and to have Siri nag you about turning around because you just missed your turn for the third time.

However, during the rapid growth of AI, there has been another, related industry that has been growing as well, one that has received much less fanfare. This is the area known as “Artificial Artificial Intelligence,” or AAI for short. Yes, “artificial” is used twice on purpose.

To understand what AAI does, let’s imagine that we have a folder of thousands of pictures we have taken over the years. We want to write a computer program that sorts each of our images into one of four categories—people, animals, plants, and buildings. That sounds simple enough. We can imagine a program that looks something like this:

Free Vector Sleeping Cat
Illustration credit: Vecteezy.com

1. Open up the folder containing the images

2. For each image in the folder

A. Look at the image

B. Decide if the primary subject of the image is a person, animal, plant, or building

C. Put the image in the right destination folder based on the primary subject

This is all straightforward, except for Step 2B. This actually involves two subtasks, each of which is very difficult for a computer:

1. Identify the primary subject

2. Determine if that subject is a person, animal, plant, or building

While artificial intelligence tools are getting better and better at each of these operations, for now, no AI system comes close to being able to do either of these two tasks well. Even to the extent that it is possible, it is certainly not easy to write programs for them.

Peter Thiel

However, humans are actually very good at both tasks. As PayPal co-founder Peter Thiel noted in Zero to One,

In 2012, one of [Google’s] supercomputers made headlines when, after scanning 10 million thumbnails of YouTube videos, it learned to identify a cat with 75% accuracy. That seems impressive – until you remember that an average four-year-old can do it flawlessly.

So what can a computer programmer do? Let’s start with the fact that every other task is easy enough to program. What if, on Step 2B, instead of trying to apply artificial intelligence to the task, we applied an “artificial” artificial intelligence instead? In other words, we replace the AI with an actual human? Therefore the new workflow would look like this:

1. Open up the folder containing the images

2. For each image in the folder

A. Look at the image

B. Ask a real person (from a pool of participants) if the primary subject of the image is a person, animal, plant, or building

C. Put the image in the right destination folder based on the primary subject

In this revised situation, artificial intelligence Step 2B is replaced by asking a real person.

AAI requires a reliable pool of humans as well as effective ways to distribute tasks to them and gather up the results. AAI won’t work if, at Step 2B, the program freezes because no one is available to do the task or there isn’t a way to get the data from that person back into the program.

There are many systems available today to facilitate these types of interactions. The best known is Amazon’s Mechanical Turk. In addition to its dominant retail business, Amazon sells many cloud-based computer services to other software developers.

The Mechanical Turk is named after a “mechanical chess” machine that startled many in the 1770s.

A1980s A 1980s reconstruction of Kempelen’s chess-playing “automaton”/Carafe (CC BY-SA 3.0)

Inventor Wolfgang von Kempelen (1734–1804) claimed to have invented a machine that would play chess for Habsburg Archduchess Maria Theresa. “Automatons” (machines that played the flute, etc.) were a common source of entertainment in pre-industrial Europe. But the Turk was something else:

The Turk was good. Really good. And it wasn’t just adept at executing a repetitive task. The Turk responded skillfully to the unpredictable behavior of humans. This machine seemed to be operating autonomously, guided by its own sense of rationality and reason. If the human opponent attempted to cheat, as Napoleon did when facing off against the machine in 1809, the Turk would move the chess piece back to its previous position, and, after repeated cheating attempts, would swipe his arm across the board, scattering pieces to the ground.

Although some suspected a hoax, for others, the Turk raised many of the questions we address today about the nature and limits of artificial intelligence. But back then, a simpler truth emerged:

… the Turk did operate via a concealed operator, who controlled each movement from inside the cabinet by candlelight, pulling levers to operate the Turk’s arm and keeping track of the moves on their own board. Von Kempelen, and his Turk-touring successor, Johann Maelzel, picked up new chess players on their travels, gave them a quick how-to orientation, then bundled them into the cabinet.

Ella Morton, “The Mechanical Chess Player That Unsettled the World” at Slate

And that is somewhat like Amazon’s practices today. Numerous systems that have been built with Mechanical Turk. Some examples include:

  •  Apps which scans a business card and turns it into a phone contact
  • Several software transcription systems
  • Apps which will rewrite paragraphs to make them shorter, livelier, etc.
  • Software which automatically tags photos when uploaded

All of these software systems look to their users like advanced artificial intelligence. Actually, they hide a lot of human effort behind a technological mask.

The ultimate goal for any product developer is to find the best split between humans and computers for any task. Using humans for tasks that can be solved by computers is wasteful, inefficient, and error-prone. However, using computers for tasks that require holistic thinking eats up money without creating value. Artificial Artificial Intelligence (AAI) can help the developer split that gap and thus provide “automatic” services to users when AI is either incapable of the task or too expensive to implement.

In short, many apparently automated services that have a considerable “wow” factor are actually outsourcing pieces of the puzzle to humans who do the creative work.

Note: Some have noted that Mechanical Turk and similar systems can lead to abusive labor conditions. For example, Alana Semuels tells us at The Atlantic, “A research paper published in December that analyzed 3.8 million tasks on Mechanical Turk, performed by 2,676 workers, found that those workers earned a median hourly wage of about $2 an hour. Only 4 percent of workers earned more than $7.25 an hour” (“The Internet Is Enabling a New Kind of Poorly Paid Hell,” January 23, 2018). Because each task is individually contracted, employers need not pay minimum wage or observe any labor regulations. In fact, employers know nothing about the workers they employ. They may be inadvertently using captive workers who are literally chained to a desk to perform tasks. That is a problem of current AI implementation, to be sure, but it is not an inevitable outcome of introducing AI technology.

Jonathan Bartlett is the Research and Education Director of the Blyth Institute.

See also: Sometimes the ‘bots turn out to be humans (Denyse O’Leary)

and

Is Bitcoin safe?: Why the human side of security is critical (Jonathan Bartlett)