As accomplished as modern-day computers are, there are some very basic things even the smartest machines have yet to master: tough judgment calls, advanced image recognition, making goofy faces, conducting psychological surveys. These are an assortment of tasks we humans can still claim as our own. Or at least that we can outsource to other, less fortunate humans. Like me.
In Amazon’s words, Mechanical Turk is “a marketplace for work that requires human intelligence”. But in reality it’s even simpler than that description implies: It’s a job board where the pay is low and the jobs are dumb. If you have a functional cerebral cortex, an internet connection, and a few minutes to spare, you can pick up a handful of odd jobs — the oddest of jobs — and make a few bucks, pennies and nickels at a time. But what’s it like to be that “human intelligence?” As I found out last year, it’s weird, fascinating, perplexing and a little depressing, all at once.
I’d always been vaguely aware of Mechanical Turk, and in the summer of last year, I decided to try it on a lark. I don’t remember exactly what spurred me on, but the concept of making a few dollars while sitting slack-jawed in front of a computer was appealing, and with Amazon’s weight behind it the whole process seemed a little more respectable than filling out surveys for gift cards.
But the real draw was curiosity: What kind of jobs are out there for 5 cents or 10 cents? What kind of massive data-driven operations would I get the chance to glimpse one strange facet of?
Weird ones.
So you want to be a Turk
Signing up for the program with your Amazon account is easy; just head to the Mechanical Turk website and click one of the “get started” buttons on the austere, seemingly Dot Com-era interface. Once you’re in, the site will tell you how much you’ve earned and how many tasks you’ve accomplished — displaying a long list of “HITs”, or “Human Intelligence Tasks”.
This is your HIT list.
HITs can be all manner of things, all necessitating uniquely human skills. These are tasks that should appeal to the soul, right? “Feel empathy for 20 cents” and “Contemplate the nature of existence for a nickel.” Not so much. In reality, it’s mostly a lot of surveys, and a lot of cruft. Choose wisely; you’ll always remember your first HIT job.
My first job was to stalk people on social media for somewhere around two cents a pop. No, really. Specifically, the task was to look at user accounts on three or four different social media sites and judge whether or not they’re owned by the same person. Is this that LinkedIn dude’s Twitter account? Is this his half-empty Google+ profile?
The indecision and doubt was bad enough, but the asshat who created this particular HIT made the experience doubly unpleasant by seeding in trick questions — planted test examples with right and wrong answers, just to make sure you aren’t an idiot. It added a layer of suspicion to the whole thing. Is this one where he’s trying to trick me?!
That unpleasant experience turned me off the whole concept for more than a year, until a few months ago, when the siren call of Turk’s lurking fascinating weirdness pulled me back. On that second, more determined pass, I started to uncover Mechanical Turk’s real secrets, and its glory.
The mind-numbing, tedious crap
When you try just one HIT, or even two or three, you’re bound to get something horribly boring. This stuff is Mechanical Turk’s bread and butter: simple, mind-numbingly tedious jobs that use your miracle of an organic brain-computer for lame and obvious tasks. It’s assembly line work, but for the mind instead of the hands.
These are your transcription HITs, your decipher-this-horrible-handwriting HITs, your survey HITs. These are the tasks where translating human ambiguity into computer-style binary is the most frustrating, because you’re likely to arrive at the answer “I don’t know” and be forced to choose anyway. I remember a few angry-keyboard-mashing examples, like the time I was told to identify whether jewellery was gold or silver based on a single black-and-white picture. Impossible to tell! Or being asked if a person’s illegible scribbling looks more like “X” or “Y”. Neither, it looks like “Z!”
The tasks are oddly more frustrating once you’ve figured out their e. Putting together gadgets on an assembly line might be slightly amusing if you can keep yourself occupied by turning them around in your hands and wondering what they’re for. But when the answer is clear and non-mysterious, there’s not much to keep you going other than those nickels and dimes.
The HIT asking you to go to a local pharmacy and take a geotagged picture of the front counter “making very sure to include any tobacco products if there are any tobacco products displayed” (a hefty $2 a pop!) is clearly part of some study on tobacco marketing. The HIT asking you to google brand names and describe the results are clearly part of some SEO scam/research. The HIT asking you to rip information about item prices from 100 Walmart receipts is probably some lazy dude’s attempt to outsource his expense report. They are boring. They are benign. They are not worth your time.
I am being watched
There’s an old Indian proverb where a king has a half dozen blind men touch an elephant. He asks each to tell him what it is they have brushed up against. The one who felt a leg guesses it’s a tree. The one who touches a tail guesses it’s, I dunno, a snake probably? None of them can ever really grasp the truth of what they have really encountered. Which is to say, the joy of Mechanical Turk isn’t just to pick up a little money for even littler jobs. It’s stepping into that role of the blind man for just a few hours at a time, and reveling in the challenge of imagining elephants.
Of the many flavours of “What… why?” I asked myself during my time with Turk, the most common stemmed from the strange, inexplicable HITs that basically just ask you to be a human. And they’re repeated framed in a way that makes it seem like the “person” behind the job doesn’t know what the hell a human is.
Take “Describe a POI Video” for instance. My mission was to watch a three minute-ish video, shot first person from the driver’s seat of a car. The job (which paid $0.10) was not just to perform a function but to actually roleplay as a human being.
I was instructed to imagine that I am the driver and that my passenger is a local who knows every place in town. Then, I am to write out conversationally-worded questions asking him to tell me what some buildings are as you drive by.
Repeat five times, collect your dime.
Was this research for some local tourism board, investigating what buildings on its main drag are interesting enough slap on a pamphlet? Maybe. But why not just come out and ask that? The roleplay bit makes it stranger, like the whole thing is a ruse. Like the job itself is just busywork and someone — or something — is observing you from behind a one-way mirror as you carry it out.
In another roleplay HIT, I was tasked with composing text messages to nonexistent characters about oddly specific events in my imaginary life. I’ve volunteered to organise the “kids party” at school, but now I have to back out. Compose a text explaining way. Or I’m going to the theatre this weekend. Compose a text to my friend explaining what we’ll be seeing and asking if they’d like to tag along.
But the questions sounded distinctly bot-like. It’s almost as if there’s a nascent AI behind this task that has just enough self-awareness to know it needs to study humans, but isn’t quite bright enough to realise when it’s being awkward.
“Today morning” is not proper English.
The questions couldn’t even nail basic grammar: “Today morning” one read, as if to say “show me how it is that you do the texts about the feelings you have! How does a human make its feelwords?”
In one HIT with the awkward title, “Do expressions in front of your webcam,” I was asked to use my human face to imitate the expression of another human face. Why? So computers could learn to recognise emotion. I’m not kidding:
“In this HIT you are asked to do expressions like closing your eyes or opening your mouth. We’ll take snapshots with your webcam so that we can teach our computers how to detect these kind of expressions,” the task description reads.
Suddenly, while I’m sitting there making faces into my webcam to teach machines to recognise hate and fear and disgust and love and surprise, I’m actually face-to-face with another human being. Am I catching a glimpse through the one-way mirror? Maybe, I think, this “fearful” lady isn’t just one of the watched like I am, she’s one of the watchers as well.
My expression is half muted bewilderment, half terror.
I’m the one watching
Looking back on the whole experience, I ran into other people pretty much anywhere I went with Mechanical Turk, starting with my very first HIT cyberstalking social media accounts. Yet somehow it’s easy to tune out the inherent creepiness of it. At least it was for me. Everything was broken up into bite-sized chores and bite-sized payments, and I slipped from one to the next without giving much thought to any of it.
I got my first taste of this when I found an HIT called “Label materials of objects in an image.” In the LMoOiaI HIT, you’re given a picture — usually of a street or someplace outside — with maybe two dozen dots on it. When you hover your curser over the dot it shows a label that indicates what the object is: “street” or “window” or so on. Your job is to click each dot and list what the object is made out of, selecting from a given list of materials.
If it sounds easy but tedious, that’s because it is. Mostly. A window is glass and metal. An awning is metal and fabric. A car is metal, fabric, leather, plastic, glass and rubber. But unlike the mind-numbing “read my messy handwriting” HITs, LMoOiaI gives you some food for thought along with the tedium. Specifically the question, What in the actual hell??!?
I spent an hour or two pondering the point of this task while completing various LMoOiaI HITs. I still have no goddamn idea what it was about. Then, while idly wondering if this was a job dreamed up by a human or a part-way sentient computer, I ran up against an even more pressing question. I found a little dot that said “person”. Well, shit, I thought to myself. What are the materials of a person? Brain matter, muscle tissue, fingernails, aspirations, nostril hair, a soul? Fortunately the options were limited, which kept me from spinning off into an existential crisis. This was the best I could come up with:
I’m still a little conflicted about the “food” part, but I stand by it.
It’s weird enough to try to classify what a person is made out of, choosing from a very limited palette of ingredients. But the challenge also begged a follow-up question: Who wants to know? Surely, somewhere out there were other Turkers looking at pictures and labelling the things in them, blissfully unaware that another human being would soon be describing that “person” they labelled as “food”. But who is orchestrating the plot from its upper levels? There’s never an answer, but the question looms with every new HIT.
In that four-hour binge I found myself accepting another perplexing HIT — a particularly juicy, $1.50 task. The job? Watch a silent video and describe it in 10 sentences. Then in five sentences. Then in three. A fun little exercise in summary for a writer, I thought to myself at the time. The video would feature a person, the instructions warned, but I was not to acknowledge anything but their deeds or include descriptions of their appearance or clothing — just how they affected the physical world around them.
Of course, I still couldn’t help but notice when the dark-haired man dressed in all black stopped slicing his apple to gesture at someone off-screen, seemingly confused. The nuance of his hand movements and facial expressions made it clear this wasn’t a one-sided video; he was having a conversation with a director, a supervisor, a… captor?
Watch it all. I dare you. (Just kidding, please don’t.).
My mind started wandering: What kind of direction could the man off-screen possibly be offering? Is the dark-haired man happy with his life of preparing fruit dishes on camera for later scrutiny by 50 cent internet mercenaries? Does he even know he’s being watched? And, again, the larger, ever-looming question: Why? What is this for?
Should you try it?
If I’ve made Mechanical Turk sound like a disturbing hole that sucks up countless dazed hours clicking away pondering the world as a strange unimaginable shape, that’s because it is. Or it’s a relatively fun way to make a couple bucks playing around on the internet.
I did this shit for hours. It’s addictive! Mechanical Turk is like gambling’s alternate universe cousin. It’s a bizarre cephalopodan slot machine, a thousand-armed bandit that pays you for the trouble of pulling one of its many strange levers. I found it alarmingly easy to slip into a Turking daze. Each strange task leaves you with a brief glimpse of some larger whole, and it’s easy to find yourself looking for just one more.
My experience is only the tip of the iceberg. At any given time there are some half a million HITs available. Of those, some require “Master Qualification”, a mysterious status that you can’t apply for but instead is bestowed upon those who are Turkin’ hard, as opposed to hardly Turkin’. Bilingual workers can generally find pretty juicy translation gigs. And then there are the folks on the assigning end, issuing their own strange tasks for the unseen ocean of Turkers to tackle in all their human glory.
And in the end, merely scratching the surface is the very essence of Mechanical Turk. To use it is to run through a library ripping books off the shelves and reading paragraphs at random, reveling in the impossible challenge of extrapolating the whole novel from the several sentences you can see. And you make a buck or two along the way.
Should you sign up and try it? Definitely. I’ll give you a nickel if you do.