Google Image Labeler

» see images.google.com/imagelabeler as it lived, on the wayback machine

~ the last words ~

In 2006 an image search was a text search wearing a costume. The engine could read the filename and whatever words happened to sit around the picture on the page. It could not look at the picture. Everybody working near that problem knew it, and the honest fix was to get a human being to look at an image and say what was in it, several million times over.

Google's answer went up on 31 August 2006 and it was a game. You were matched with an anonymous partner somewhere in the world, both of you shown the same image with ninety seconds on the clock. You typed what you saw. When your word matched theirs, the pair scored and the next image came up. Some images carried a list of off-limits words already collected, which pushed you past the obvious answer toward something more specific.

The mechanic was not Google's. Luis von Ahn had built it at Carnegie Mellon as the ESP Game, part of a body of work he called human computation, resting on the argument that people will do useful labor for nothing if the labor has been shaped like a game. He licensed it to Google, then went on to reCAPTCHA, where the unpaid work was digitizing scanned books, and after that to Duolingo.

This one lived on a Labs page, which is where most people met it and where most people left it. I cannot claim it as a habit. What is worth keeping is how strange the arrangement was and how normal it looked: for five years, part of the world's largest image index was assembled by pairs of strangers agreeing on a noun, for points that could not be spent on anything.

~ the record ~

~ the matching rule ~

The design problem with asking people to label things is that you cannot check the answers. Verifying one person's label costs about as much as producing it, which puts you back where you started. Agreement solves it cheaply. Two strangers with no way to communicate, looking at the same picture, arriving at the same word, is a much stronger signal than either of them typing it alone.

It also polices itself. Nonsense never matches, so there is no reward in typing nonsense, and the incentive to cheat evaporates without anyone having to moderate anything. The off-limits list did the other half of the job. Once enough people had said dog, the word was retired for that image and the next pair had to reach for the breed or the collar or what the animal was standing on.

~ who was really working ~

The players thought they were playing and they were also the workforce. This was never hidden, which matters: the page said plainly that the labels improved image search, so it was a trade rather than a trick. The wage was points, a leaderboard and the pleasant feeling of matching a stranger in four seconds.

What it demonstrated is that the exchange rate on that wage is astonishingly good when the task is genuinely fun. Nobody has ever been able to reproduce it reliably. Most attempts at gamified labor since have been an unpleasant job with a progress bar bolted onto it, and the players notice within a minute.

~ the second version ~

Google brought the name back in 2016 inside Crowdsource, and the new thing kept the label and threw away the game. No partner, no clock, no competition. You were shown a category and asked whether it applied, one tap at a time, which is data entry with a friendly interface.

The change is a good measure of what happened in between. In 2006 the machine needed a vocabulary from scratch and the only source was people, so it was worth building a game to get them. By 2016 the machine could already guess, and what it needed was somebody to say yes or no to its guesses. The original closed on 16 September 2011, at roughly the point where a stranger typing dog for ninety seconds had stopped being worth the ninety seconds.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »