Overkill

absurdly advanced tech, explained badly, used for dumb stuff

All overkills

CAPTCHAs

or: proving you’re human to robots who do it better

By 2010 people were solving more than 100 million squiggly-word puzzles a day just to prove they were human, so a computer scientist turned each one into a tiny unpaid job, and the whole internet helped digitize the New York Times archive and piles of old books one smudged word at a time1. Then the puzzles switched to Street View photos, and we all spent years clicking crosswalks and traffic lights2, until a 2023 study found that bots solve the distorted-text kind in under a second with 99.8% accuracy, while people take 9 to 15 seconds and get 50 to 84% right3.

And we use it to prove we’re not robots to robots that are better at it than we are.

How it works, badly

  1. #
    w0rdthanks!
    Every squiggly word you typed was a word from an old book.
  2. #
    select all traffic lights
    Then we clicked traffic lights for years.
  3. #
    doneis that a bus?
    The robots got better at it than us.
  4. #
    I’m not a robotI promise
    Now we prove to the robots that we aren’t one.

Wait, is that real?

Yes, you worked for free, and yes, the bots are better at it now.

The unpaid typing job↩

reCAPTCHA came out of Carnegie Mellon University in 2007. Instead of showing a random squiggly word, it showed words that scanning software had failed to read, so each person who typed one was transcribing a scrap of an old document. It helped digitize the New York Times archive, and Google Books later used it the same way. Google bought reCAPTCHA in 2009, and by 2010 it was reported to show more than 100 million CAPTCHAs a day.

The traffic lights↩

In 2012, reCAPTCHA began using Street View photos, asking people to pick out things like crosswalks and street lights. In 2014, Google moved to the "I’m not a robot" checkbox, which watches how you behave and only shows image puzzles when it suspects a bot.

The robots won↩

A 2023 study from the University of California, Irvine, presented at the USENIX Security Symposium, had about 1,400 participants solve 14,000 CAPTCHAs. For distorted-text CAPTCHAs, people took 9 to 15 seconds and were 50 to 84% accurate, while bots took under a second and were 99.8% accurate.

The fine print↩

Press coverage of the study gives the participant count as both 1,000 and 1,400. The image puzzles helped label photos, but Google hasn’t said exactly how much each click taught its software. Newer CAPTCHAs mostly judge how you move and browse rather than what you click, so many people now see a checkbox or nothing at all. The bots in the study were researchers’ tools, not every bot on the internet.

Sources: Wikipedia: reCAPTCHA; UC Irvine: new findings on CAPTCHAs; The Register: bots beat humans at CAPTCHAs; USENIX Security 2023: An Empirical Study & Evaluation of Modern CAPTCHAs

Last fact-checked .