Page 437 -

428 CHAPTER 14 Online and ubiquitous HCI research

CAPTCHA AND reCAPTCHA—CONT'D
Subsequent work merged CAPTCHA's goal of using human intelligence
tasks as security with the ESP games notion of using these tasks to
accomplish useful work, leading to the reCAPTCHA tool (von Ahn et al.,
2008). reCAPTCHA was designed to solve the problem of digitizing
text that had proven challenging for optical-character recognition (OCR)
systems. reCAPTCHA provides users with images including text that has
proven difficult for computer vision systems to interpret. Specifically, the
original reCAPTCHA asked users to decipher words that have each failed
to be consistently recognized by two different OCR programs. Each time a
reCAPTCHA is used, the user is asked to interpret images containing two
words: one for which the interpretation is known, and another which has not
yet been classified. If the user provides a correct answer for the known word,
the answer for the other word is assumed to be correct. Each word is presented
to multiple users, and words can be promoted to become known words if
sufficient accurate human guesses are provided. All words are distorted in an
attempt to defeat computer vision programs (Figure 14.1A) (von Ahn et al.,
2008). reCAPTCHA has been used on many web sites to provide the security
that motivated the design of the original CAPTCHA, primarily verification
of user registration and login on web sites. reCAPTCHA was purchased by
Google in 2009 (Zlatos, 2009), with subsequent evolution of the tool including
variants for labeling images (Figure 14.1B) and predictive tools capable of
identifying users as human based on interactions with the widget, without the
need for image labeling (Shet, 2014).
reCAPTCHA's use of images highlights a key design challenge. The image-
labeling tasks in the ESP game were purely entertainment on the part of the
users. CAPTCHAs, on the other hand, are often used on sites that might be the
sole route for users to access functionality needed for personal or professional
purposes. As a result, accessibility becomes a key concern, as some users—
particularly those with low vision or blindness—might struggle with some of
the images used in tools like reCAPTCHA. This problem is magnified by the
nature of the tools—by definition, the images used in reCAPTCHA are those
that have been in some ways hard to process. reCAPTCHA has always had an
audio option, which has generally asked users to type a sequence of spoken
digits. Alternative CAPTCHA tests have been the subject of multiple research
efforts (Sauer et al., 2010; Davidson et al., 2014).

Although reCAPTCHA is likely the most familiar human computation task, the
notion of using games to motivate participation has been used in many different do-
mains. Online games have been particularly successful in scientific fields, with the
Fold.It game (http://fold.it) harnessing the power of multiple users to generate high-
quality protein models (Khatib et al., 2011; Eiben et al., 2012) and bioinformatics