Page 434 -

14.3 Human computation 425

careful attention to appropriate rules for protection of human research participants.
Detailed discussions of human subject protections can be found in Chapter 15.
The considerable challenges and headache associated with deceptive online research
provide a strong argument against this sort of approach. If you find yourself tempted to try
this sort of study, consider a lab-based study instead. You may still use deception in this
case but the use of prior informed consent can help you avoid many difficult questions.
As with any HCI research, online research can be particularly challenging if there
is potential harm involved or when dealing with special cases, such as research in-
volving children. Technical measures such as encryption of transmitted data may
be useful for privacy protection and for verifying parental consent in the case of
minors (Kraut et al., 2004). Laws such as the Children's Online Privacy Protection
Act of 1998 (COPPA) in the United States may limit the amount of information that
can be collected from minors. Researchers working in these areas should construct
study materials carefully; consult with appropriate authorities responsible for human
research participant protection (known as Institutional Review Boards in the United
States—see Chapter 15) and external experts to review proposed procedures; and use
traditional studies as opposed to online studies when appropriate (Kraut et al., 2004).

14.3 HUMAN COMPUTATION
14.3.1 INTRODUCTION TO HUMAN COMPUTATION
What can people do more effectively than computers? Despite the frustrations as-
sociated with seemingly endless bugs and glitches, most people who use computers
frequently would probably agree that computers do many jobs more quickly and
more accurately than humans (if you ever talk to someone who disagrees, see what
happens if you ask them to give up their smartphone or laptop). However, there
are some areas where humans continue—at least for the time being—to outperform
computers. Tasks requiring detailed interpretation of complex inputs are a prime
example. Despite recent improvements in computer vision, natural-language pro-
cessing, and other fields of artificial intelligence, software systems often struggle to
identify objects in digital images or to interpret written text, even when such tasks are
straightforward for many humans.
Given these differing—and often complementary—strengths of both humans and com-
puters, many observers have argued for the use of computers to augment human cognition
(Shneiderman, 2002). This line of inquiry dates back to the prehistory of HCI, in specula-
tive designs such as Vannevar Bush's Memex (Bush, 1945) and Douglas Englebart's work
on augmenting human intellect (Engelbart, 1962), which led to the famous 1968 demos of
the first computer mouse, early graphical user interface, and word processor.
Human Computation takes the opposite approach. Given a task that might be hard
for a computer but relatively easy for a human, a human computation strategy might
ask multiple humans to complete small pieces of that task. For example, consider a
computer vision algorithm for identifying numerals in digital photographs. A ma-
chine learning tool for such a task might be challenged by the range of sizes, fonts,