an under–emphasis on quality” [62]. Complex work, which is
often subjective to evaluate, falls victim to this pitfall.
Comparing the phenomena
The research on piecework tells us that we should expect it to
thrive in industries where the nature of the work is limited in
complexity [22], and become less common as work becomes
more complex. Has computation shifted piecework’s previous
limits of expertise, measurement, and evaluation?
In some ways, yes: technology increases non–experts’ lev-
els of expertise by giving access to information that would
otherwise be unavailable. For example, taxi drivers in Lon-
don endure rigorous training to pass a test known as “The
Knowledge”: a demonstration of the driver’s comprehensive
familiarity with the city’s roads. This test is so challenging that
veteran drivers develop significantly larger the regions of the
brain associated with spatial functions such as navigation [101,
102, 140, 141, 160, 159]. In contrast, with on–demand plat-
forms such as Uber, services such as Google Maps and Waze
make it possible for people entirely unfamiliar with a city to op-
erate professionally [139, 65]. Other examples include search
engines enabling information retrieval, and word processors
enabling spelling and grammar checking. By augmenting the
human intellect [43], computing has shifted the complexity of
work that is possible with minimal or no training.
Algorithms have automated some tasks that previously fell to
management. Computational systems now act as “piecework
clerks” [52] to inspect and modify work [72, 106]. However,
these algorithms are less competent than humans at evalu-
ating subjective work, as well as in their ability to exercise
discretion, causing new problems for workers and managers.
Implications for on–demand work
Algorithms are undoubtedly capable of shepherding more
complex work than the linear processes available to Airy and
Ford. However, as work becomes more complex, it becomes
increasingly difficult to codify a process to achieve it [44,
41]. So, while algorithms will increase the complexity ceiling
beyond what was possible previously with piecework, there is
a fundamental limit to how complex such work can become.
Technology’s ability to support human cognition will enable
stronger assumptions about workers’ abilities, increasing the
complexity of on–demand work outcomes. Just as the shift
to expert crowdsourcing increased complexity, so too will
workers with better tools increase the set of tasks possible.
Beyond this, further improvements would most likely come
from replicating the success of narrowly–slicing education
for expert work as Hart and Roberts and later Grier described
in their piecework examples of human computation [55] and
drastically reformulating macro–tasks given the constraints of
piecework [63]. An argument might be made that MOOCs
and other online education resources provide crowd workers
with the resources that they need, but it remains to be seen
whether that work will be appropriately valued, let alone prop-
erly interpreted by task solicitors [7]. If we can overcome this
obstacle, we might be able to empower more of these workers
to do complex work such as engineering, rather than doom
them to “uneducated” match–girl reputations [134]. However,
many such experts are already available on platforms such as
Upwork, so training may not directly increase the complex-
ity accessible to on–demand work unless it makes common
expertise more broadly available.
Evaluation remains as difficult for crowd work as it did for
the efficiency experts. Reputation systems for crowdsourcing
platforms remain notoriously inflated [67]. Ultimately, many
aspects of assessment remain subjective: whether a logo made
for a client is fantastic or terrible may depend on taste.
So, in the case of complexity, the history of piecework does
not yet offer compelling evidence that on–demand work will
achieve far more complex outcomes than piecework did. Im-
provements in workflows, measurement, and evaluation have
already been made, and it’s not immediately clear that the re-
maining challenges are readily solvable. However, on–demand
work will be far more broadly distributed than piecework his-
torically was — reaching many more tasks and areas of exper-
tise by virtue of the internet.
Decomposing Work
At its core, on–demand work has been enabled by decom-
position of large goals into many small tasks. As such, one
of the central questions in the literature is how finely–sliced
these microtasks can become, and which kinds of tasks are
amenable to decomposition. In this section, we place these
questions in the context of piecework’s Taylorist evolution.
The perspective of on–demand work
Many contributions to the design and engineering of crowd
work consist of creative methods for decomposing goals. Even
when tasks such as writing and editing cannot be reliably per-
formed by individual workers, researchers have demonstrated
that the decomposition of these tasks into workflows can suc-
ceed [84, 14, 147, 110]. These decompositions typically take
the form of workflows, instantiated as algorithmically man-
aged sequences of tasks that resolve interdependencies [17].
Workflows often utilize a first sequence of tasks to identify an
area of focus (e.g., a paragraph topic [84], an error [14], or a
concept [163, 164]) and a second sequence of tasks to execute
work on that area. This decomposition style has been success-
fully applied across many areas, including food labeling [112],
brainstorming [137, 163], and accessibility [90, 87, 88].
If decomposition is key to success in on–demand work, the
question arises: what can, and can’t, be decomposed? More
pointedly, how thinly should work be sliced and subdivided
into smaller and smaller tasks? The general trend has been
that smaller is better, and the microtask paradigm has emerged
as the overwhelming favorite [148, 146]. This work illus-
trates a broader sentiment in both the study and practice of
crowd work, that microtasks should be designed resiliently
against the variability of workers, preventing a single errant
submission from impacting the agenda of the work as a whole
fully exploiting the abstracted nature of each piece of work
[71, 89, 149]. In this sense, finer decompositions are seen as
more robust — both to interruptions and errors [32] — even if
they incur a fixed time cost. At the extreme, recent work has
demonstrated microtasks that take seconds [150, 23] or even
fractions of a second [85]. However, workers perform better