Professional Documents
Culture Documents
DOI:10.1145/ 2597886
!13.3% of BS degrees, 28.7% of MS
It’s not just about being fair. degrees, and 19.2% of Ph.D.’s were
awarded to female candidates, down
BY BERYL NELSON from a high of 37% of BS degrees in
computer science in 1986.
! 5.3% of BS degrees were awarded
The
to African American candidates, as
were 2.7% of MS degrees, and 2% of
Ph.D.’s.
Among computing professionals,
about 20% of CS faculty in U.S. uni-
versities are women, and 1.6% are
Data on
African American.50 Similar numbers
exist in industry.
Diversity, bias, and stereotypes
have traditionally been discussed
in very relativistic terms: surveys of
whether people thought there was
bias, and so on. In recent years,
Diversity
imaginative researchers have devel-
oped ways to gather quantitative data
about the benefits of, as well as the
challenges to, having a diverse work-
force. This article explores the bene-
fits that diversity can bring to teams,
and the cognitive factors—namely,
stereotypes based on social group
membership—that keep us from
achieving optimal levels of diversity.
Benefits of Diversity
Diverse teams are more effective: they
can achieve more than
P EOPLE WOR KIN G TO G E T H E R produce better financial results and
better results in innovation. These re-
they can alone; this is a fundamental principle upon sults show that having a diverse orga-
which organizations are founded. Social scientists nization is a business imperative.
have shown that teams and organizations whose
key insights
members are heterogeneous in meaningful ways, for
" Teams and organizations whose
example, in skill set, education, work experiences, members are heterogeneous in
meaningful ways have a higher potential
perspectives on a problem, cultural orientation, for innovation than teams whose
and so forth, have a higher potential for innovation members are homogeneous.
" Social science experiments using
than teams whose members are homogeneous. quantitative methods show bias,
These findings are not without controversy, yet the stereotype threat, and methods to
combat them.
implications for the computing industry are profound, " Effectiveness of diverse teams depends
given the relative homogeneity of the field along a on trusting and supportive cultures.
Data publication is one of the most
few important dimensions. Take, for example, the important tools to identify and
combat identity threat and biased
composition of degrees awarded in computer science, decision making.
computer engineering, and informatics in 2012 at " There is hope! There are tools that have
been shown to combat bias and identity
research institutions in the U.S. threat effectively.
from 2007–2009 had 41% greater re- sales was 16.8%, as opposed to an aver- Similar results have been found
turn on equity, and 56% greater earn- age 11.5%; return on invested capital for race: organizations with greater
ings before interest and taxes than was 10%, as opposed to an average 6.2% racial diversity were associated with
companies with no women in the (see Figure 2).26 greater sales revenue, a larger number
executive committee, for companies This is also true in other geogra- of customers, greater market share,
within the same industrial sector (see phies. Profits of Indian companies and greater profits. In the same study,
N OV E MB E R 2 0 1 4 | VO L. 57 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 87
review articles
greater gender diversity was also found causation can be inferred, but not tion that occurs when organizations
to be associated with better results in proven. Researchers have posited are diverse. The next section reviews
sales revenue, number of customers, that diversity may actually be one this evidence.
and profitability.23 causal determinant of firm perfor- Innovation. Diversity has been
These studies are correlational: mance due to the increased innova- shown to create a cognitive and social
environment that is a positive indica-
Figure 1. Women at the top of corporations: Making it happen. Courtesy of McKinsey tor for innovation and a negative indi-
Report.15
cator for routine tasks. These dynam-
ics may have real-world consequences
Companies in the top quartile for the women for scholarship in our field.
representation in executive committee vs. sector In a study of collective intelligence
Companies with 0 women in executive committee and creativity, researchers gave sub-
in that specific sector jects aged 18 to 60 standard intelli-
gence tests and assigned them ran-
domly to teams. Each team of 3–5
Average return on equity Average EBIT margin people was asked to complete several
2007–2009 2007–2009
tasks, including brainstorming, deci-
Percent Percent sion making, and visual puzzles, and
22
to solve one complex problem that
was too difficult for one brilliant indi-
vidual to solve: that is, a team was re-
17 quired. Teams were given intelligence
+41%
scores based on their performance.
15 The only predictor of team collec-
+56% tive intelligence was whether there
11 were women on the team. Note that
all the high-scoring groups are close
to 50% women; some of the low-scor-
ing teams are also near 50%, but the
groups with little gender mix did not
score as highly. This was a surprise re-
sult to the researchers. With more in-
vestigation, it was found the difference
was having the social skills that made
it possible to use the contributions of
all the team members, and these corre-
late more with women than with men.
Figure 2. Financial performance at companies with three or more women board directors Figure 3 illustrates the relationship of
(WBD).
team composition to success.48,49
Patent authorship is another area
16.7% 16.8%
that shows a benefit of diversity. More
than 90% of all computer technology
patents issued in the U.S. since 1980
have been granted to men only. Yet
Average Average mixed gender patents are cited 26%
11.5% 11.5% to 42% more than any single gender
10.0% patent.4 An update to this report in
2012 showed that mixed gender pat-
ents typically have a large number of
Average
6.2%
authors. The higher citation rate (30%
to 40% more in the 2012 update) is
associated with higher numbers of
authors. The reasons for this are not
well understood.
These results show a correlation be-
tween diverse organizational compo-
Return on equity Return on sales Return on
sition, financial success, and innova-
investment capital tion. While there is not a clear causal
Courtesy of Catalyst Report26 relationship shown between diversity
and success, the results have been
shown with varying methodologies tions: the user’s implicit association of age and experience.5 Blindspot also
and in varying geographies, to a degree between two concepts is measured via considers the evidence for any predic-
that demands attention. user response time. There are reports tive link between measured biases and
for many associations, and the results outcomes or behaviors; results are not
Challenges Faced by Diverse Teams are stunning.5 yet conclusive in this area.
Unfortunately, it is not easy to make ! Almost everyone has measurable In individual cases, bias can be
diverse teams effective.a There are a biases (for example, 70%–80% have very difficult to verify. In aggregate,
number of forces that work against the biases against women in technology, however, bias and other forms of un-
desired effect: having the entire team or preferring white to African Ameri- conscious decision-making are read-
productive. There can be potential neg- can, or preferring young people). ily apparent:
ative effects of any of the following: ! Almost no one reports such biases ! About 58% of CEOs of Fortune 500
! unconscious bias, (for example, 15% report a preference companies are taller than 6ft. (about
! stereotype threat, for white people). 183cm), and almost a third are taller
! exclusion from critical social net- ! Even the people who are the sub- than 6ft. 2in. (about 188cm). In the
works, ject of a bias may have that bias. For population in general, about 14.5%
! lack of role models, and example, I tested as moderately bi- are taller than 6ft., and 3.9% taller
! unaware managers. ased against women in science and than 6ft. 2in.17 Most people would not
The following sections address technology, and this is totally against believe height predicts competence,
primarily unconscious bias and stero- my self-interest. and yet, these choices are frequently
type threat. The book Blindspot5 explains the made. In fact, height is strongly cor-
Unconscious bias. One of the factors development of the implicit asso- related with career success.
that both limits the diversity present ciation test and its results, exploring ! In a laboratory situation, appli-
in a team or organization, and that in- the reasons for differences between cants were seen in the waiting room
hibits the potential success of diverse unconscious perception and our as- either alone, or sitting next to an-
teams, is the unconscious bias, that sumptions. For example, it has been other applicant. Applicants who were
is, stereotypes, that we hold toward found that age is one of the strongest seen sitting next to an overweight
people based on the social groups to biases tested; this is true even among applicant were less likely to be hired
which they belong. Stereotypes are, the elderly, and in societies in Asia that than an applicant sitting alone or
simply, a constellation of traits, char- traditionally have valued the wisdom next to an average weight applicant,
acteristics, skills, and values that we
ascribe to members of social groups, Figure 3. Collective Intelligence vs. Team Composition.
such as gender, race, age, religion,
nationality, and others. These are
learned through cultural messages
and stories, comments from family
and friends, portrayals in the media,
and so forth. These stereotypes, de-
spite our best intentions, can bias our High
impressions of, and affect our actions
toward, others in our environment.
The stereotypes especially relevant in
work situations include those charac-
teristics that are visible, such as sex,
race, weight, and age; but also those
not visible but relatively easy to dis-
cern, such as educational background Average
and nationality.
Project Implicit at Harvard hosts
an online test of implicit associa-
N OV E MB E R 2 0 1 4 | VO L. 57 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 89
review articles
and regardless of their own weight.46 parently heavily influenced by the saying that we create a set of mental
! In an examination of data on wag- photos, at the end of the trial, no one shortcuts, or heuristics, for the situ-
es of a given population over the course remembered the photos changing.7 ations we experience every day, and
of five years, it was observed that there Clearly, there was unconscious deci- that we actively work against those
is a correlation between weight and sion making in each of these cases, biases for the most part, to the extent
income: for women, this is a negative but the data alone does not give tools we think the bias is not “good.” For
correlation; for men, a positive correla- to overcome bias. example, you see very small children,
tion, excluding obesity.27 Science faculty at respected U.S. as well as people with dementia, ex-
! A disturbing finding is that African research institutions show measur- pressing strong bias; or politicians
Americans and other minorities re- able bias toward male students: the may make a slip of the tongue when
ceive less effective healthcare interven- faculty were given applications for lab they are tired.46
tions and outcomes, even when con- manager that varied only in the gen- Until this point, I have been de-
trolling for socioeconomic class and der of the name. Male students were scribing the stereotyping process as
insurance coverage.5 rated significantly more competent, a negative force for individual and
! At a coffee station at a company and were given higher average salaries team functioning. However, this pro-
in Northern England, the sign for and more mentoring opportunities cess actually stems from an adaptive,
the amount to pay for milk (to put than women students. Importantly, often functional psychological pro-
in tea or coffee, on an honor sys- even the female faculty showed this cess. Mental heuristics and cogni-
tem) had a photo of flowers or eyes bias.32 Male faculty offered $30,520.83 tive shortcuts enable us to process
on alternate weeks. The amount of to male students on average, and information without conscious de-
money received was about double in $27,111.11 to female students. Female liberation: they fill in the informa-
the weeks with eyes over the weeks faculty offered $29,333.33 to male stu- tional gaps we often experience when
with flowers. And if the eyes were dents on average, and $25,000.00 to making decisions. In other words,
looking directly out, the amount re- female students. habits of mind help us to save brain
ceived was much higher; see Figure Shankar Vedantam explains the power for more difficult tasks. Joseph
4. Although their behavior was ap- bias illustrated in these examples by Pieper’s classic Leisure, the Basic of
Culture35 sets out this theory well. In
Figure 4. Coffee station pictures and money collected. Courtesy of Bateson et al.7 the 1980s, people thought we could
create expert systems by interview-
ing experts like brain surgeons or oil
Image
Time (weeks) exploration specialists, and creating
a rule chaining prolog environment
10 that would recreate their decision-
making ability. The problem was the
9 experts did not know how they knew
what they knew. That is, experts are
creating associations between dispa-
8
rate experiences and pieces of knowl-
Eye weeks
edge, using the subconscious brain.
7 Flower weeks In software development, we see this
in our most skillful engineers: they
can debug a complex problem, but
6
they will probably not be able to spec-
ify a set of rules that would work as
5 well as consulting the expert.
What it does say, however, is that we
4 should have better knowledge of our
own biases and unconscious decision-
making.
3
Stereotype threat. In Whistling Viv-
aldi,47 Claude Steele published a his-
2 tory of research about stereotype ef-
fects and identity threat. The original
1
problem he tried to solve was, why
were college entrance scores not pre-
dictive of college success for African
Americans? Steele asserts we each
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
have multiple identities (such as I
£ paid/liter milk consumed
am a woman; I am an MIT alumna; I
am American); and that we respond
N OV E MB E R 2 0 1 4 | VO L. 57 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 91
review articles
them, and we are going to work to- assist, then the rating of his work was
gether to make this happen.” The dif- not changed from the norm; if a wom-
ference is engagement.13 an refused to assist, her performance
! Communities of support: a key rating decreased.21
difference between African American
students and Asian students at U.S. col- Most people would If a man and a woman perform a
male-stereotyped task together, the
leges turned out to be that the Asian
students would get together, share
not believe that majority of people will attribute the
success of that task to the man, unless
their experiences, and study together. height predicts ! the contributions of the woman
People who feel isolated sometimes re-
fuse help, as they think that it will con-
competence, and are specifically attributed, or
! there is information about the way
firm the stereotype. The African Ameri- yet, these choices the task was structured, or
can students did not tend to form study
groups naturally; when these commu-
are frequently ! there is some clear example of pri-
or competence by the woman.22
nities were introduced, their perfor- made. In fact, There are differences in the com-
mance increased dramatically.45
! Fostering intergroup conversa- height is strongly munication styles considered ac-
ceptable in men and women leaders:
tions as learning opportunities. These correlated with women have a very narrow band of ac-
conversations are difficult, as people
are afraid of appearing to be biased or career success. ceptable behaviors. In general, women
are given less time to speak, are more
ill informed, so they often steer away likely to be interrupted, and if they do
from the situation. Subconsciously, interrupt someone, that is most likely
they pull their chairs further apart. to be another woman.43 As a conse-
When a facilitator starts a conversa- quence, women are less likely to be
tion by saying that tensions are natu- able to hold the floor in meetings: an
ral, but the conversation should be important quality in a leader.
treated as a learning experience, that Career advancement strategies
signals to people that differences can that work for men do not always work
be learned. People pull their chairs for women. When comparing only
closer together. This does not hap- women who have not taken career
pen when the facilitator says the par- breaks for family with men, there are
ticipants will not be judged, nor when still significant differences in achieve-
the facilitator says all points of view ment levels.10 The study on the Myth
are welcome.18 of the Ideal Worker says, “Women
Consequences of stereotyping. As benefit most by making their achieve-
mentioned earlier, there are a number ments known. Men benefit most by
of other forces that can hinder the ability scanning for external opportunities
of diverse teams to function optimally. and blurring work-life boundaries.
Stereotypes make us feel as though Both benefit by gaining access to pow-
we have useful information about peo- erful others.” The report also says,
ple’s strengths, weaknesses, and per- “changing jobs accelerated compen-
sonal characteristics. But they operate sation growth for men, but slowed it
in a more prescriptive way as well: They for women.” It has been shown that
shape our expectations of what people sponsorship, in which a sponsor ac-
should be doing, especially at work. tively promotes and takes risks for the
Thus, women are expected to be nur- sponsee, is more effective for women
turing and collaborative at work, in ac- than mentorship, in which the men-
cord with their stereotyped strengths. tor merely gives advice.24
When they deviate from that, they in- There are other career differences
cur penalties. Men are not expected to observable between men and women
be altruistic, but if they are, then they as well: women are more than 10% less
are given credit for giving assistance. likely than men to change jobs for a
This was demonstrated in an experi- raise in salary or for a promotion; but
mental setting in which a person was they are almost 10% more likely than
asked for help with a technical task: men to change jobs because of a bad
men were given an increased perfor- manager (see Figure 5).9 The numbers
mance rating if they gave assistance, who leave jobs to take care of fam-
whereas a woman who assisted was ily are very small for both men and
not; her performance rating remained women. In Japan as well, significant-
at the base level. If a man refused to ly fewer women report they have off-
ramped because of childcare-related Roughgarden, on the other hand, has cares about is a big motivator. A CS
reasons (32%) than those who have said that as a woman, “You get inter- class is required for all students in the
off-ramped because they feel stymied rupted when you are talking, you can’t first semester.
and stalled at work (49%).11 One in- command attention, but above all you ! First-year women students are
terpretation: It is difficult to balance can’t frame the issues.” encouraged to attend the Grace
both working and parenting, and one Ben Barres also wrote a strong com- Hopper Celebration of Women in
is not likely to want to stay unless the mentary in Nature6 refuting Lawrence Computing, a large (about 4,000 at-
work is rewarding. Moreover, career Summers’ speech, in which Summers tendees) conference, regardless of
priorities differ: A one-size-fits-all implied that women inherently have major; a large percentage of the stu-
management structure will not work less aptitude for science. He cited dents choose to attend. This provides
well for a diverse organization. studies showing that many selection role models and breaks the stereo-
This theory is supported by a study processes set an extremely high bar type about what computing is about:
of why women stay in engineering: peo- for competence for women and mi- hundreds of happy, successful, tech-
ple who choose to stay are engaged and norities: 2.5 times in the case of a re- nical women attend, at all stages of
basically hopeful, and are proactive search grant proposal. And yet, most their careers. Even the women who
about the problems they see.8 people have a strong desire to believe choose not to study computer sci-
Even a small bias can result in a the world is fair, so there is widespread ence say this event has a huge impact
large difference in the representa- belief that little discrimination exists. on their career choices, and many do
tion of minorities at the top levels of a He gave a number of recommenda- decide to study computer science as
company. A simulation of promotions tions to reduce discrimination. a result.
was performed with only 1% of bias in ! A research program was offered
ratings of women, and a posited eight Case Studies: What Works after the first year for about 10 female
levels for promotion. The simulation Many of the issues described so far students each year for four years. At
starts with 50% women at each level, are major societal issues, which lead this time, it is still true that some stu-
and ends when the entire organiza- to a diffuse pipeline of minority stu- dents (both male and female) have ac-
tion has been replaced with new em- dents into the workforce. Despite the cess to research after the first year.
ployees. At the end of the simulation, challenges posed by a narrow pipe- None of these suggestions offer a par-
at the lowest level of the organization line, bold leaders have made signifi- ticularly new insight alone, but in com-
there are 53% women, and the top lev- cant strides in reducing the gender bination they have had a dramatic and
els of an organization go to only 35% gap in computing, as we will show lasting effect. Consider these results:
women.29 through a few examples. ! In the same period, enrollment
Transgender studies show some At Harvey Mudd College, a liberal by women at Harvey Mudd increased
fundamental differences in the ways arts college of science and engineer- from 33% to 46%.
that society treats the same person ing, the college went from the average ! These changes have attracted
as either a man or a woman. These 12% of women in CS to about 40% in more computer science majors over-
studies are particularly interesting five years by taking these actions:19 all: Harvey Mudd graduated over 70
in that the same person experiences ! The introductory CS class was CS majors in 2014, compared formerly
life as both genders. The studies con- changed to be more holistic, and not with 25 to 30.
sistently show that men who change assume the students already know ! In 2014, 56% of those graduating
to women have an average lower sal- how to program. The act of designing with engineering degrees from Mudd
ary after the change, and women who systems to solve problems that one were female. The proportion of grad-
change to men have a higher salary, as
well as finding everyday life situations Figure 5. Reasons for leaving a first job. Courtesy of Carter et al.9
easier.38 A very powerful example is the
history of two biologists at Stanford
Women
University, Ben Barres (a man, former-
50% Men
ly a woman) and Joan Roughgarden (a
woman, formerly a man).46
38% 38%
Ben Barres has said, “When it
comes to bias, it seems that the desire 29%
27%
to believe in a meritocracy is so pow- 26% 25%
erful that until a person has experi-
16%
enced sufficient career-harming bias
themselves they simply do not believe
it exists … By far, the main difference 3% 2%
that I have noticed is that people who
don’t know I am transgendered treat Faster More Money/ Career Difficult Child
Advancement Better Benefits Change Manager Rearing
me with much more respect: I can
even complete a whole sentence with-
out being interrupted by a man.” Joan
N OV E MB E R 2 0 1 4 | VO L. 57 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 93
review articles
N OV E MB E R 2 0 1 4 | VO L. 57 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 95
review articles
N OV E MB E R 2 0 1 4 | VO L. 57 | N O. 1 1 | C OM M U N IC AT ION S OF T HE ACM 95