Not Learning From Failure—the Greatest Failure of All

Abstract

Our society celebrates failure as a teachable moment. Yet in five studies (total N = 1,674), failure did the opposite: It undermined learning. Across studies, participants answered binary-choice questions, following which they were told they answered correctly (success feedback) or incorrectly (failure feedback). Both types of feedback conveyed the correct answer, because there were only two answer choices. However, on a follow-up test, participants learned less from failure feedback than from success feedback. This effect was replicated across professional, linguistic, and social domains—even when learning from failure was less cognitively taxing than learning from success and even when learning was incentivized. Participants who received failure feedback also remembered fewer of their answer choices. Why does failure undermine learning? Failure is ego threatening, which causes people to tune out. Participants learned less from personal failure than from personal success, yet they learned just as much from other people’s failure as from others’ success. Thus, when ego concerns are muted, people tune in and learn from failure.

Full text

https://doi.org/10.1177/0956797619881133
Psychological Science
1 –12
© The Author(s) 2019
Article reuse guidelines:
sagepub.com/journals-permissions
DOI: 10.1177/0956797619881133
www.psychologicalscience.org/PS
ASSOCIATION FOR
PSYCHOLOGICAL SCIENCE
Research Article
Our society celebrates failure as a teachable moment
(Brown, 2014; Maxwell, 2007). Research appears to sup-
port this point. People react more strongly—physiologi-
cally, cognitively, and emotionally—to negative events
than positive ones (Baumeister, Bratslavsky, Finkenauer,
& Vohs, 2001; Kahneman & Tversky, 1979; Rozin &
Royzman, 2001; Taylor, 1991) in ways that arguably
enhance learning. For example, compared with positive
stimuli, negative stimuli command more attention
(Öhman, 2007; Pratto & John, 1991) and increase infor-
mation processing (Bless & Fiedler, 2006; Ohira,
Winton, & Oyama, 1998; Puig & Szpunar, 2017; Taylor,
1991). It follows that people may pay attention to fail-
ure, process it, remember it, and thus learn from it—as
much or more than they learn from success.
Here, we explored the alternative: that people learn
less from failure. Whether people learn from failure
depends not only on whether failure is attention grab-
bing but also on people’s motivation to attend to it.
If people are motivated to ignore their failures, then
they will not attend to them and will not learn from
them. For example, if researchers are motivated to
ignore failed experiments, they will learn nothing
from them.
Motivation research demonstrates that failure often
undermines goal commitment, leading people to dis-
engage from their goals (Cochran & Tesser, 1996;
Soman & Cheema, 2004). Failure has this undermining
influence when people interpret it personally (Hattie
& Timperley, 2007). For example, novices infer from
negative feedback that they are not committed to the
goal in question (Fishbach & Finkelstein, 2012), and
many students interpret failure to mean they lack apti-
tude, which discourages subsequent goal pursuit
(Yeager & Dweck, 2012). According to several motiva-
tional theories, negative feedback lowers people’s
881133PSSXXX10.1177/0956797619881133Eskreis-Winkler, FishbachNot Learning From Failure
research-article2019
Corresponding Authors:
Lauren Eskreis-Winkler, Booth School of Business, University of
Chicago, 5807 S. Woodlawn Ave., Chicago, IL 60637
E-mail: leskreiswinkler@uchicago.edu
Ayelet Fishbach, Booth School of Business, University of Chicago,
5807 S. Woodlawn Ave., Chicago, IL 60637
E-mail: ayelet.fishbach@ChicagoBooth.edu
Not Learning From Failure—the Greatest
Failure of All
Lauren Eskreis-Winkler and Ayelet Fishbach
Booth School of Business, University of Chicago
Abstract
Our society celebrates failure as a teachable moment. Yet in five studies (total N = 1,674), failure did the opposite: It
undermined learning. Across studies, participants answered binary-choice questions, following which they were told
they answered correctly (success feedback) or incorrectly (failure feedback). Both types of feedback conveyed the
correct answer, because there were only two answer choices. However, on a follow-up test, participants learned less
from failure feedback than from success feedback. This effect was replicated across professional, linguistic, and social
domains—even when learning from failure was less cognitively taxing than learning from success and even when
learning was incentivized. Participants who received failure feedback also remembered fewer of their answer choices.
Why does failure undermine learning? Failure is ego threatening, which causes people to tune out. Participants learned
less from personal failure than from personal success, yet they learned just as much from other people’s failure as from
others’ success. Thus, when ego concerns are muted, people tune in and learn from failure.
Keywords
learning, feedback, failure, ego threat, motivation, open data, open materials, preregistered
Received 5/16/19; Revision accepted 9/5/19

2 Eskreis-Winkler, Fishbach
confidence in their overall ability to pursue their goals,
as well as their general expectations of success
(Atkinson, 1964; Bandura & Cervone, 1983; Lewin, 1935;
Weiner, 1974; Zajonc & Brickman, 1969). Only experts
appear able to sustain commitment in failure’s aftermath
(Louro, Pieters, & Zeelenberg, 2007).
Although past work has documented that failure
undermines goal commitment and future goal pursuit,
it is unclear how or whether failure affects motivation
in the moment of failure itself. Past research has not
explored how quickly the motivational system disen-
gages following a failure. It is possible that failure gen-
erates an immediate motivational shutdown, undermining
the individual’s motivation to attend to the task at hand.
This response—“tuning out”—would imply, for exam-
ple, that batters who strike out stop paying attention to
the game and do not think about the way the pitch
crossed the plate, the way they swung the bat, or why
they struck out—which means they cannot learn from
the experience.
We explored the possibility that failure—specifically,
failure feedback on a task—compromises people’s
motivation to learn during the failed experience itself.
We predicted that the way people responded to failure
feedback would undercut their ability to learn from it.
Because people find failure ego threatening, they will
disengage from the experience, which means they stop
paying attention. Such tuning out has direct conse-
quences for learning because people cannot learn infor-
mation that they have not attended to. Because failure
undermines people’s motivation to attend to a task, we
hypothesized that people would learn less from failure
than from success, even when failure and success were
equally informative.
We tested this prediction using a novel research para-
digm. In this paradigm, participants completed a learn-
ing phase, in which they guessed the correct answer to
a binary-choice question (“is X ‘A’ or ‘B’?”), and then
received (on the next screen) either failure feedback
(“Incorrect!”) or success feedback (“Correct!”). Next,
they were tested on the content of the initial questions,
to see whether they learned from the feedback.
To explore why people tune out after failure feed-
back, we examined whether ego involvement mediated
and moderated the effect of failure on learning. Specifi-
cally, we examined whether drops in self-esteem medi-
ated the effect of negative (vs. positive) feedback on
learning. We also compared how much people learned
from their own failures and successes versus how much
they learned from others’ failures and successes. Unlike
personal failures, others’ failures do not involve the
ego. Thus, we expected people to learn the same
amount from others’ failures and others’ successes. It
follows that people will learn more from others’ failures
than from their own.
Here, we report five studies that tested whether fail-
ure undermines learning and whether ego involvement
mediates and moderates this effect. First, we examined
whether employees (telemarketers) learned less from
failure (vs. success) when completing a learning task
that was relevant to their profession. Next, we tested
whether people learned less from failure (vs. success)
in a learning task that contained language and relation-
ship stimuli (i.e., learning the meaning of linguistic
symbols and guessing which couples were real or not
real). Across studies, we varied learning incentives in
order to examine whether the inability to learn from
failure generalizes across low and high levels of motiva-
tion. This range of stimuli and contexts allowed us to
assess whether the failure to learn from failure is a
generalizable, meaningful, real-world phenomenon.
To maximize power, we honed our measures and
manipulations in pilot studies. These pilot studies
yielded medium to large main effects (d = 0.76, d =
0.65, and d = 0.63; see the Supplemental Material avail-
able online). Accordingly, we targeted a minimum sam-
ple of 50 participants per cell. Power analyses conducted
in G*Power software (Version 3.1; Faul, Erdfelder, Lang,
& Buchner, 2007) on respective sample sizes and target
alpha level (.05) revealed that power was sufficient
across all studies (i.e., ≥ .80) to detect a medium to
large effect (e.g., d = 0.60, ηp2 = .08). Sample sizes were
determined prior to data collection. All studies con-
ducted for this research, including pilot studies, are
reported in either the manuscript or the Supplemental
Material (all materials and data are posted on the Open
Science Framework at https://osf.io/5kbx6).
Study 1: Not Learning From Failure
in the Workplace
In Study 1, we examined whether telemarketers at a
call center learned less from failure than from success.
In the learning phase, telemarketers completed multiple-
choice questions about customer service—a topic of
relevance to their profession. We randomly assigned
them to receive success feedback on the questions they
got right or failure feedback on the questions they got
wrong. Both success and failure feedback contained
full information on the correct answers. Nevertheless,
we predicted that on a test, participants would demon-
strate higher levels of learning following success than
following failure.
Method
Participants. All telemarketers at a call center in the
Midwest were e-mailed an invitation to complete a sur-
vey, which was described as an opportunity to gauge
employee attitudes and work ethic for the purpose of

Not Learning From Failure 3
improving company culture. In this study, rather than
predetermining a sample size, management invited all
company employees to participate. In total, 422 telemar-
keters were randomly assigned to condition (success: n =
218, failure: n = 204). A sizable portion of telemarketers
(success: n = 46, failure: n = 47) dropped out before
completing the survey; however, the between-condition
difference in attrition was not significant, χ2(1) = 0.23, p=
.631. We ran analyses on the subset of telemarketers who
completed the survey (success: n = 172, failure: n = 157;
63.8% female; age: M = 30.44 years, SD = 10.97).
Procedure. Telemarketers read that they would be tak-
ing a survey about customers (“Today, we will ask you
little-known facts about customers—what they like, dis-
like, how their experiences influence their attitudes to
your company, etc. Your goal is to learn as much as you
can”). To make the relevance of the survey salient, we
reminded participants that their daily jobs are primarily
about satisfying customers (“Much of your job involves
pleasing customers”). Participants learned that they
would answer a series of questions and that because of
time constraints, they would receive performance feed-
back on only a few questions.
Participants then answered 10 trivia questions about
customer satisfaction and customer service. Each ques-
tion had two answer choices (e.g., “How much money,
annually, do U.S. companies lose due to poor customer
service? A. Approximately $90 billion, B. Approximately
$60 billion”). The questions were based on customer-
service facts taken from recent polls and research stud-
ies (Help Scout, 2018).
For this learning phase, participants were randomly
assigned to a failure or a success condition. Participants
in the success condition received success feedback
(“Your answer was correct”) on the first four questions
they answered correctly. They did not receive feedback
on any other questions. If participants had fewer than
four correct answers, they received feedback on just
the questions answered correctly (94% received feed-
back on four questions, 4% received feedback on three
questions, and 2% received feedback on two questions).
Participants in the failure condition received failure
feedback (“Your answer was incorrect”) on the first four
questions they answered incorrectly. They did not
receive feedback on any other questions. If participants
had fewer than four incorrect answers, they received
feedback on just the questions they answered incor-
rectly (75% received feedback on four questions, 16%
received feedback on three questions, 7% received
feedback on two questions, and 2% received feedback
on one question). The fact that each question had just
two answer choices meant that both failure and success
feedback provided participants with full information
on the correct answer. Success feedback communicated
the correct answer directly; from failure feedback, par-
ticipants could infer that the answer they did not select
was the correct one. Feedback was presented on the
screen that followed the participants’ choice.
We made the quiz 10 questions long and used
obscure trivia facts to ensure that most participants
would get 4 questions correct and 4 questions incor-
rect—a response distribution that would allow us to
deliver similar amounts of feedback across participants
in the two conditions.
Next, participants completed a distractor activity, in
which they described an impactful, positive experience
they had at the company. Then they entered the test
phase. The test consisted of only the questions on
which participants had previously received feedback
(i.e., up to four questions in total). These questions had
the same answer choices as the initial quiz questions,
but the questions themselves were phrased in the
reverse (e.g., “Which of the following amounts is NOT
the amount that U.S. companies lose annually due to
poor customer service? A. Approximately $90 billion,
B. Approximately $60 billion”). We operationalized
learning as the percentage of questions on this test that
the participant answered correctly.
Results
Overall, participants performed slightly better than
chance (50%) at identifying the correct answers to the
quiz questions in the learning phase (they got 56%
right, and that hit rate was similar in both conditions).
As a result, participants in the failure condition received
feedback and were tested on fewer questions (M = 3.64,
SD = 0.70) than participants in the success condition
(M = 3.92, SD = 0.32), t(327) = 4.75, p < .001. This may
have made learning from failure easier than learning
from success because participants in the failure condi-
tion received, had to remember, and were tested on
less information.
Yet as predicted, participants in the success condi-
tion scored higher on the test (M = 62% correct, SD =
26%) than participants in the failure condition (M = 48%
correct, SD = 28%), t(327) = 4.71, p < .001, d = 0.52,
95% confidence interval (CI) = [0.30, 0.74]. Whereas
participants in the success condition learned at a rate
above chance, t(171) = 5.84, p < .001, those in the fail-
ure condition did not, t(156) = −1.03, p = .303 (see Table
1 and Fig. 1).
Study 2: Not Learning From Failure
in Controlled Studies
Although Study 1 lent preliminary support to the
hypothesis that failure undermines learning motivation,
it is possible that participants’ prior knowledge of the

4 Eskreis-Winkler, Fishbach
(admittedly obscure) trivia questions biased the results.
Further, it is possible that participants in the failure
condition got feedback on (and thus had to learn) ques-
tions that were more difficult. Accordingly, in Study 2,
we developed a script task that contained researcher-
invented script symbols of which participants had no
prior knowledge. We randomly assigned participants
to receive success or failure feedback on the same
questions, and then we tested participants to see if they
learned from the feedback. Participants received a
monetary bonus for each correct answer. We report four
iterations of this task.
Method
Participants. Individuals of any nationality were invited
to participate as long as their Amazon Mechanical Turk
(MTurk) approval rating was at or above 50%. We recruited
four samples. In each one, we opened the survey to 100
participants, except for the Study 2a replication, which was
Table 1. Comparison of Results From the Test Phases of Studies 1, 2, 4, and 5
Study N
Correct
answers
in failure
condition
Correct
answers
in success
condition
Between-conditions
comparison Cohen’s d95% CI
Study 1 329 48% (28%) 62% (26%)* t(327) = 4.71, p < .001 0.52 [0.30, 0.74]
Study 2a 99 59% (41%) 80% (35%)* t(97) = 2.80, p = .006 0.55 [0.15, 0.95]
Study 2a replication 325 66% (36%)* 88% (25%)* t(323) = 6.17, p < .001 0.71 [0.49, 0.94]
Study 2b 102 77% (31%)* 90% (22%)* t(100) = 2.57, p = .012 0.51 [0.11, 0.90]
Study 2c 114 51% (44%) 81% (38%)* t(112) = 3.86, p < .001 0.72 [0.34, 1.10]
Study 2d 103 67% (34%)* 91% (21%)* t(101) = 4.30, p < .001 0.85 [0.44, 1.25]
Study 4 100 68% (36%)* 88% (27%)* t(298) = 5.65, p < .001 0.65 [0.42, 0.88]
Study 5
Self condition 202 69% (38%)* 83% (33%)* F(1, 400) = 25.68, p < .001 0.71 [0.43, 1.00]
Other condition 200 80% (32%)* 82% (31%)* F(1, 400) = 0.81, p = .369 0.04 [−0.04, 0.13]
Note: In the two columns showing the average percentage of correct answers, standard deviations are given in parentheses. Asterisks
denote learning that exceeded chance. CI = confidence interval.
*p ≤ .001.
62% 80% 88% 90%81% 91%88% 83% 82%48% 59% 66% 77% 51% 67% 68%69% 80%
0%
20%
40%
60%
80%
100%
Study 1 Study 2a Study 2a
Rep
Study 2b Study 2c Study 2d Study 4 Study 5
(Self )
Study 5
(Other )
Correct Answers in Test Phase
Success Condition Failure Condition
Fig. 1. Average percentage of correct answers in the test phases of Studies 1, 2, 4, and 5 as a function of the success- and failure-
feedback conditions. Error bars represent ±1 SE.

Not Learning From Failure 5
preregistered and opened to 300 participants, in accor-
dance with a reviewer’s advice (http://aspredicted.org/
blind.php?x=it2ej3). In Study 2a, MTurk returned 99 respon-
dents (46.5% female; age: M = 32.71 years, SD = 10.39); in
the Study 2a replication, MTurk returned 325 respondents
(49.8% female; age: M = 40.77 years, SD = 12.82); in Study
2b, MTurk returned 102 responses (45.1% female; age: M =
30.21 years, SD = 10.61); in Study 2c, MTurk returned 114
responses (44.7% female; age: M = 36.96 years, SD = 11.40);
and in Study 2d, MTurk returned 103 responses (51.5%
female; age: M = 34.98 years, SD = 10.16).
Procedure. Prior to randomization, all participants were
asked to respond to the following question in an open-
text response field: “Please tell us what is your favorite
book, and why.” We included this question because
online participants who are not willing to invest effort
tend to drop out when they see an open response ques-
tion.1 Participants were randomly assigned to a condition
only if they answered this question. Next, participants
took the script task.
The script task in Study 2a consisted of three ques-
tions in Round 1 (learning phase). Each question asked
participants to guess which of two symbols had a spe-
cific meaning in an invented language (e.g., “Which of
the following characters in an ancient script represents
an animal? or ”). Notably, as in Study 1, success and
failure feedback contained equivalent amounts of infor-
mation. Feedback was presented on the screen after
participants chose each answer.2 (Note that unlike
Study 1, in which participants received feedback on
only some questions, participants received feedback on
every question they answered.) Both types of feedback
provided full information on the correct answer because
each question had only two answer choices. Prior to
the beginning of Round 1, participants read that per-
formance on the later test (Round 2) was incentivized:
“Whether you get the answers right or wrong in Round
1, try to learn. Round 2 will test how much you learned
and bonus you $0.10 for each question you get correct!”
We incentivized performance in order to ensure that
participants were sufficiently motivated to learn in an
online context.
After completing Round 1, participants completed a
brief distractor activity—an open text box in which they
reflected on their favorite music (“Tell us: what is your
favorite music to listen to?”). Next, participants took a
test that measured their learning. The three questions
on the test (Round 2) paralleled each of the initial ques-
tions (Round 1) but were phrased in the reverse. For
example, in Round 1, one question read, “Which of the
following characters in an ancient script represents an
animal?” On the test, participants had to answer a ques-
tion with the same two symbol choices, but the question
was rephrased as, “Which of the following characters
represents a non-living, stationary object?” Because
there were only two symbol choices, all participants
could deduce that the symbol that was an animal in the
first round was not the “non-living, stationary object”
on the test, and vice versa. Thus, from the feedback, all
participants received the information that would allow
them to answer the test questions correctly. See the
Appendix for the exact manipulation.
The Study 2a replication was a direct, preregistered
replication of Study 2a. Study 2b used the same task as
Study 2a, except with higher learning incentives ($1.50
instead of $0.10), in order to test whether participants
would continue to learn less from failure under condi-
tions that spurred higher motivation.
Study 2c used an iteration of the script task in which
learning from failure required fewer mental inferences
than learning from success. One reason that people
may learn less from failure is that learning from failure
(vs. from success) requires the participant to make
more mental inferences. To learn from success, partici-
pants must remember the correct answer they were
presented with, but to learn from failure, participants
must infer the correct answer from the incorrect one.
Accordingly, in Study 2c, we modified the test-phase
questions (Round 2). In this new test phase, participants
were shown the initial questions from Round 1 and
were instructed to select the incorrect response to each
of the questions. On this modified test, participants in
the failure condition had to reselect their initial answer
choices to answer correctly, whereas participants in the
success condition had to infer the incorrect response
on the basis of their initial correct response. Thus,
failure-condition participants had to make fewer mental
inferences than their success-condition counterparts to
answer the test correctly.
Finally, in Study 2d, we tested whether participants
would continue to learn less from failure than success
when the content was social in nature. People are better
at cognitive-reasoning skills when content is social
(Cosmides, 1989). In Study 2d, participants completed
a relationship game that was structurally similar to the
script task (it featured a learning round with feedback,
followed by a test), but the script symbols were replaced
with social content. Each question in the relationship
game asked, “Which of the following two couples are
engaged?” Participants had to choose from one of two
couples, following which they received feedback.
Depending on condition, participants were randomly
assigned to receive success feedback (“You are cor-
rect”) or failure feedback (“You are incorrect”) on each
of the three questions. After the learning phase (three
questions in Round 1), participants completed a short
distractor activity that asked them to reflect on one of

6 Eskreis-Winkler, Fishbach
their favorite couples. Following this, they completed
a test that assessed whether they had learned which
couple was engaged.
Results
Study 2a. Learning was operationalized as the percent-
age of Round 2 questions the participant answered cor-
rectly (out of three). The results supported our hypothesis:
Participants in the failure condition (M = 59%, SD = 41%)
learned less—that is, had fewer correct answers—than
participants in the success condition (M = 80%, SD =
35%), t(97) = 2.80, p = .006, d = 0.55, 95% CI = [0.15,
0.95].
Another question is whether participants in either
condition learned anything at all—that is, whether their
performance exceeded chance level. Participants in the
success condition learned at a level that exceeded chance
(i.e., 1.5 out of 3), t(49) = 6.06, p < .001. By contrast,
learning did not exceed chance in the failure condition,
t(48) = 1.45, p = .154 (see Table 1 and Fig. 1).
Study 2a replication. As in Study 2a, participants in
the failure condition (M = 66%, SD = 36%) learned less
than participants in the success condition (M = 88%,
SD = 25%), t(323) = 6.17, p < .001, d = 0.71, 95% CI =
[0.49, 0.94]. Participants in the success condition learned
at a level that exceeded chance, t(165) = 19.14, p < .001,
as did participants in the failure condition, t(158) = 5.82,
p < .001 (see Table 1 and Fig. 1).
Study 2b. The main effect was replicated: Participants
in the failure condition (M = 77%, SD = 31%) learned
less—they had fewer correct answers on the test—than
participants in the success condition (M = 90%, SD =
22%), t(100) = 2.57, p = .012, d = 0.51, 95% CI = [0.11,
0.90]. In the success condition, participants learned at a
level that exceeded chance (1.5 out of 3), t(51) = 13.10,
p < .001, and the same was true in the failure condition,
t(49) = 6.07, p < .001 (see Table 1 and Fig. 1). It appears
that with a higher financial incentive, participants still
learned from failure, albeit less than from success.
Study 2c. Again, the main effect was replicated: Partici-
pants in the failure condition (M = 51%, SD = 44%)
learned less than participants in the success condition
(M = 81%, SD = 38%), t(112) = 3.86, p < .001, d = 0.72,
95% CI = [0.34, 1.10]. Test scores in the success condition
exceeded chance, t(58) = 6.18, p < .001; in contrast, test
scores in the failure condition did not, t(54) = 0.15, p =
.880 (see Table 1 and Fig. 1). Thus, in Study 2c, partici-
pants learned less from failure than from success, despite
the fact that failure feedback was technically easier to
learn from than success feedback because doing so
required fewer mental inferences.
Study 2d. The main effect was replicated: Participants
in the failure condition (M = 67%, SD = 34%) learned less
than participants in the success condition (M = 91%,
SD = 21%), t(101) = 4.30, p < .001, d = 0.85, 95% CI = [.44,
1.25]. Learning in the success condition exceeded chance
(2.5 out of 5), t(50) = 13.78, p < .001, as did learning in
the failure condition, t(51) = 3.50, p = .001 (see Table 1
and Fig. 1). (Note that the failure condition had notably
more variance than the success condition in this study,
which is likely the result of ceiling effects in the success
condition.)
Study 3: Comparing Failure Feedback
With No Feedback
Studies 1 and 2 compared learning following failure
with learning following success. Consequently, these
studies left open the possibility that success motivates
people to tune in, not that failure motivates people to
tune out. In Study 3, we examined whether people
learn less following failure feedback (failure condition)
compared with an experience that offers no feedback
at all (control condition). Thus, Study 3 tested whether
participants remember fewer answers after failure feed-
back than after no feedback.
Measuring memory for one’s initial answer choices
also allowed us to examine a different dimension of
learning, which, arguably, more closely corresponds to
tuning out. We expected participants who received fail-
ure feedback to have less memory of their initial answer
choices than participants who received no feedback.
Method
Participants. We opened the survey to 100 partici-
pants on MTurk. Individuals of any nationality were
invited to participate as long as their approval rating was
at or above 50%. MTurk returned 100 responses (49%
female; age: M = 37.41 years, SD = 12.32).
Procedure. We followed the procedure outlined in
Study 2a; for example, participants were again incentiv-
ized. However, instead of comparing failure with success,
we compared the failure condition with a condition that
received no feedback. In the test phase (Round 2), par-
ticipants saw the same multiple-choice questions from
Round 1 and had to recall the answers they gave in
Round 1 to each of these questions. Because we thought
a test on three questions might be too easy, we expanded
both Round 1 and Round 2 to include five questions.

Not Learning From Failure 7
Results
Supporting our hypothesis, results showed that partici-
pants in the failure condition (M = 59%, SD = 39%)
remembered fewer of their initial answer choices than
participants who received no feedback (M = 94%,
SD = 16%), t(98) = 5.91, p < .001, d = 1.18, 95% CI =
[0.78, 1.61]. Participants in the failure condition did not
remember their answer choices at a rate above chance
(2.5 out of 5), t(50) = 1.61, p = .113, whereas partici-
pants who received no feedback did remember their
answer choices at a rate above chance, t(48) = 19.61,
p < .001. It appears that failure—more than a no-
feedback experience—led people to tune out. These
results suggest that in prior studies, over and beyond
any effect that success may have had on motivating
people to tune in, failure led people to tune out.
Study 4: Mediation by Ego Threat
Why does failure feedback undermine learning? We
hypothesized that failure hurts the ego, which leads
people to tune out and not learn from the experience.
In Study 4, we asked participants to report on their
self-esteem following feedback. We assumed that self-
reported self-esteem would capture ego threat. We
hypothesized that failure feedback (vs. success feed-
back) would undermine self-esteem and that this would
explain the lower levels of learning.
Method
Participants. We opened the experiment to 300 par-
ticipants on MTurk. We recruited a slightly larger sample
in Study 4, compared with Studies 2 and 3, to ensure that
the study was powered to test for mediation. Individuals
of any nationality were invited to participate as long as
their MTurk approval rating was at or above 50%. MTurk
returned 300 responses (49.3% female; age: M = 35.77
years, SD = 10.72).
Procedure. The procedure was the same as in Study 2a,
with one exception: After the learning phase (Round 1),
we inserted a self-report question that asked participants
to report the degree to which the task undermined their
self-esteem (“To what extent would you say that complet-
ing Round 1 undermined your self-esteem?”; 1 = not at all,
5 = very much). Following this question, participants com-
pleted the distractor activity and the test from Study 2a.
Results
In support of our main hypothesis, results showed that
participants in the failure condition (M = 68%, SD =
36%) learned less—they had fewer correct answers—
than participants in the success condition (M = 88%,
SD = 27%), t(298) = 5.65, p < .001, d = 0.65, 95% CI =
[0.42, 0.88]. Learning in the success condition exceeded
chance (2.5 out of 5), t(152) = 17.41, p < .001, as did
learning in the failure condition, t(146) = 5.96, p < .001
(see Table 1 and Fig. 1).
Next, we tested for ego threat—whether participants
in the failure condition reported lower levels of self-
esteem than participants in the success condition. In
support of our hypothesis, results showed that partici-
pants in the failure condition felt that the task had
undermined their self-esteem (M = 3.22, SD = 1.22)
more than participants in the success condition did
(M = 1.70, SD = 1.15), t(298) = 11.15, p < .001, d = 1.29,
95% CI = [1.04, 1.54].
Finally, we tested whether ego threat mediated the
effect of failure on learning. Supporting our hypothesis,
results showed that the effect of failure on learning was
significantly reduced when ego threat was added to the
model, t(297) = 3.24, p = .001. Ego threat mediated
the indirect effect of condition on learning, β = −0.20,
SE = 0.07, 95% CI = [–0.37, –0.03] in an analysis based
on 10,000 bootstrap samples. In sum, participants who
received failure feedback were significantly more likely
than participants who received success feedback to feel
that their self-esteem had been compromised. The
sense that failure was ego threatening in turn under-
mined learning.
One limitation of Study 4 is that we relied on self-
reports of ego threat. It is possible that participants did not
have the insight to report their true reactions, even when
we translated “ego threat” into the more familiar concept
of “self-esteem.” To address this, in Study 5, we tested
whether people’s ability to learn from failure improved
when ego threat was removed experimentally.
Study 5: Moderation by Ego Threat—
Learning From Other People’s Failures
In Study 5, we examined whether people learn from
failure when ego threat is removed. We did this by
comparing learning following personal successes and
failures to vicarious learning following other people’s
successes and failures.
Method
Participants. We opened the experiment to 400 par-
ticipants. Again, individuals of any nationality were
invited to participate as long as their MTurk approval
rating was at or above 50%. MTurk returned 402 respon-
dents (52.5% female; age: M = 36.21 years, SD = 11.12).

8 Eskreis-Winkler, Fishbach
Procedure. This study used a 2 (feedback: success vs.
failure; within participants) × 2 (perspective: self vs.
other; between participants) mixed design. Specifically,
each participant received failure feedback on one set of
three questions and success feedback on another set of
three questions, in counterbalanced order. Unlike in prior
studies, where success and failure were between partici-
pants, participants in the current study experienced both
success and failure, albeit in separate question sets.
The script task shown to the self condition was iden-
tical to the script task used in Study 2a. In the learning
phase, participants answered script questions, following
which they received feedback on their answer choices
(Round 1). In the test phase (Round 2), we measured
learning.
In contrast, in the other condition, the script task
showed someone else’s performance. Prior to each set,
participants in the other condition read, “In this set, you
will see how someone else answered three questions
and get feedback on this other person’s answers.” Par-
ticipants in the other condition then clicked through
Round 1 questions and saw the answer choices someone
else had selected. After each answer choice, the observ-
ing participant had to reselect the answer choice, which
made vicarious learning more active. Following this, the
participant received feedback on the other person’s
answer choice (success or failure feedback, depending
on condition). He or she then completed the same test
(Round 2) as participants in the self condition.
Results
A 2 (feedback) × 2 (perspective) analysis of variance
revealed no main effect of perspective, F(1, 400) = 3.17,
p = .076, and a main effect of feedback, F(1, 400) = 17.75,
p < .001. Participants learned more from success than
from failure. In support of our hypothesis, results showed
a Feedback × Perspective interaction, F(1, 400) = 8.63,
p = .004. Replicating the effect from prior studies,
results also showed that participants learned less from
personal failures (M = 69%, SD = 38%) than from per-
sonal successes (M = 83%, SD = 33%), F(1, 400) = 25.68,
p < .001. However, participants learned just as much
from other people’s failures (M = 80%, SD = 32%) as other
people’s successes (M = 82%, SD = 31%), F(1, 400) = 0.81,
p = .369.
We also calculated simple contrasts for each feedback
condition. Participants learned significantly more from
others’ failures than their own failures, F(1, 400) = 9.23,
p = .003, but learned the same amount from personal
successes and others’ successes, F(1, 400) = 0.10, p =
.752. Learning was above chance level in all four cells—
self-success: t(201) = 14.45, p < .001; self-failure: t(201) =
7.28, p < .001; other-success: t(199) = 14.94, p < .001;
and other-failure: t(199) = 13.39, p < .001 (see Table 1
and Fig. 1). In sum, the more failure is dissociated from
the self, the less people tune out, and the more they
learn from failure.
We chose vicarious learning as a moderator because
this moderator eliminates ego threat. That said, vicari-
ous learning can have other effects as well—for exam-
ple, it can lead people to adopt the other person’s
perspective (vs. a self-perspective; Grossmann & Kross,
2014; Libby & Eibach, 2011; Pronin, Gilovich, & Ross,
2004), it can decrease overall involvement in a task
(Bertsch, Pesta, Wiscott, & McDaniel, 2007), or it can
prompt more abstract processing (Trope & Liberman,
2010). Nevertheless, it is difficult to see how these
alternative processes could account for the results. Spe-
cifically, because perspective taking, decreasing overall
task involvement, and prompting more abstract pro-
cessing are less likely to differentially affect people’s
ability to learn from failure (vs. success), we conclude
that vicarious learning eliminated ego threat, thus
increasing people’s ability to learn from failure.
General Discussion
To paraphrase the celebrated political theorist Antonio
Gramsci, history teaches, but it has no pupils (Gramsci,
1977). We found that something similar happens with
failure. Across five studies, participants learned less
from failure feedback than from success feedback—
even when both types of feedback contained full infor-
mation on the correct answer. Failure feedback
undermined learning motivation because it was ego
threatening: It caused participants to tune out and stop
processing information.
Our findings advance motivation theory and, in par-
ticular, past theoretical and empirical work that argues
that negative feedback undermines goal commitment
(Atkinson, 1964; Bandura & Cervone, 1983; Cochran &
Tesser, 1996; Fishbach & Finkelstein, 2012; Hattie &
Timperley, 2007; Lewin, 1935; Soman & Cheema, 2004;
Weiner, 1974; Yeager & Dweck, 2012; Zajonc &
Brickman, 1969). Complementing this past work, which
describes how failure affects motivation in the future,
our studies explored how failure feedback affects moti-
vation in the present—the moment of failure itself. Our
key result is that people find failure feedback ego
threatening, which leads them to tune out and miss the
information the feedback offers. In other words, failure
undermines learning. It is possible that these immediate
effects underlie the longer-term demotivating effects of
failure on goal commitment. Tuning out from a pursuit
in the moment of failure could be the first step in a
chain reaction that distances and discourages people
from the goal they are pursuing.

Not Learning From Failure 9
It is possible that this tune-out reaction depends on
the size of the failure. In the well-documented phenom-
enon of aversion learning, animals that taste poison,
receive shocks, or experience other “large” failures
learn to avoid these threats in the future (Garcia, Lasiter,
Bermudez, & Deems, 1985). It is possible that for large
failures, the attentional pull of the negative experience
overrides the motivation to tune out. Nevertheless,
there are many large failures—for example, failures in
close relationships—that people might be bad at detect-
ing over long periods of time, despite their size and
importance. Similarly, many failures are small (e.g., a
failed experiment that a researcher discards as nonin-
formative), yet they accumulate a significant amount of
information that people might fail to learn from.
We found that people struggle to learn from failure
feedback in the field, using a task that presented employ-
ees with relevant professional information, and in online
samples, using tasks involving language and relationship
stimuli. We found the effect in both the United States
and in the United Kingdom, though these cultures are
admittedly similar. It is still an open question whether
people’s failure to learn from their mistakes would gen-
eralize to individuals in other cultures—and in particular,
to individuals in cultures that have different attitudes to
failure. For example, Japanese individuals persist longer
after they fail than after they succeed, whereas Ameri-
cans do the opposite (Heine et al., 2001). Thus, it is
unclear whether Japanese individuals, like the American
and British participants in our samples, would learn less
from failure than from success. Another open question
is whether certain failures are more easily taught than
others. Future research is needed to determine whether
the observed effect generalizes when feedback is more
personalized, more detailed, or delivered in a different
way (i.e., by a caring mentor). We have no reason to
believe that the results depend on other characteristics
of the participants, materials, or context.
Our results have practical implications. People who
want to learn may be better able to do so via successful
experiences than via unsuccessful experiences. When
failure feedback is inevitable, our results suggest that
people will learn more if failure feedback can be sepa-
rated from the ego. No matter the precise method for
reducing ego involvement—for example, positioning
people as vicarious learners or instructing people to
reappraise feedback in less threatening terms—our
results suggest that reducing the degree to which failure
involves the ego will promote learning.
Appendix
This appendix contains the text for the manipulations
used in Study 2a and the Study 2a replication. The
manipulations for the other studies are posted at https://
osf.io/5kbx6.
Failure condition Success condition
Welcome! Today you will answer some language questions about a
researcher-manufactured ancient script.
When you click to the next page, you will begin Round 1.
Whether you get the answers right or wrong in Round 1, try to learn. Round 2 will test
how much you learned and bonus you $0.10 for each question you get correct!
YOU ANSWERED THIS QUESTION INCORRECT! YOU ANSWERED THIS QUESTION CORRECT!
YOU ANSWERED THIS QUESTION INCORRECT! YOU ANSWERED THIS QUESTION CORRECT!

10 Eskreis-Winkler, Fishbach
Action Editor
Michael Inzlicht served as action editor for this article.
Author Contributions
Both authors developed the study concept and study designs.
L. Eskreis-Winkler collected and analyzed the data. L. Eskreis-
Winkler drafted the manuscript; A. Fishbach provided critical
revisions. Both authors approved the final version of the
manuscript for submission.
ORCID iD
Lauren Eskreis-Winkler https://orcid.org/0000-0003-1977-5677
Acknowledgments
We are grateful to the participants and to Carman Fowler,
who made this work possible. We thank Shane Frederick and
lab members for indispensable feedback. L. Eskreis-Winkler
extends special thanks to Ari Lustig, who learns from all his
successes—he has no failures.
Declaration of Conflicting Interests
The author(s) declared that there were no conflicts of interest
with respect to the authorship or the publication of this
article.
Supplemental Material
Additional supporting information can be found at http://
journals.sagepub.com/doi/suppl/10.1177/0956797619881133
Open Practices
All data and materials have been made publicly available via
the Open Science Framework and can be accessed at https://
Failure condition Success condition
YOU ANSWERED THIS QUESTION INCORRECT! YOU ANSWERED THIS QUESTION CORRECT!
Thanks for answering those initial questions. Right now take a short breather. Tell us:
what is your favorite music to listen to?
This is Round 2.
Based on what you learned about the researcher-manufactured ancient script in Round 1,
answer the final three questions below. For each question you get correct, you will earn
bonus cash ($0.10).

Not Learning From Failure 11
osf.io/5kbx6. The design and analysis plans for the Study 2a
replication were preregistered at http://aspredicted.org/blind
.php?x=it2ej3. The complete Open Practices Disclosure for
this article can be found at http://journals.sagepub.com/doi/
suppl/10.1177/0956797619881133. This article has received the
badges for Open Data, Open Materials, and Preregistration. More
information about the Open Practices badges can be found at
http://www.psychologicalscience.org/publications/badges.
Notes
1. We began the survey with this question in all online panel
studies reported in this article (Studies 2–5), and we randomly
assigned participants to a condition only if they provided an
answer.
2. Nevertheless, results were consistent when feedback was
presented on the same page, alongside the answer choices. See
Study S4 in the Supplemental Material.
References
Atkinson, J. W. (1964). An introduction to motivation. Oxford,
England: Van Nostrand.
Bandura, A., & Cervone, D. (1983). Self-evaluative and self-
efficacy mechanisms governing the motivational effects of
goal systems. Journal of Personality and Social Psychology,
45, 1017–1028.
Baumeister, R. F., Bratslavsky, E., Finkenauer, C., & Vohs, K. D.
(2001). Bad is stronger than good. Review of General Psy-
chology, 5, 323–370.
Bertsch, S., Pesta, B. J., Wiscott, R., & McDaniel, M. A. (2007).
The generation effect: A meta-analytic review. Memory &
Cognition, 35, 201–210.
Bless, H., & Fiedler, K. (2006). Mood and the regulation of
information processing and behavior. In J. P. Forgas (Ed.),
Affect in social thinking and behavior (pp. 65–84). New
York, NY: Psychology Press.
Brown, J. (2014). 27 inspirational Bill Gates quotes. Retrieved
from https://addicted2success.com/quotes/27-inspira
tional-bill-gates-quotes/
Cochran, W., & Tesser, A. (1996). The “what the hell” effect:
Some effects of goal proximity and goal framing on per-
formance. In L. L. Martin & A. Tesser (Eds.), Striving
and feeling: Interactions among goals, affect, and self-
regulation (pp. 99–120). Hillsdale, NJ: Erlbaum.
Cosmides, L. (1989). The logic of social exchange: Has natural
selection shaped how humans reason? Studies with the
Wason selection task. Cognition, 31, 187–276.
Faul, F., Erdfelder, E., Lang, A.-G., & Buchner, A. (2007).
G*Power 3: A flexible statistical power analysis program for
the social, behavioral, and biomedical sciences. Behavior
Research Methods, 39, 175–191.
Fishbach, A., & Finkelstein, S. R. (2012). How feedback influ-
ences persistence, disengagement, and change in goal
pursuit. In H. Aarts & A. J. Elliot (Eds.), Frontiers of social
psychology: Goal-directed behavior (pp. 203–230). New
York, NY: Psychology Press.
Garcia, J., Lasiter, P. S., Bermudez, F., & Deems, D. A. (1985).
A general theory of aversion learning. Annals of the New
York Academy of Sciences, 443, 8–21.
Gramsci, A. (1977). Selections from political writings (1910–
1920) (Q. Hoare, Ed., J. Mathews, Trans.). London,
England: Lawrence and Wishart.
Grossmann, I., & Kross, E. (2014). Exploring Solomon’s para-
dox: Self-distancing eliminates the self-other asymmetry
in wise reasoning about close relationships in younger
and older adults. Psychological Science, 25, 1571–1580.
Hattie, J., & Timperley, H. (2007). The power of feedback.
Review of Educational Research, 77, 81–112.
Heine, S. J., Kitayama, S., Lehman, D. R., Takata, T., Ide,
E., Leung, C., & Matsumoto, H. (2001). Divergent con-
sequences of success and failure in Japan and North
America: An investigation of self-improving motivations
and malleable selves. Journal of Personality and Social
Psychology, 81, 599–615.
Help Scout. (2018). 75 customer service facts, quotes &
statistics. Retrieved from https://www.helpscout.com/75-
customer-service-facts-quotes-statistics/
Lewin, K. (1935). A dynamic theory of personality. New York,
NY: McGraw-Hill.
Libby, L. K., & Eibach, R. P. (2011). Visual perspective in men-
tal imagery: A representational tool that functions in judg-
ment, emotion, and self-insight. In J. M. Olson & M. P.
Zanna (Eds.), Advances in experimental social psychology
(Vol. 44, pp. 185–245). San Diego, CA: Academic Press.
Louro, M. J., Pieters, R., & Zeelenberg, M. (2007). Dynamics
of multiple-goal pursuit. Journal of Personality and Social
Psychology, 93, 174–193.
Maxwell, J. C. (2007). Failing forward: Turning mistakes
into stepping stones for success. Nashville, TN: Thomas
Nelson.
Ohira, H., Winton, W. M., & Oyama, M. (1998). Effects of stim-
ulus valence on recognition memory and endogenous eye-
blinks: Further evidence for positive-negative asymmetry.
Personality and Social Psychology Bulletin, 24, 986–993.
Öhman, A. (2007). Has evolution primed humans to “beware
the beast”? Proceedings of the National Academy of
Sciences, USA, 104, 16396–16397.
Pratto, F., & John, O. P. (1991). Automatic vigilance: The
attention-grabbing power of negative social information.
Journal of Personality and Social Psychology, 61, 380–391.
Pronin, E., Gilovich, T., & Ross, L. (2004). Objectivity in the
eye of the beholder: Divergent perceptions of bias in self
versus others. Psychological Review, 111, 781–799.
Puig, V. A., & Szpunar, K. K. (2017). The devil is in the
details: Comparisons of episodic simulations of positive
and negative future events. Emotion, 17, 867–873.
Rozin, P., & Royzman, E. B. (2001). Negativity bias, nega-
tivity dominance, and contagion. Personality and Social
Psychology Review, 5, 296–320.
Soman, D., & Cheema, A. (2004). When goals are counterproduc-
tive: The effects of violation of a behavioral goal on subse-
quent performance. Journal of Consumer Research, 31, 52–62.

12 Eskreis-Winkler, Fishbach
Taylor, S. E. (1991). Asymmetrical effects of positive and
negative events: The mobilization-minimization hypoth-
esis. Psychological Bulletin, 110, 67–85.
Trope, Y., & Liberman, N. (2010). Construal-level theory of psy-
chological distance. Psychological Review, 117, 440–463.
Weiner, B. (1974). Achievement motivation and attribution
theory. Morristown, NJ: General Learning Press.
Yeager, D. S., & Dweck, C. S. (2012). Mindsets that promote
resilience: When students believe that personal charac-
teristics can be developed. Educational Psychologist, 47,
302–314.
Zajonc, R. B., & Brickman, P. (1969). Expectancy and feed-
back as independent factors in task performance. Journal
of Personality and Social Psychology, 11, 148–156.