Unlearning implicit social biases during sleep

See allHide authors and affiliations

Science  29 May 2015:
Vol. 348, Issue 6238, pp. 1013-1015
DOI: 10.1126/science.aaa3841

Sleep on it: Consolidating implicit learning

A good night's sleep is one of the best ways to fix recently learned information into long-lasting memory. Recent evidence suggests that recent memories are reactivated during sleep and woven into existing representations of stored information. Hu et al. now demonstrate that triggering memory consolidation during sleep can help set into place recently learned anti-bias training (see the Perspective by Feld and Born). Changes in people's stereotypical attitudes toward race and gender were maintained for up to 1 week after training.

Science, this issue p. 1013; see also p. 971


Although people may endorse egalitarianism and tolerance, social biases can remain operative and drive harmful actions in an unconscious manner. Here, we investigated training to reduce implicit racial and gender bias. Forty participants processed counterstereotype information paired with one sound for each type of bias. Biases were reduced immediately after training. During subsequent slow-wave sleep, one sound was unobtrusively presented to each participant, repeatedly, to reactivate one type of training. Corresponding bias reductions were fortified in comparison with the social bias not externally reactivated during sleep. This advantage remained 1 week later, the magnitude of which was associated with time in slow-wave and rapid-eye-movement sleep after training. We conclude that memory reactivation during sleep enhances counterstereotype training and that maintaining a bias reduction is sleep-dependent.

Social interactions are often fraught with bias. Our preconceptions about other people can influence many types of behavior. For example, documented policing errors have repeatedly shown the potential harm of racial profiling (1). In experiments that used a first-person-shooter videogame, both White and Black participants were more likely to shoot Black than White individuals, even when they held a harmless object rather than a gun (2). When hiring potential research assistants, both male and female faculty members were more likely to hire male than equally qualified female candidates (3).

Although the tendency for people to endorse racist or sexist attitudes explicitly has decreased in recent years (4), social biases may nevertheless influence people’s behavior in an implicit or unconscious manner, regardless of their intentions or efforts to avoid bias (5). Ample evidence indicates that implicit biases can drive discriminatory behaviors and exacerbate intergroup conflict (58). For instance, implicit racial biases decrease investments given to racial out-group members in a trust game (6). At a broader level, the gender gap in science achievement in a nation is correlated with the level of implicit stereotyping of females as not having an aptitude for science (8).

Whereas discriminatory behaviors can be detrimental to individuals and society, implicit social bias can be difficult to correct because of a range of affective, cognitive, motivational, and social factors, as follows (9, 10). First, out-group members can be perceived as threatening, and the fear response to those individuals can resist extinction (11). Second, biases are acquired over many years of exposure to stereotypes, and they can efficiently operate without occupying cognitive resources (5, 10). Third, motivation to seek higher status or self-enhancement commonly results in out-group derogation (9, 10). Last, perceived social norms can prescribe people’s expression of stereotyping and prejudice (12). Despite such challenges, implicit biases can be reduced by learning about counterstereotype cases (13). However, benefits of this counterbias training can be fragile, subject to reversal when the original stereotypes are again reinforced in typical circumstances, such as through the media (14). Longer-term reductions in implicit social biases may necessitate that counterbias training be followed by further memory consolidation, as is the case for many other types of learning (15).

Recent findings suggest that memory consolidation during sleep may be essential for preserving newly acquired information, such as declarative and procedural memories (1519). During sleep, information recently stored in the brain can be integrated with other information and transformed into stable representations through a process known as systems-level consolidation (15). The mechanisms of this transformation are thought to involve repeated reactivation of information, particularly during sleep, leading to subsequent improvement in postsleep memory performance (1924).

Taking into consideration the role of sleep in memory consolidation, we adapted procedures for (i) reducing implicit social biases and (ii) reactivating this training during sleep. We were particularly interested in factors that can influence whether such training procedures produce transient or persistent effects. Because pervasive stereotypes in the media and broader culture could function to regenerate a bias that is momentarily reduced (14), maintaining the benefits of training is crucial for the ultimate usefulness of potential bias-reducing interventions.

We reactivated counterbias information during sleep using subtle auditory cues that had been associated with counterbias training. Participants were White males and females from a university community (N = 40) and were recruited as two subsamples that allowed for a direct replication (25). First, biases were quantified using two versions of the implicit association test (IAT) (26). The IAT allows for an assessment of the strength of implicit associations between social groups and attributes (26). One test examined the degree to which female faces were preferentially associated with art versus science words—or the reverse for male faces (gender-bias IAT). The other test examined the degree to which Black faces were preferentially associated with bad versus good—words or the reverse for White faces (racial-bias IAT). Results were quantified by using a conventional scoring procedure (27), in which zero indicates no bias and larger scores indicate greater bias. Consistent with previous research (7), IAT scores showed that participants held implicit social biases for both gender and race, with both scores significantly greater than zero [mean ± SEM, 0.559 ± 0.044; gender t(39) = 9.076, P < 0.001; race t(39) = 8.388, P < 0.001].

After this confirmation of baseline levels of implicit bias, participants engaged in training designed to reduce gender and racial bias (13). In both cases, bias reduction was expected because participants intentionally selected counterstereotype information intermixed with other information. Participants viewed several types of face-word pairing but were required to attend and respond only to pairings that countered the typical bias (Fig. 1A) (25). Two unusual frequency-modulated sounds were presented during training, one after correct counter–gender bias responses and the other after correct counter–racial bias responses. To reinforce these associations, we administered another task wherein the same two sounds prompted participants to form a corresponding face-word pairing (25). Training thus established a strong association between each sound and one type of counterbias training.

Fig. 1 Experimental procedures and results.

(A) Procedures for counterbias training with sound cues. (B) Implicit bias reduction was found for both counter–racial bias training and counter–gender stereotype training. Bias was measured using the IAT before training (baseline) and after training (prenap). Error bars indicate ±1 SEM adjusted for within-subject comparisons. (C) Procedures for the nap phase of the experiment, when one sound was repeatedly played to participants during SWS, using a low intensity to avoid arousal from sleep. (D) The change in implicit bias from prenap to postnap diverged as a function of cueing condition, with a further reduction only for the cued social bias. (E) The change in implicit bias from prenap to the 1-week delay diverged as a function of cueing condition, with a significant increase only for the uncued social bias. (F) The change in implicit bias from baseline to the 1-week delay diverged as a function of cueing condition, with a significant reduction only for the cued social bias. Significant pairwise differences are indicated: *P < 0.05 or **P ≤ 0.01.

Biases were reduced compared with baseline levels (Fig. 1B) [within-subject analysis of variance (ANOVA), F1,39 = 15.453, P < 0.001, ηp2 = 0.284]. The mean IAT score was 0.559 at baseline and 0.335 at the prenap test. This bias reduction did not differ as a function of bias type (F1,39 = 1.840, P = 0.183).

Next, participants were invited to take a 90-min afternoon nap (Fig. 1C; see table S1 for sleep-stage information). When electroencephalographic signals showed clear signs of slow-wave sleep (SWS), we repeatedly played one auditory cue, randomly selected as the counter–gender bias sound (n = 21) or the counter–racial bias sound (n = 19). Stimulation was discontinued at any sign of arousal from sleep. The number of presentations averaged 258 ± 24 (SEM).

Implicit biases were measured again after waking. Bias change from prenap to postnap varied with cueing condition as predicted [substantiated by a two-way interaction (cued or uncued by prenap or postnap), F1,39 = 14.612, P < 0.001, ηp2 = 0.273]. As shown in Fig. 1D, implicit bias was significantly reduced from prenap to postnap when cued [t(39) = 2.698, P = 0.010] and unchanged when not cued [t(39) = –1.378, P = 0.176] (fig. S1 and S2). This differential bias reduction was not moderated by bias type (fig. S3). Thus, reactivating counterbias learning during sleep can selectively reduce implicit racial or gender bias, depending on which form of counterbias training was cued.

Implicit biases were measured again after 1 week, revealing that the differential bias reduction endured (Fig. 1E) (n = 38; F1,37 = 4.672, P = 0.037, ηp2 = 0.112). Cueing during sleep resulted in sustained counterbias reduction, such that the cued bias did not differ between prenap and delayed testing [t(37) = –0.774, P = 0.444], whereas the uncued bias increased during the delay [t(37) = –3.078, P = 0.004]. When compared with baseline (Fig. 1F), cued biases were weaker after 1 week [t(37) = 2.203, P = 0.034], whereas uncued biases were not [t(37) = 0.524, P = 0.603], although the interaction was not significant (F1,37 = 0.471, P = 0.497).

Neurophysiological activity during sleep—such as sleep spindles, slow waves, and rapid-eye-movement (REM) duration—can predict later memory performance (17). Accordingly, we explored possible relations between cueing-specific bias reduction and measures of sleep physiology. We found that only SWS × REM sleep duration consistently predicted cueing-specific bias reduction at 1 week relative to baseline (Fig. 2) [r(38) = 0.450, P = 0.005] (25).

Fig. 2 Relation between long-term bias reduction and sleep physiology.

The quality of sleep after training as indexed by the product of (minutes in SWS) × (minutes in REM sleep), predicted differential bias change, quantified as follows. Given that standardized implicit bias scores were preferentially reduced for the cued relative to uncued condition overall (Fig. 1F), we computed the reduction separately for cued and uncued conditions [(baseline score) – (delayed score) in both cases]. Differential bias change was taken as the cued reduction minus the uncued reduction, such that higher values indicated larger bias reduction over this interval for the cued compared to the uncued bias.

Past research indicates that by pairing learning episodes with auditory or olfactory stimuli and then presenting these stimuli again during postlearning SWS, learned information can be specifically reactivated and strengthened (19). Benefits of this targeted memory reactivation (TMR) have been documented for declarative, procedural, and emotional memories (19). Such learning typically does not challenge preexisting knowledge nor compete with daily experiences outside the laboratory. In contrast, we examined learning-induced changes in long-standing social biases. We showed that selectively reactivating counterbias learning during sleep weakened preexisting implicit social biases immediately after the nap and facilitated the retention of this learning going forward. Without TMR during sleep, training effects tended to dissipate, and the bias returned to baseline levels. These results thus enlarge our conception of sleep’s role in socially relevant learning.

Observed relations between sleep neurophysiology and behavior further reinforced the conclusion that bias reduction is sleep-dependent. Current thinking about consolidation emphasizes sets of cortical networks that can become integrated through interactions with hippocampal networks, possibly by means of cyclic SWS-REM periods (1517, 28). The correlation with SWS × REM duration implicates a benefit from REM-based processing subsequent to SWS-based reactivation, perhaps to integrate learning within associative knowledge networks. These findings support the notion that both SWS and REM are operative in sleep-dependent memory consolidation (16, 17, 28, 29).

Future research is needed to address many outstanding questions in relation to our findings. For example, how much training is needed to make implicit benefits persist for long periods of time and transfer to explicit benefits in interpersonal interactions? To what extent do persistent benefits depend on repeated training, the nature of other waking activities after training, and repeated memory reactivation during sleep? Although IAT measures are imperfect and may sometimes reflect knowledge of cultural stereotypes rather than implicit bias per se (30), prior research has demonstrated consequences for social behavior, such that low implicit bias as measured with the IAT may indeed be linked with egalitarianism (6, 7). Given that training to reduce implicit bias can be conceptualized as a type of habit learning (31), perhaps novel sleep manipulations could be adapted to aid people in changing various unwanted or maladaptive habits, such as smoking, unhealthy eating, catastrophizing, or selfishness (32).

Supplementary Materials

Materials and Methods

Supplementary Text

Figs. S1 to S3

Table S1

Sound Files

References (3336)


  1. Materials and methods are available as supplementary materials on Science online.
  2. Acknowledgments: Support was provided by Northwestern University and by NSF grants BCS-1025697, BCS-1461088, and DGE-0824162, and by NIH grants T32-AG020418 and F31-MH100958. Data are archived and available on the Open Science Framework ( K.A.P. served as an advisor to Sheepdog Sciences (a company developing educational technologies related to sleep) in exchange for equity compensation. The authors declare no conflict of interests.
View Abstract

Stay Connected to Science


Navigate This Article