Do anti-discrimination policies work? Updated

Elevator pitch

Labour market discrimination is widespread and raises ethical, societal, and efficiency concerns. It not only results in the unfair treatment of individuals with comparable skills, but also imposes broader costs on society by eroding trust and weakening cohesion. Moreover, discrimination limits the full potential of the working-age population by excluding talent or trapping people in roles below their abilities. These effects are amplified by feedback loops: fewer opportunities lower labour market participation and productivity, while the harm discrimination inflicts on mental and physical health further reduces economic output.

Key findings

Author's main message

To combat labour market discrimination, governments should adopt a multi-pronged strategy. In addition to passing anti-discrimination laws, they should invest in data collection to monitor employers’ hiring and career management practices, and thus create incentives for change among discriminatory employers. Yet incentives alone are not enough. Employers must also be equipped with evidence on what works to reduce statistical discrimination and prevent (un)conscious bias from influencing recruitment and career management decisions, including through the responsible use of AI. Statistical discrimination can be further addressed through policies that mitigate concerns about potential productivity constraints affecting certain groups. Meanwhile, affirmative action policies should be approached with caution, as they may inadvertently reinforce the very drivers of discrimination they seek to dismantle.

Motivation

According to the 2023 Eurobarometer survey on Discrimination in the European Union, discrimination on the grounds of race and ethnicity is perceived as the most widespread form of discrimination, with 60% of respondents across the EU considering it frequent (Figure 1). This is followed closely by discrimination based on sexual orientation, gender identity or sex characteristics (SOGISC), which a majority of respondents (53%) also view as widespread. Perceived prevalence is substantial as well for discrimination based on disability (49%), being seen as too young or too old (45%), religion (42%), and being born as a man or a woman (38%). With the exception of religion, perceptions of discrimination across all these grounds have worsened since 2019.

The existence of discrimination on the grounds listed in Figure 1 is firmly established when more objective measures are used, most notably correspondence studies. These field experiments consist of sending to real job vacancies the applications of fictitious yet equally qualified candidates who differ only on one potential ground of discrimination and comparing their chances of being invited to a job interview. Systematic differences in callback rates reveal hiring discrimination.

A plethora of such studies documents, for instance, substantial racial/ethnic discrimination across OECD countries. With identical CVs, white natives are up to twice as likely to be invited to a job interview as white non-white natives, i.e., individuals of Asian, Middle Eastern and North African, or sub-Saharan African background. The picture is even more troubling when experiments extend to audit studies, in which actors attend real interviews, since these audits reveal significant discrimination at later stages of the hiring process. Most concerning, a meta-analysis of correspondence studies conducted since the late 1960s finds no meaningful decline in hiring discrimination against racial/ethnic minorities, suggesting a persistent failure to invest in anti-discrimination policies that work. 

Discussion of pros and cons

Discrimination in the labour market arises when individuals with comparable productivity are treated differently by employers because they belong to different demographic groups, such as those illustrated in Figure 1.

This unequal treatment can result from a rational economic calculation. Employers rarely have full information about a candidate’s productivity since CVs offer limited and unverifiable signals. To minimise hiring risks and maximise expected profit, employers may rely on indirect information, drawing on the average group-level characteristics they observe to make inferences about individual applicants. For example, if racial/ethnic minority candidates are observed to be, on average, more likely to come from lower socio-economic backgrounds, employers may infer these candidates are also more likely to have lower levels of human capital, even when their CVs appear equivalent to those of majority applicants. In doing so, they may systematically screen out minority candidates and favour their majority counterparts holding CVs constant. Although this practice aligns with an employer’s goal of profit maximisation, it comes at a cost to equity and efficiency. It penalises all minority candidates whose individual productivity equals or exceeds that of majority applicants with similar CVs. And because productivity is not uniform within groups but widely distributed around the average, the number of such unfairly excluded candidates can be substantial, even if group-level averages differ. Economists refer to this form of discrimination as “statistical discrimination” because it relies on inter-group comparisons of key statistics, most notably average productivity, drawn from the empirical evidence available to employers.

Labour market discrimination can also stem from bias against outgroups, which combines negative prejudice and stereotypes that reinforce one another. Here, prejudice refers to an immediate, often unconscious emotional reaction of discomfort towards certain individuals simply because they belong to a different group. Stereotypes, in turn, refer to the process of attributing unfavourable traits to all members of an outgroup, resulting in generalisations that exaggerate perceived weaknesses and downplay strengths compared to the ingroup. In this context, stereotypes are distinct from the beliefs underlying statistical discrimination: while stereotypes are ill-grounded because they are driven by prejudice and further reinforce it, statistical discrimination is supposed to be based on empirically observed group differences. Yet, the fact that prejudice and stereotypes prevent employers from recruiting candidates from the outgroup undermines the chances for this bias to fade over time. Without interacting with outgroup employees, employers miss opportunities to realise that their stereotypes are unfounded and to benefit from intergroup contact, which has been shown to reduce prejudice especially in settings such as the workplace where cooperation is in the interest of all parties. Economists refer to this form of discrimination as “taste-based discrimination” as it is driven by “distaste” towards the outgroup.

In theory, employers who engage in taste-based discrimination should be at an economic disadvantage, since they reject outgroup candidates who are as capable as, or more capable than, those they recruit. In a truly competitive market, such inefficiencies should cause biased employers to lose out to their unbiased peers, and eventually disappear. But in practice, markets are rarely competitive enough for this to happen. What’s more, employers may discriminate not only because of their own preferences, but also by anticipating those of colleagues or customers. If some customers refuse to be served by certain workers, or if colleagues are reluctant to work with them, it may become rational for employers not to hire them. Overall, there may simply be too few unbiased actors to drive the “cleansing” effect of competitive labour markets.

Consistent with this surmise, (un)conscious bias is widespread. One cannot rely on surveys to measure it, since people are not necessarily aware of their bias or may be reluctant to admit socially unacceptable views. Instead, tools like the Implicit Association Test (IAT) reveal automatic mental associations that individuals cannot easily fake. For instance, the IAT measures whether people are quicker to associate Black faces with negative words (like unpleasant or lazy) than White ones, and slower to associate them with positive words. When compiled across thousands of EU test-takers on the Project Implicit website, the findings are troubling: seven in ten show a pro-White bias, and one in four a strong one. 

Combating labour market discrimination requires a multi-pronged approach that addresses both its taste-based and statistical components. The following outlines key steps to achieve this objective.

If anti-discrimination laws are a vital start, enforcing them is what truly counts

The Universal Declaration of Human Rights (1948) affirms the right to work along with the right to equal pay without discrimination (Article 23). A review of anti-discrimination legislation across all 193 UN countries shows that many aim to uphold this principle. Despite remaining gaps in both the grounds covered (notably sexual orientation, gender identity, and sex characteristics) and the areas of employment addressed, significant achievements have been made. Nearly all UN member states provide at least some protection from employment discrimination based on sex at birth, and large majorities also do so on the grounds of disability, religion, and race/ethnicity.

Anti-discrimination laws are foundational. One cannot meaningfully combat discrimination if the legal system does not first recognise it as unlawful. Yet even where strong legal protections exist, discrimination often goes unpunished, not because it is rare, but because it is difficult to prove. 

Demonstrating discrimination in career progression, for example, requires showing that one’s professional trajectory was less favourable than that of colleagues who differ with respect to the protected characteristic in question, but are otherwise comparable (meaning they were hired at the same time, in similar roles, with similar levels of performance throughout their careers). Such evidence is typically not accessible to individuals who suspect they have been discriminated against, and in many cases, it is not systematically collected or retained in a way that would allow it to be examined at all. Showing discrimination in hiring is even more challenging. Candidates typically do not observe the profiles of others considered for the job. As a result, even when discrimination is plausible, there is often no direct proof to substantiate a claim.

In this setting, passing anti-discrimination laws is only the first step. What ultimately matters is enforcing them by easing the burden of proving discrimination. Without this, under-reporting remains the norm and the law stays largely symbolic.
 
Poor enforcement can make anti-discrimination laws not only ineffective but also potentially counterproductive, as it may discourage the hiring of the very groups they aim to protect. It is generally easier to prove discrimination in career progression or dismissal (where comparisons with colleagues, though difficult, are possible) than in recruitment, where rejected candidates have no information about who else applied. As a result, anti-discrimination laws can effectively raise the perceived cost of dismissing workers from protected groups (because employers may fear or face unfair dismissal claims), without creating a comparable risk of sanction for discriminating at the hiring stage. This imbalance can inadvertently create incentives for some employers to avoid hiring protected groups in the first place. 

This asymmetry helps explain why empirical findings on the employment effects of anti-discrimination laws prohibiting discrimination in hiring and dismissal remain mixed. By contrast, once individuals are hired (thus when information is richer and enforcement more effective), anti-discrimination provisions targeting pay and promotion tend to deliver clearer gains, leading to improved relative labour earnings, at least for protected groups who suffer from wage penalties in the first place, including women, LGB people, and racial and ethnic minorities – inconclusive effect was found for the relative labour earnings of people with disabilities, however.

Most of this evidence comes from the US, where the federal structure has allowed researchers to exploit cross-state variation in the timing of anti-discrimination laws using difference-in-differences designs, prior to the adoption of nationwide protections. This approach has been used to assess the employment and labour earnings effects of prohibiting discrimination for various groups (though for women only labour-earnings effects can be examined, as early state-level sex anti-discrimination laws focused exclusively on pay), including based on race/ethnicity (prior to Title VII of the Civil Rights Act of 1964), age (before the Age Discrimination in Employment Act of 1967), disability (prior to the Americans with Disabilities Act of 1990), and sexual orientation and/or gender identity (before the 2020 Bostock v. Clayton County decision). 

While exceptions exist for each ground listed, a large body of evidence points to the following patterns when it comes to the impact of prohibiting discrimination in hiring and dismissal on the employment of protected groups: disability anti-discrimination laws have most often been associated with lower employment among people with disabilities; race/ethnicity anti-discrimination laws have yielded inconclusive employment effects for racial/ethnic minorities; and age anti-discrimination laws have generally been found to increase employment among older workers. The mostly negative employment effects of disability anti-discrimination laws likely arise because their potential unintended consequences are exacerbated by reasonable accommodations required by these laws (e.g. workplace adjustments such as wheelchair access), which can further increase the cost of hiring persons with disabilities. By contrast, the largely positive employment effects of age anti-discrimination laws likely reflect the fact that unintended consequences associated with anti-discrimination legislation are less prevalent when it comes to age. Even though these laws raise termination costs, those costs may carry limited weight in employers’ hiring decisions, as many older workers may not expect to remain in a given job for long (except, as evidence shows, during downturns, when employers may wish to dismiss workers before they decide to leave voluntarily). 

Research on the impact of sexual orientation anti-discrimination laws stands out for its scarcity, with only one paper examining these effects over the entire period preceding the 2020 Bostock v. Clayton County decision. This study finds an overall positive impact on the employment and labour earnings of LGB individuals, reflecting not only direct effects on labour-market outcomes but also improvements in public attitudes towards LGB people, as these laws seem to be viewed as signals of a positive shift in social acceptance to which individuals are eager to conform. 

In a context where enforcement is key to reaping the full benefits of anti-discrimination laws, one primary step taken by several countries to ease the burden of proving discrimination has been the introduction of burden-shifting frameworks in discrimination claims in civil and administrative proceedings. A second key step has been the establishment of institutions with a dedicated mandate to enforce anti-discrimination law. In many countries, particularly EU member States, this has taken the form of creating specialised equality bodies. In others, it has involved expanding the mandate of existing National Human Rights Institutions (NHRIs), encouraged globally by the 1993 Paris Principles, so they also act as enforcers of anti-discrimination legislation. 

Holding employers accountable is critical to enhance enforcement

Anti-discrimination laws and their enforcement play a crucial preventive role by signalling that discrimination is unlawful and punishable. But legal deterrence alone is not enough: even with stronger enforcement mechanisms, proving discrimination will remain challenging for many people and in many situations. This is why additional tools are needed to further foster employer compliance, with systematic monitoring of employer practices emerging as an urgent priority for action.

Monitoring employer career management behaviour

In recent years, monitoring employer career management behaviour has become increasingly common. More than half of OECD countries now require private-sector employers to report gender pay gaps, with many extending these obligations to public-sector organisations.
 
Pay transparency policies often require employers to report both overall gender pay gaps and gender pay gaps by job category, using gender-neutral job classifications. The second indicator helps detect wage discrimination (women being paid less for work of equal value), while the first helps identify discrimination in access to higher-paid jobs when the second indicator does not reveal anything (e.g., women being less likely to access higher-paid managerial positions). However, it is important to remember that a gap in itself is no proof of discrimination. Rather, it signals the need for deeper analysis into its causes, including the possibility of discriminatory practices.
 
For employers to be compelled to conduct such analysis when a gap emerges (and to take action to remedy it) public disclosure of pay gaps is essential, since transparency creates reputational pressure on employers to address unjustified differentials. The possibility of “name and shame” if employers do not act often drives stronger change than financial penalties alone, which tend to be modest and infrequently enforced.

Overall, the evidence suggests that pay transparency policies reduce the gender wage gap [1]. In five countries (Austria, Canada, Denmark, Switzerland, and the UK) researchers have been able to identify the causal effects of these reforms using quasi-experimental methods. They rely on a difference-in-differences approach, comparing changes in the gender wage gap before and after the reform between firms subject to the policy (those above the employee-size threshold triggering compliance) and firms not subject to it (those below the threshold). To strengthen robustness, these analyses are typically complemented by a regression discontinuity design, which places greater weight on comparisons between firms that are otherwise very similar, i.e., those just above and just below the threshold. Except for Austria where no effect is detected, the gender wage gap declined following the introduction of pay transparency reforms in Canada, Denmark, Switzerland, and the UK. Notably, these reductions are mainly driven by slower wage growth among men rather than faster wage growth among women (Switzerland is the sole exception, where the effect is concentrated among new hires, with no clear evidence on whether female recruits gain or male recruits lose in wage terms). One plausible explanation is that, in a more transparent environment, firms become more cautious about granting high-pay exceptions, which tend to be requested more frequently by men, because such exceptions could set precedents for future negotiations with women.

A few countries or subnational jurisdictions have taken (or are currently taking) steps to require employers to report pay gaps disaggregated by characteristics beyond sex at birth. These extensions concerns age (e.g., Australia, France, Portugal), disability (e.g., Canada for federally regulated employers, and the UK for employers with more than 250 employees), LGBTIQ+-related identity (with employers with more than 50 employees in British Columbia in Canada to report pay data across a broad range of gender identities, including cisgender and transgender women, cisgender and transgender men, non-binary individuals, and those who prefer not to disclose their gender), and race/ethnicity (e.g., again Canada and the UK).

However, requiring employers to report pay gap along all these dimensions is unrealistic. Not only would it create an unbearable administrative burden, but characteristics such as race/ethnicity, LGBTIQ+ identity, and disability are not systematically collected in HR records (unlike sex assigned at birth and age) and would therefore necessitate dedicated data collection efforts. Moreover, because these characteristics are highly sensitive, such data collection would involve strict confidentiality safeguards that are difficult to guarantee at the employer level and would likely result in high non-response rates among employees.

An alternative approach is to assign a state public body the responsibility of regularly computing and publishing firm-level pay gaps across protected characteristics. This could rely on administrative data (at least for sex assigned at birth and age) and/or census data in countries where information on disability, LGBTIQ+ identity, and race/ethnicity is collected. This model already exists. Since 2021, Lithuania’s State Social Insurance System has been calculating and publicly disclosing gender pay gaps for employers with at least eight employees. Centralising data production in this way increases accuracy, ensures comparability across employers, and thus strengthens the likelihood that monitoring efforts translate into real progress for at-risk groups. Evidence shows that when employers self-report, gaps tend to be understated, weakening incentives for change. 

Monitoring employer hiring behaviour

Holding employers accountable also requires monitoring their hiring behaviour. Beyond addressing the unintended consequences of anti-discrimination laws, focusing on the hiring stage is particularly important, as efforts to reduce wage gaps can also produce unforeseen consequences. For example, employers who discriminate in pay or promotion may respond to equal pay for work of equal value legislation or new pay-transparency requirements by reducing the hiring of under-represented groups. This may occur because the labour of these groups becomes relatively more expensive and, in the case of pay-transparency policies, because employers are typically not required, owing to confidentiality and statistical robustness concerns, to report pay gaps in job categories where the number of employees in either group falls below a minimum threshold. At the same time, however, pressure to narrow wage gaps may prompt employers to revise their overall practices, with positive effects not only on the wages of under-represented groups but also on their employment. Reflecting these opposing mechanisms, the empirical evidence to date is mixed. For instance, the Equal Pay Act in the US has been shown to reduce women’s employment and promotion prospects [2], whereas pay-transparency policies appear to have had an overall positive effect on women’s employment [1].

One option (central to a bill adopted in 2023 by the French National Assembly but not taken up by the Senate and therefore stalled), would be to task a national public body with conducting large-scale correspondence studies, including in the labour market, building on research showing how hiring discrimination can be estimated at the employer level [3]. The objective would be to identify employers engaging in discriminatory hiring across a wide range of grounds. Where discrimination is detected (albeit in an understated manner since correspondence studies can only capture the first stage of the hiring process), employers would be required to implement corrective measures within a defined timeframe, with the possibility of public disclosure in cases of non-compliance.

Yet testing the entire universe of employers is not realistic. A more feasible approach could be to build and regularly update a “risk dashboard” using multiple data sources to identify employers more likely to discriminate, and to run correspondence studies only on this subset. These could include employers with unusually large pay gaps revealed through pay-transparency tools, those where under-represented groups remain scarce in quality jobs (such as permanent contracts), both in absolute terms and relative to comparable employers in the same local labour market, or those whose job advertisements use coded language that may signal bias. For example, job ads containing ageist language, even when not overtly discriminatory, have been shown to partly predict age discrimination in hiring.

Looking ahead, complementary monitoring approaches could be explored. Large-scale correspondence studies remain a powerful tool to detect discrimination, but they are costly, not only to design and implement, but also for employers, who devote time and resources to reviewing applications they reasonably believe to be genuine (a cost and ethical concern that Institutional Review Boards have traditionally deemed acceptable given the societal value of measuring discrimination). In addition, the growing reliance of recruitment platforms and email providers on authenticated user accounts is making such studies inceasingly difficult to conduct when applicants are fictitious, both in practice and legally.

In this context, an alternative and potentially scalable avenue would be to require firms above a certain employee threshold to maintain application registers covering all job applicants to their vacancies. These data could then be analysed, ideally by independent third-party research bodies to ensure impartiality, to assess how candidates progress through successive stages of the recruitment process, from initial screening and interview invitations to final hiring decisions. Such an approach, already implemented by some employers on a voluntary basis, would make it possible to identify whether certain characteristics protected by law are associated with lower selection probabilities, holding other observable characteristics constant, including the skills signalled in CVs. 

Law enforcement is not enough, employers also need to know what works

Enforcement of anti-discrimination laws, including via the monitoring of employer practices, can act as a strong incentive for employers to take action, by reducing the likelihood that discrimination goes undetected and thus unchallenged. However, these measures are unlikely to spur significant change unless employers are also informed about which hiring and career-management practices limit statistical discrimination and the influence of (un)conscious bias.

One strategy that appears particularly promising in this context is the adoption of a skills-first approach, whose uptake has accelerated since the pandemic since acute labour shortages pushed employers to think about more effective ways of identifying talent. This approach refers to human resources practices that prioritise a candidate’s demonstrated skills, regardless of how they were acquired, over traditional credentials such as formal education levels (e.g., default university degree requirements), specific qualifications (e.g., mandatory IT or software certifications), or linear work histories.

Because statistical discrimination stems from employers’ uncertainty about candidates’ productivity, restructuring recruitment processes around the objective measurement of skills should mechanically reduce statistical discrimination against under-represented groups. This positive consequence is even more likely since these groups may disproportionately show non-traditional educational, qualification, or career paths. 

Furthermore, when skills are treated as the primary criterion for selection and progression, opportunities for conscious and unconscious bias to influence decisions at both recruitment and career-management stages are substantially reduced. To further ensure that HR processes remain free from bias, research points to the importance of promoting the responsible use of AI over human decision-making for HR tasks that can be automated, while strengthening the regulation of tasks that cannot easily be automated, such as performance evaluations.

A skills-first approach…

A skills-first approach is typically organised around three steps. First, employers translate job titles into the set of skills they entail, and rewrite job ads so they no longer rely on traditional credentials but are instead based on the skills required. Second, employers screen CVs and letters of application based on the skills they reveal. Third, for screened-in candidates, employers develop simulation tasks and/or standardised interviews designed to verify whether they possess the skills needed for the role.

…combined with the responsible use of AI for automatable HR tasks…

Randomised controlled trials (RCTs) show that, for HR tasks that can be automated, including most of those involved in skills-first hiring, AI, when used responsibly, identifies talent more effectively than human decision-making and, in doing so, reduces discrimination. For example, in the US, it has been shown to lower barriers faced by women entering male-dominated tech roles [4].

Responsible use has two key dimensions. First, AI tools must be designed to minimise bias. For example, an AI screening system that aims not only to identify candidates whose skills match job requirements but also to predict hiring success or future performance (in a way blind to their demographics and other irrelevant characteristics) must be trained on unbiased data. If candidates at risk of discrimination are underrepresented in the training dataset (as is often the case when data reflect historically biased human decisions), the AI may replicate or even amplify this bias, systematically disadvantaging underrepresented groups [5]. Second, responsible use requires continuous auditing of AI tools, even when they are designed to be unbiased at the outset. Seemingly neutral configuration choices can inadvertently disadvantage certain groups. For instance, an RCT run in the US found that a simulation task requiring candidates to solve a problem while being observed penalised women disproportionately, not because of their lower technical skills, but because of greater susceptibility to performance anxiety: not a single female completed the task correctly under observation, whereas all did so when unobserved. 

When AI is managed responsibly, evidence suggests the need to adopt a comply-or-explain rule, whereby hiring managers are expected to follow AI recommendations unless they provide a clear justification for deviating from them. In the absence of such safeguards, research conducted across 15 US firms has found that managers’ departures from the results of skills-based simulations at the hiring stage reintroduce human bias and ultimately lead to lower-quality and thus discriminatory recruitment decisions [6].

…and enhanced evidence-based regulation for non-easily automatable HR tasks

For HR tasks that cannot be easily automated (such as performance appraisals which rely on individualised managerial judgment), the process should be tightly guided by rules grounded in robust empirical evidence (most of it stemming from the United States) on what effectively prevents bias from taking hold. Such rules include avoiding open-ended questions that invite subjective judgments, and instead requiring managers to provide specific, concrete examples to substantiate their assessments. Where rating scales are used in addition to motivated assessments, research emphasises the importance of carefully designing and testing them in advance, as some formats, such as 10-point scales, have been shown to amplify bias against under-represented groups [7].

Evidence also points to the value of managers completing evaluations, where possible, before reviewing employees’ self-assessments. Members of groups at risk of discrimination tend to underrate their own performance, often due to the internalisation of negative bias, which can create anchoring effects that distort managers’ perceptions and undermine final evaluations [8]. 

Finally, research shows that exposing managers to short, targeted online diversity training immediately before they make HR decisions can meaningfully reduce discriminatory behaviour [9]. These brief interventions, which refocus attention on skills assessment and common sources of bias to be avoided, are far more effective than generic diversity training delivered at onboarding or as part of annual programmes, which are often disconnected from the moment when hiring and career-management decisions are actually made and too broad in scope to directly inform them [10].

Addressing concerns about the productivity of some at-risk groups

A skills-first approach is expected to mechanically reduce statistical discrimination by shifting attention to skills, which are more reliable proxies for productivity than traditional credentials. However, further progress requires also addressing the pre-existing factors that shape negative expectations about the productivity of some at-risk groups. 

For instance, the persistence of traditional gender norms that place a disproportionate share of domestic and caregiving responsibilities on women can place them at a disadvantage in the labour market. This is particularly the case when women apply for senior managerial roles, as employers may statistically discriminate against them, perceiving their unpaid workload as incompatible with the demands of such positions. For immigrants and their descendants in countries characterised by low-skilled immigration, employers’ expectations of lower productivity may be driven by these populations’ lower socio-economic backgrounds and their transmission across generations.

Addressing these concerns requires policies that directly tackle their underlying drivers. In the case of women, this objective notably entails promoting a more equal sharing of caregiving responsibilities from a child’s birth onwards. Evidence shows that policies encouraging fathers’ early involvement have lasting ripple effects, fostering more gender-balanced parenting as children grow. Such outcomes can be achieved by extending paternity leave and increasing its uptake through mechanisms such as father quotas (periods of leave reserved exclusively for fathers) or bonus periods, which grant additional paid leave when fathers commit to taking a minimum share. For immigrants and their descendants in countries characterised by low-skilled immigration, the priority lies in strengthening equality-of-opportunity policies from early childhood onwards. Ensuring equal access to high-quality early education, sustained support throughout the school trajectory, and targeted interventions where needed can help children from immigrant backgrounds close inherited socio-economic gaps and leave the education system with skill levels comparable to those of peers from more advantaged backgrounds.

However, productivity constraints affecting some at-risk groups are not always pre-existing. They may result from repeated exposure to discrimination. Such discrimination can not only discourage investment in the hard and soft skills needed to succeed in the labour market but also foster forms of withdrawal or oppositional identity-building, whereby members of at-risk groups define themselves in contrast to the expectations of the rest of the population. This dynamic has been documented, for example, among Muslim populations living in Christian-heritage societies, who face particularly high levels of discrimination [11]. Evidence suggests a risk of a discriminatory equilibrium in which anti-Muslim bias fuels community withdrawal, which in turn ultimately ignites statistical discrimination and reinforces taste-based discrimination, locking both sides into a self-perpetuating vicious circle. Breaking out of these harmful dynamics is far from straightforward and requires action on both fronts: not only policies aimed at preventing radicalisation among discriminated groups, but also ambitious, sustained efforts to combat discrimination across society as a whole.

Affirmative action policies can trigger backlash

Beyond raising concerns about equality before the law, affirmative action policies may inadvertently reinforce the very drivers of discrimination they seek to dismantle. This risk arises for both main forms of affirmative action: quotas or targets, which require employers to allocate a minimum share of jobs to individuals from at-risk groups (or to demonstrate progress towards a benchmark) under the threat of penalties; and hiring or employment subsidies, which offer financial incentives to encourage employers to hire or retain candidates from these groups.

Quotas and targets

Quotas and targets do mechanically increase the representation of the groups they target and are therefore effective at improving hiring and retention outcomes for beneficiaries. For instance, this has been shown in Austria, using a regression discontinuity design that exploits the requirement that Austrian firms hire at least one individual with a disability per 25 employees without disabilities, creating quota thresholds at firm sizes of 25, 50, and so on [12]. However, they have largely failed to deliver on their core promise: acting as a temporary mechanism that would ultimately eliminate workplace discrimination and render affirmative action unnecessary. The original rationale was twofold. First, quotas and targets were expected to reduce (un)conscious bias by fostering meaningful contact between employers, co-workers, and individuals from at-risk groups. Second, they were meant to curb statistical discrimination by incentivising employers to invest in better screening tools, to more accurately assess individual productivity. 

While there is some evidence that the second objective has been at least partially achieved [13], research suggests the first has fallen short. A key reason lies in an unintended consequence of quotas and targets: they generate resentment and perceptions of unfairness, thereby exacerbating bias, possibly to an extent that outweighs any debiasing effect of increased contact. In practice, the share of jobs that workers from at-risk groups would hold in the absence of discrimination is unknown and varies across firms, depending on skill requirements and the local availability of qualified candidates. When uniformly imposed across employers, quotas and targets are therefore often perceived by those who do not benefit from them as arbitrary and as granting undue advantages to certain groups, which can ultimately lead to a net increase in bias against their intended beneficiaries. This mechanism may in turn reinforce statistical discrimination by fostering the belief that individuals from targeted groups are recruited not because of their skills, but simply because of their group membership.

Hiring and employment subsidies

Like quotas and targets, hiring and employment subsidies aimed at groups at risk of discrimination may unintentionally exacerbate discrimination against these groups. These policies operate by compensating employers for hiring or retaining minority candidates, rather than sanctioning discriminatory behaviour. As a result, they risk reinforcing the belief that beneficiaries suffer from a productivity shortfall for which employers must be compensated, thereby fuelling statistical discrimination. At the same time, by framing some hires as “subsidised”, such measures may also strengthen negative perceptions towards beneficiaries, contributing to increased taste-based discrimination. Consistent with these mechanisms, available assessments find little evidence that hiring and employment subsidies improve the hiring or retention of their intended beneficiaries [14].

Limitations and gaps

A growing body of evidence identifies hiring and career-management practices that limit statistical discrimination and leave minimal scope for (un)conscious bias to operate. Yet, far less is known on how to eliminate bias itself and thereby curb taste-based discrimination. Social psychology research shows that debiasing is fostered when individuals develop a sense of shared humanity with those they initially perceive as belonging to an “outgroup”. Yet translating this insight into concrete, scalable policies that durably reduce prejudice and stereotypes in the broader population remains a major challenge.

Schools are a natural entry point, as shaping attitudes early is critical to building more cohesive societies. However, evidence on how best to embed effective debiasing techniques into school curricula remains scarce, with few RCTs conducted in this setting. While experiments run in laboratories and real-world non-school settings suggest that non-judgmental exchanges focused on empathy building (such as through perspective-giving and perspective-taking) are effective, the first classroom-based RCT on the topic indicates that such approaches may also backfire [15]. Increasingly frequent negative group dynamics, possibly fuelled by students’ growing exposure to online disinformation and hate speech, can undermine these interventions, highlighting the need to better equip students to navigate digital environments and disengage from hateful content, including misogyny, homophobia, and racism.

Overall, substantially more research is needed to understand how to sustainably foster positive attitudes towards outgroups, both in schools and in society at large.

Summary and policy advice

A promising strategy to end labour-market discrimination rests on four mutually reinforcing policy pillars.

First, employers must face clear and credible incentives not to discriminate. This aim requires not only comprehensive anti-discrimination laws (covering a wide range of grounds and all areas of employment), but also their effective enforcement. Strengthening incentives also entails the systematic monitoring of employers’ hiring and career management practices. This objective could be achieved by requiring firms above a certain employee threshold to maintain application registers covering all job applicants to their vacancies, and by extending pay-transparency policies beyond gender to other protected grounds, where data allow.

Second, employers must be equipped with evidence-based practices that reduce discrimination. Incentives alone will not suffice unless employers are informed about HR practices that reduce statistical discrimination and limit the influence of (un)conscious bias. This goal includes adopting a skills-first approach combined with the responsible use of AI for automatable HR tasks, and enhanced regulation of non-automatable decisions such as performance evaluations.

Third, policies must address the structural factors that fuel statistical discrimination. Where employers’ productivity concerns reflect real pre-existing constraints affecting some groups, these barriers must be tackled at their source. For example, reducing the burden of unpaid domestic and caregiving work borne by women, through policies that promote more equal sharing of care from a child’s birth, can help close perceived productivity gaps and improve women’s access to senior managerial positions.

Fourth, when productivity constraints are themselves the product of repeated exposure to discrimination, policy must break the resulting vicious circle. This requires action on both sides: preventing withdrawal or oppositional identity-building among discriminated groups, while simultaneously fostering attitude change among those who discriminate. Although research has yet to fully identify scalable solutions in this area, ignoring either side risks entrenching self-perpetuating discrimination dynamics.

Finally, affirmative action policies should be used with caution, as evidence suggests they may generate backlash.

Acknowledgments

The author thanks an anonymous referee and the World of Labour editors for many helpful suggestions on earlier drafts. Version 2 of this article updates all parts of the text, adds new Further readings and Key references.

Competing interests

The World of Labour project is committed to the European Code of Conduct in Research Integrity. The author declares to have observed these principles.

© Marie-Anne Valfort