What’s in a Label? Public Use and Perceptions of Labeling Alternatives in Criminology

Abstract

Research indicates that crime-first language (“criminal”) increases stigma, but there is limited evidence comparing person-first language (“person with a conviction”) to other non-deviant terminology. Using two survey experiments, we test whether using person-first language in the employment context has a mimicry effect (i.e. people adopt language they are exposed to) and whether language has a stronger influence when paired with positive employment credentials. We do not find consistent evidence of mimicry. Respondents viewing a positive credential were more likely to use person-centered or professional language and positive credentials generally improved perceptions of the applicant. However, person-first language can cancel out beneficial credential effects compared to alternative language. When considering alternatives to crime-first language, two implications emerge: positive information is more consistently influential than the terminology used, and after establishing a person has a criminal record, substituting other identity labels (e.g. person, applicant) can further reduce stigma.

Keywords:

• Criminal record stigma
• labeling
• person-first language
• positive credentials
• employment

Introduction

Criminal record labels and their accompanying consequences can have social, economic, and public safety implications (Bernburg & Krohn, 2003; Boppre & Reed, 2021; Chiricos et al., 2007; Denver et al., 2017; Tran et al., 2018). At its core, the labeling process is about shaping identity, and social interactions and public perceptions play a pivotal role. A negative label connects undesirable attributes to a person, and stereotypes associated with a label can influence decisions to distance from—and discriminate against—members of the stigmatized group (Goffman, 1963; Link & Phelan, 2001). In turn, those labeled might internalize perceived traits and the stigmatized identity (e.g. Cooley, 1902; Mead, 1934). In other words, social rejection can lead to the “ultimate acceptance of deviant social status and efforts at adjustments based on the associated role” (Lemert, 1951, p. 77; see also Becker, 1963; Tannenbaum, 1938).

In articulating the influence of social perceptions, early labeling theorists indicated that language plays a key role in the “othering” process because words become synonymous with anticipated behaviors. The criminal label “carries a number of connotations specifying auxiliary traits characteristic of anyone bearing the label” (Becker, 1963, p. 33) and can define who the person is—not just what the person has done. For example, people may assume that a person labeled a “criminal” will continue to commit crimes “because he has shown himself to be a person without ‘respect for the law’” (p. 33). Furthermore, a powerful label can become a “master status” or dominant characteristic that overrides other identity roles (Becker, 1963; Goffman, 1963; Hughes, 1945).

It stands to reason, then, that the modification or replacement of stigmatizing language in everyday interactions and decisions could disrupt the labeling process by improving perceptions of stigmatized groups. Recent research has supported the idea that seemingly minor changes to labels can meaningfully influence how stigmatized groups are perceived by others more broadly (e.g. Chong & Druckman, 2007). Within criminology, simply replacing crime-first terms (or noun labels, such as “felon”) with person-first language (“person with a felony”) can notably improve public opinions and perceptions about the referenced person or group. For example, the public is more likely to support social reintegration efforts post-incarceration when person-first language is used instead of crime-first language (Jackl, 2023).

Acknowledging the potential harm of using crime-first language, a variety of terms have been proposed and adopted, but testing alternative labels is still a nascent effort. While person-first language (e.g. “formerly incarcerated person”) is a popular and important alternative (e.g. La Vigne, 2016, 2023), new research suggests terms that avoid mentioning the criminal record altogether (e.g. “returning citizen”) can have more positive effects on public perceptions (Jackl, 2023).¹ Further, it is unclear whether and how the public adopts alternative language labels. Despite a lack of research on alternative label usage among the public, early labeling theorists observed “a tendency to imitate among men” (Mead, 1934, p. 59; see also Cooley, 1902, p. 4) and there is some evidence supporting alternative labeling language diffusion within the scientific community (e.g. Dickinson et al., 2023). Two open public opinion questions are whether alternative labels that shift the focus away from the criminal record can meaningfully shape criminal record stigma in different decision contexts, including employment, and whether people adopt destigmatizing language through exposure. Building from Goffman’s (1963) concept of a disidentifier, or the idea of additional information creating ambiguity in the veracity and strength of a stigmatized status, the current study tests whether labels portraying different master statuses or identities, such as a job “applicant,” might shape language usage and perceptions (and therefore, opportunities) in society.

We contribute to this theoretical and policy-relevant discussion through two connected survey experiments in the employment context. In Study 1, we examine the language people opt to use when describing a hypothetical job applicant with a conviction record after a vignette experiment and whether randomizing conviction type, type of positive credential, or applicant race affected respondent language choices. Next, to detect whether respondents were simply mimicking the language we selected and presented in the survey vignette, we created a new survey experiment. In Study 2, we test whether randomly priming respondents with one of three common language categories (person-first, professional, or person) impacts the language they use to describe the applicant and their perceptions of the applicant. In the new experiment, we also randomized the presence of two positive credentials—with varying levels of perceived value—to test if language priming has different effects for strong and weak desistance signals. While we are unable to determine the long-term significance people place on different labeling categories through our study design, an improved understanding of language usage, the link between language and perceptions of others, and whether responses to language priming interact with positive credentials can provide key insights into labeling processes.

Stigma Management, Identity, and Labels

Criminal record stigma management centers on “strategic control over the image of [self]” (Goffman, 1963, p. 10) by attempting to influence the type and timing of information that others can access. Unlike fully concealed stigmas that require disclosure to be publicly known, modern criminal record histories could be considered a quasi-concealed form of stigma; while not outwardly visible, information can be gleaned from formal (e.g. background check) or informal (e.g. private mugshot website) methods. Quasi-concealed stigma can lead to a change in behavior as people actively avoid social and economic opportunities that may reveal the criminal record (e.g. Lageson 2016). When criminal records are (or will be) revealed, popular stigma mitigation strategies can involve controlling the narrative (e.g. Ali et al., 2017; Denver & Ewald, 2018) or contextualizing the criminal record. Contextualization involves providing positive information that serves as a “disidentifier,” emphasizes other “master statuses,” or otherwise calls into question the assumptions underlying the stigmatizing characteristic (DeWitt & Denver, 2020; Goffman, 1963, p. 44).

However, criminal labels complicate contextualization efforts. By “tagging” the person a criminal and linking them to a “criminal credential” through language, the label can impede the ability of other components of the person’s identity to contextualize the criminal behavior and effectively neutralize stigma. At the extreme, the label can become a dominant identity (Hughes, 1945) and lead to social and economic exclusion (e.g. Bernburg & Krohn, 2003; Denver et al., 2017). While research tends to examine either the impact of labels on individuals’ outcomes or how positive credentials mitigate criminal record stigma, there is not clear evidence on how labels and credentials might work in tandem (or in opposing ways) when appearing simultaneously.

How Language Shapes Public Perceptions

Proponents of person-first language policies argue that the language used to describe social groups can impact public perceptions of those groups in meaningful ways. Although different descriptors may appear equivalent on the surface, certain terminology is connected to positive or negative sentiments, an idea known as the equivalency framing effect (Druckman, 2001). For instance, Rucker et al. (2019) find that “illegal immigrant,” “illegal alien,” and “undocumented alien” are all perceived as having more negative connotations than “undocumented immigrant.” In the only study we are aware of that directly tests the widespread adoption of a labeling policy, Djourelova (2023, p. 802) finds that an Associated Press (AP) media ban on using the controversial and politically charged phrase “illegal immigrant,” which led the term to “instantaneously disappear” from AP content, reduced support for several restrictive immigration policies.

The idea that language impacts cognition and perception is rooted in the theory of linguistic relativity—sometimes referred to as the Sapir-Whorf hypothesis. Although it is generally agreed upon that language operates as a conduit through which we express our thoughts, theories of linguistic relativism suggest that the relationship between language and cognition is bi-directional; in addition to conveying our thoughts, language may play a role in shaping the way we think and perceive the world (Lucy, 2001). Although research supporting linguistic relativity is mixed, evidence that label-use impacts cognitive perception is relatively robust. Using noun labels, in particular, has been found to trigger essentialist thinking—or the assumption that group members share “hidden underlying essences that determine category membership” (Ritchie, 2021, p. 465)—relative to adjectives or verbs (Witkowska et al., 2022). This concept has been established especially well in the child development literature. In Gelman and Heyman’s (1999) classic study of children, participants were presented with a description of a child that used either a noun classificatory label (such as “Rose is a carrot eater”) or a verbal predicate (“Rose eats carrots whenever she can”) and asked follow-up questions regarding the stability of Rose’s defining trait (eating carrots). Children presented with the label condition believed the key trait was dispositional and a fundamental part of Rose’s identity (see also Baron et al., 2014).

In addition to encouraging generalizations about the characteristics of group members, label categories are enduring in another way: individuals have an easier time remembering categorizing details about other people when these details are associated with a label, even after the label itself has been forgotten (Aslanov et al., 2021). As such, labeling is an important “causal ingredient” of stereotyping (Bigler & Liben, 2007, p. 166). Social categorization has also been found to be a key component in the development of biases in adults. As Tajfel and Turner (1986, p. 13) describe, “the mere perception of belonging to two distinct groups…is sufficient to trigger intergroup discrimination favoring the in-group.”

Evidence on Labeling Effects

Research in the fields of medicine and mental health further illustrates that subtle changes in wording—and noun label use in particular—can impact social appraisal by triggering inferences beyond what is explicitly conveyed about the subject. Compared to person-first language, noun labels have been found to elicit negative reactions in addiction research (Kelly & Westerhoff, 2010), mental health studies (Granello & Gibbs, 2016), and research that merges psycho-medical labeling with criminal labeling, such as “pedophilia” labels (Imhoff & Jahnke, 2015). Similarly, in the field of criminology, criminal labels can impact social appraisal. In a study testing a person-first language policy announced by the Department of Justice in 2016, Denver et al. (2017) found that using a crime-first category label (“convicted criminals”) relative to a person-first descriptor (“people convicted of crimes”) increased perceived recidivism risk for violent crime types. However, the label choice did not have a meaningful difference when respondents considered recidivism risk for those convicted of property or drug crimes (Denver et al., 2017). Instead, negative connotations from criminal labels appear to be particularly damaging for the most stigmatized: those who have committed or are convicted of violent or sex offenses (Denver et al., 2017; Harris & Socia, 2016; Huebner et al., 2019; Lowe & Willis, 2020).

In a recent study examining how language choices might impact opportunities for individuals with criminal records, Jackl (2023) finds using the term “returning citizen” has more positive effects on the public’s preferences for reintegration policies than person-first language (“person who was formerly incarcerated”). A consideration is that the term “returning citizen” may be confusing to the broader public without additional context, and convict criminologists—i.e. those centering the lived experiences of those with criminal justice system involvement in research, often with incarceration records themselves—have considered the terms to be a misnomer because this population does not “carry all the rights of citizenship” (Cerda-Jara et al., 2019, p. 3). However, it is also possible that using a person-first label that explicitly references criminal legal system involvement could cause people to perceive this group more negatively than a label that does not. For example, in the substance use field, researchers similarly find a stark difference in public perceptions between noun-labels and what they call positive counterterms, but there are subtle differences depending on the replacement term used. For example, while various counterterms had clear advantages over noun-labels such as “addict” or “alcoholic,” referring to the “recurrence of use” had a meaningful improvement in positive connotations over “relapse” (Ashford et al., 2018). These studies support the notion that words can carry meaning, and adjustments to language can influence public opinion and social rejection.

In sum, a key takeaway from the criminology public opinion literature is that people more often associate criminal labels with negative traits (e.g. “dangerousness”) compared to person-first descriptors when they are randomly assigned to view different descriptions (Elderbroom et al., 2021).² However, the underlying perceptions behind personal language choices are not well understood. Similarly, while there have been broader language change movements to advance the goal of reducing criminal record stigma (e.g. National Institute of Justice Style Guide, 2022; McDowall et al., 2022), it is unclear whether alternative language choices have a meaningful effect on everyday decision-making. It is also unclear whether exposure to certain labels alters how people label; in other words, whether there is a “priming” effect or transmission of language. Evidence of priming and mimicry may support policies requiring or encouraging the use of positive counterterms at an organizational level (e.g. the Associated Press, American Medical Association, the National Institute of Justice). Rather than only seeing such language in unique or special situations, if the public is increasingly exposed to destigmatizing language in ordinary exchanges or decision processes, this may have the benefit of reminding people of different linguistic options and encourage them “to use new words, expressions or constructions, and moreover to remember them for subsequent use” (Pickering & Garrod, 2017, p. 173). New language adoption, in turn, could improve public perceptions of stigmatized groups and subsequently discourage discrimination and ostracization at an individual level.

To explore the potential impact of such initiatives, the current study examines whether people tend to use person-first, crime-first, or other language to describe a hypothetical job applicant with a conviction record; tests whether randomly assigning language alters personal language choices; and then examines the relationship between language priming, positive credential types, and perceptions of the applicant.

Methods

Data and Experimental Design

We present two survey experiments in the current study. The original data for Study 1 were collected in April and May of 2018 using a nationwide survey of adults in the United States. The survey was administered through Amazon’s Mechanical Turk (MTurk), a nonprobability platform that connects respondents to researchers. In Study 1, we used the following factorial design, with a total of 22 possible conditions (DeWitt & Denver, 2020):

• Race: A Black or white male applicant. We assigned one of 10 names used by Bertrand and Mullainathan (2004), which appeared at the top of the job application.

• Criminal record type: No criminal record, a drug felony conviction, or a violent felony conviction. Both convictions referenced an 18-month prison sentence. We only consider applicants with a criminal record in the current study.

• Credentials: One of five credential conditions: no credential, involuntary completion of a job skills training program, voluntary completion of a job skills training program, a state-issued occupational license to be a barber, and a reference letter from a prior employer.

See Appendix A for the exact wording in the vignette. The language used in the credential descriptions varied in both terms and dosage and included direct descriptions (e.g. the applicant), indirect descriptions (e.g. individuals with felony convictions), professional titles from prior jobs (barber), and positive descriptors (good worker).³

Building on Study 1, we launched a new survey experiment (Study 2) in July 2022 through the Prolific platform.⁴ Prolific screens respondents upfront on multiple dimensions, allowing researchers to select from a variety of sample eligibility criteria. For the follow-up survey, we restricted the pool of eligible respondents to those who reported prior hiring or interviewing experience, as our follow-up study is specifically interested in language and positive credentials in the hiring context.

The content of Study 2 largely mirrors our original vignette in Study 1 with three exceptions. First, based on null effects in the main factorial survey findings (DeWitt & Denver, 2020), we held race and criminal record type constant. All hypothetical applicants had a violent felony record (aggravated assault) and were Black men.⁵ Second, we narrowed the positive credentials to three categories: no credential, involuntary job training program completion, and voluntary job training program completion. We made minor adjustments to the two credential descriptions to ensure they were as similar as possible except for the voluntary/involuntary nature of the program. Third, the program descriptions were modified to incorporate a language label condition (applicant, person, or person with a felony conviction) in three locations per description. Language labels were randomized across respondents, but consistent within them. See Appendix C for the exact credential language.

Third, we retained several follow-up questions and included a new one. We used identical language as the original survey to ask the respondent’s perceptions of the likelihood of the person/applicant/person with a felony conviction committing a workplace-related crime, committing a crime in the future, being a trustworthy employee, and regularly showing up for his shifts and arriving on time. We also added an item that asked about perceptions of “good moral character,” a controversial descriptor used in some policy discussions and legislation (CCRC, 2019).⁶

Design for Open-Ended Responses

In each survey, we asked respondents “What is the likelihood that you would call [applicant’s first name] to come in for an in-person interview?” Response options ranged from 1 “Very unlikely” to 5 “Very likely.” Respondents were then asked to rank-order the elements of the application that influenced their decision (1 “Most influential” to 5 “Least influential”) with four elements that everyone saw (criminal record, education, references, and work history) and a fifth—the positive credential—displayed if that condition appeared in the respondent’s vignette. Directly following this question, we asked respondents to provide a brief justification for why they ranked an element as the most influential. Responses to this question were content-coded into five language categories used to describe the subject of the vignette: no code (i.e. no reference to the subject), deviant or crime-first (i.e. stigmatizing criminal record labels), person (i.e. person-focused language that does not reference the criminal record), person-first (i.e. person-first language), and professional (i.e. job-identity language, such as applicant or candidate). Table 1 provides examples for each.

Table 1. Distribution of coded language types by randomized positive credential condition (Study 1).

Language	Examples	Full sample (n = 2891)	No credential (n = 696)	Positive credential (n = 2195)
No code	Assault is a serious crime. I’m afraid he would do it again.	58.42%	60.78%	57.68%
Deviant	Because it is important to make sure he has moved beyond being a repeat offender. I would never hire a felon. Criminals can ruin a business.	3.63%	3.88%	3.55%
Person	It’s important to know whether the person has experience working in a similar environment. This person has a relevant work history which is good.	14.91%	14.08%	15.17%
Person-first	I would not feel safe hiring this person with a history of assault. The fact that the person has a prior felony, no matter for what reason, shows the person is not of upstanding character, and extra time would be spent watching them.	2.73%	3.88%	2.37%
Professional	I would want to know the reason behind the assault as further qualifications, because it seems like he would be an excellent candidate for a long-term stay. This could let us know an idea of what the employee was doing before joining my organization. There could be no better way to predict the prospective employee’s job performance.	20.30%	17.39%	21.23%

Notes: We used a chi-square test of independence (9.70, p < .05) to compare the no credential and positive credential conditions. Fisher’s exact test reveals similar results (p < .05).

Samples

For Study 1, our sample consists of respondents who saw hypothetical job applicants with a criminal record and passed our data quality checks, as detailed in Appendix B ($N_1$ = 2891).

Study 2 includes 2,775 respondents who reviewed job applications and/or conducted interviews and passed the data quality checks. Table 2 provides descriptive statistics for both samples. Perceptions of the applicant’s auxiliary traits—dependability, recidivism risk, trustworthiness, and risk of crime in the workplace—are similar across studies and tend to cluster around neutral responses (3 = “Neither Agree nor Disagree”). The question about the applicant’s moral character only appears in Study 2 but also conforms to the trend toward neutral values (mean = 2.96, SD = 0.84).

Table 2. Descriptive statistics for Study 1 (n = 2891) and Study 2 (n = 2775) samples.

	Study 1—MTurk		Study 2—Prolific
	Mean/%	Std Dev	Mean/%	Std Dev
Perceptions of applicant
Dependability	3.64	0.89	3.69	0.79
Recidivism risk	2.99	0.89	2.92	0.84
Trustworthiness	3.18	0.92	3.27	0.84
Workplace crime	3.05	1.10	3.06	1.11
Moral character	—	—	2.96	0.84
Language type (coded) (%)
Deviant	3.63	—	4.32	—
Person-first	2.73	—	4.94	—
Person	14.91	—	14.09	—
Professional	20.30	—	14.52	—
No code	58.42	—	62.13	—
Experimental conditions
Credential type (%)
No credential	24.07	—	33.51	—
Involuntary job training	18.89	—	34.02	—
Voluntary job training	19.09	—	32.47	—
Occupational license	19.30	—	—	—
Reference letter	18.64	—	—	—
Criminal record type (%)
Drug felony	50.54	—	—	—
Violent felony	49.46	—	100.00	—
Language type (assigned) (%)
Person	—	—	33.05	—
Person-first	—	—	33.55	—
Professional	—	—	33.41	—
Race of applicant (%)
Black	50.81	—	100.00	—
Respondent characteristics
Age	36.06	11.56	41.47	13.02
Arrested (%)	12.18	—	13.66	—
Convicted (%)	7.71	—	8.97	—
Education (%)
Less than HS	0.62	—	0.32	—
HS Deg/GED	8.82	—	6.41	—
Some college	24.73	—	16.83	—
Associate’s degree	11.83	—	9.23	—
Bachelor’s degree	39.29	—	42.38	—
Graduate degree	14.53	—	24.83	—
Employer experience
Interview applicants? (%)	40.23	—	94.49	—
Review job applications? (%)	48.70	—	96.86	—
Employment status (%)
Unemployed	7.13	—	6.91	—
Part-time	15.46	—	12.22	—
Full-time	59.53	—	68.07	—
Other, not unemployed	17.85	—	12.79	—
Household income (%)
≤$14,999	16.64	—	8.72	—
$15,000–$29,999	18.26	—	11.06	—
$30,000–$44,999	20.30	—	14.99	—
$45,000–$59,999	17.16	—	16.29	—
$60,000–$74,999	11.28	—	13.15	—
$75,000–$99,999	8.75	—	15.14	—
≥$100,000	7.37	—	20.58	—
Political orientation
Conservativism	2.64	1.04	2.48	1.11
Republicanism	2.63	1.16	2.48	1.24
Race/Ethnicity (%)
White	80.32	—	83.53	—
Black	8.13	—	5.91	—
Hispanic*	7.44	—	7.39	—
Other	11.55	—	10.56	—
Sex (%)
Female	56.21	—	54.16	—

Notes: Five observations were missing education and seven were missing income in Study 1. Two observations were missing household income, and conviction and arrest history in Study 2. Fifty observations are missing for age in Study 2; however, Prolific screened for age to restrict the survey to adults (age 18 and older), and we do not include age in any of our analyses.

* Hispanic ethnicity is measured separately from race, so the percentages within the Race/Ethnicity category add up to over 100%.

Language use within respondents’ open-ended replies is also similar across studies. A small percentage of respondents used crime-first (Study 1 = 3.63%; Study 2 = 4.32%) or person-first (Study 1 = 2.73%; Study 2 = 4.96%) language while a higher percentage of respondents used person (Study 1 = 14.91%; Study 2 = 14.09%) or professional (Study 1 = 20.30%; Study 2 = 14.52%) language. The majority of respondents from both studies did not use language that fit into these categories (Study 1 = 58.42%; Study 2 = 62.13%).

Regarding respondent characteristics, most demographics are similar across studies, although Study 2 respondents are slightly older (Study 1 mean = 36.06; Study 2 mean = 41.47), more likely to have a graduate degree (Study 1 = 14.53%; Study 2 = 24.83%) and more likely to have annual income tiers at or above $75,000 (Study 1 = 16.12%; Study 2 = 35.72%). These differences likely reflect the use of a public sample in Study 1 and an employer sample in Study 2. According to the 2022 Current Population Survey, human resource managers—who are often responsible for making hiring decisions—more often tend to be women (72.9%), white (81.5%), and middle-aged (U.S. Census Bureau, 2022), and the typical entry-level education for those management positions is at least a bachelor’s degree (Bureau of Labor Statistics, 2022). Given these differences between the two samples, we also analyze the subset of employers in our public sample from Study 1 in supplemental analyses (n = 1408), which are closer to the demographics of the Prolific sample.

Analysis Plan

We have three analyses in the current study. Using data from Study 1, we first explore how respondents describe the subject of the vignette. Every respondent was exposed to professional language (“applicant”) in the vignette, and some were primed with additional language through the credential conditions. As such, we separate out the full sample by any credential/no credential to compare subgroups. If a subject used multiple language types to refer to the subject of the vignette (Study 1: n = 70; Study 2: n = 114) we either coded the most frequent language type they used or, if they used different language types at the same frequency, we coded the first language type that appeared in their open-ended response. We test the sensitivity of this decision by excluding all respondents using multiple language types from our analyses (Appendix D) and discuss any substantial departures from the main results in the next section.

Second, we examine whether randomly assigning different credentials or language labels to describe the hypothetical job applicant impacts the language respondents chose in their open-ended responses. We used multinomial logistic regression models to predict the probability that a respondent’s open-ended response used crime-first, person, person-first, or professional language relative to the respondent using no such language (“No Code”) to refer to the job applicant. The multinomial logistic regression in Study 1 includes covariates for the randomly assigned positive credential (or lack of one), conviction type reported on the application, and the race of the applicant to investigate whether positive credentials affect the language type used by respondents in their open-ended responses. In Study 2, the multinomial logistic regression includes covariates for one of three randomly assigned positive credentials (or the lack of one) and language primes. This enables us to test the effects of a more limited set of positive credentials (Involuntary or Voluntary Job Training) while accounting for language dosage and also assessing if respondents are simply mimicking the language of the vignette or are using different language variants of their own accord.

For our third analysis, we consider whether different language choices are associated with perceptions about the individual described in the vignette. We created a factor score (eigenvalue = 3.41, all factor loadings meet or exceed 0.78) and reverse-coded negative perceptions so that increases in this factor variable represent improved perceptions of the applicant (i.e. they are less risky/dangerous, more trustworthy/honest, and have better moral character). We then use an ordinary least-squares (OLS) model to examine the relationship between the randomized language labels and applicant perceptions for the Study 2 sample, controlling for the randomized credential conditions.

Results

Language Choice Categories by Positive Credential Information Access

First, we display in Table 1 how respondents used different language categories in the Study 1 full sample (column 1), and for just those who saw no positive credential (column 2) or any positive credential (column 3). We identified five categories: no code (a subject is never referenced), deviant or crime-first (e.g. “felon”), person (e.g. “this person”), person-first (e.g. “person with a conviction”), and professional (e.g. “applicant”). A majority of respondents across each sample made “no code” statements that did not contain any crime-first, person, person-first, or professional language use, such as “No work history or gaps indicate possibly having been fired or bad work ethic.” Yet, a sizable percentage, between 39 and 42% of respondents, do use one or more of these language variants in their open-ended responses. In addition, a chi-square test of independence indicates significant (p < 0.05) variation in the distributions of language type across samples.

Approximately a fifth of the coded responses fit within the “professional” (applicant) language category. There is a 3.84 percentage point difference in the prevalence of professional language between the no credential (17.39%) and positive credential (21.23%) conditions. This is suggestive of a credential effect, where viewing a positive credential increases respondent propensity to use terms like “applicant” or “good worker” to refer to the subject of the experimental vignette. By contrast, the prevalence of the remaining language types is very similar across groups. For example, the second most frequent type is “person” language (“this person”), and there is only slightly more than a percentage point difference across subgroups.

Table 3 further unpacks the impact of positive credentials on language type using a multinomial logistic regression. All coefficients are reported as both average marginal effects (AMEs), or the average change in the probability that an observation belongs to a given outcome category when an experimental condition changes by one unit, and relative risk ratios (RRRs), which represent the ratio of the probabilities that an observation falls into the language category for that column as opposed to the reference category (no code) for two observations that differ by one unit in that regressor. Because each coefficient also has a reference category comparison (e.g. involuntary job training compared to no credential), the interpretation in a multinomial logit is slightly complicated. As an example, the AME and RRR coefficients for the reference letter condition in the model predicting if a respondent falls into the professional language category are 0.12 (or 12 percentage points) and 2.02, respectively (p < .001). The former indicates that seeing a reference letter from a prior employer, as opposed to seeing no credential, increases the probability that a respondent will use professional language by 12 percentage points, on average, and holding all other experimental conditions constant. By contrast, the latter indicates that the probability of a respondent being in the professional language category is 2.02 times as high as the probability that they are in the no code category if the respondent is assigned the reference letter rather than no credential. In other words, viewing a reference letter credential—which uses the word “applicant” five additional times in the description than the no credential condition—notably increases our respondents’ propensity to use professional language.

Table 3. Multinomial logit predicting coded language type (Study 1, n = 2891).

	Deviant			Person			Person-first			Professional
	RRR	SE	AME	RRR	SE	AME	RRR	SE	AME	RRR	SE	AME
Credentials
No credential (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Invol job training	1.10	0.32	0.01	0.90	0.16	−0.01	0.59	0.20	−0.02	0.97	0.15	0.00
Vol job training	0.74	0.24	−0.01	1.01	0.17	0.00	0.50	0.18	−0.02	1.07	0.16	0.02
Occ license	1.35	0.38	0.01	1.29	0.21	0.03	0.79	0.26	−0.01	1.24	0.19	0.02
Reference letter	0.64	0.23	−0.02	1.45*	0.24	0.03	0.69	0.25	−0.02	2.02***	0.29	0.12
Criminal record
Violent (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Drug	0.98	0.20	−0.00	1.17	0.13	0.02	0.87	0.20	−0.00	1.03	0.10	0.00
Applicant race
White (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Black	1.21	0.24	0.00	1.15	0.13	0.01	0.84	0.20	−0.01	1.18	0.11	0.02

AME: average marginal effect; Invol: involuntary; Occ: occupational license; Ref: reference category; RRR: relative risk ratio; SE: standard error; Vol: voluntary.

Notes: The reference category for the coefficients reported in this table is the “No Code” language type. Marginal effects represent the average change in probability to be in the selected language category, relative to the reference category for that regressor. Huber-White standard errors are reported.

*p < .05; **p < .01; ***p < .001.

In addition to increasing the use of professional language, the reference letter credential also increases the use of person language (RRR = 1.45, AME = 0.03, p < .05) relative to no coded language in a response, but this finding is sensitive to excluding respondents with multiple coded language types (See Appendix D, Table D1). The results within Table D1 also indicate that the voluntary job training credential can decrease the probability of using person-first language, though we hesitate to ascribe much weight to this result given its sensitivity to model specification. When we focus on the subset of employer respondents, defined as those with experience interviewing and/or reviewing job applications, the reference letter credential is not statistically significant for the person language category (RRR = 1.27, AME = 0.00, p = .313); although the sample size is smaller (n = 1408), this provides additional caution against the person language finding in the main results.⁷ With respect to conviction type and applicant race, we find no effects across model specifications.

Is the Effect Driven by Positive Credentials or Language Primes (Mimicry)?

Consistent with the findings from Table 2, the results reported in Table 3 demonstrate that a reference letter from a prior employer can increase the use of person and professional language. A potential cause of this effect is that this credential contains a positive evaluation of the applicant from the standpoint of their prior employer. Another plausible interpretation of this finding is that respondents are mimicking the language we use in the description of this credential condition, though it is unclear why they would only do so for one of the four randomized credential conditions.

Our analysis in Table 4 allows us to determine which of these explanations is more plausible in a new sample and using a more limited set of credentials by experimentally manipulating the language within the credential description. We specifically selected job training credentials because, with the exception of whether the program was voluntary or involuntary, the two credentials are otherwise identical. Each respondent receives an equal dosage of language, but the type of language they view (person, person-first, or professional) is randomized.

Table 4. Multinomial logit predicting coded language type (Study 2, n = 2775).

	Deviant			Person			Person-first			Professional
	RRR	SE	AME	RRR	SE	AME	RRR	SE	AME	RRR	SE	AME
Credentials
No credential (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Invol job training	0.94	0.21	−0.01	1.37*	0.19	0.03	0.97	0.20	−0.01	1.40*	0.19	0.04
Vol job training	0.87	0.20	−0.01	1.51**	0.21	0.05	0.80	0.18	−0.01	1.12	0.16	0.01
Label language
Person (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Person w/FC	1.37	0.32	0.01	1.09	0.15	0.01	1.00	0.22	−0.00	0.91	0.13	−0.01
Professional	1.10	0.27	0.00	1.13	0.16	0.01	1.02	0.22	−0.00	1.29	0.17	0.03

AME: average marginal effect; Invol: involuntary; Person w/FC: person with a felony conviction; Ref: reference category; RRR: relative risk ratio; SE: standard error; Vol: voluntary.

Notes: The reference category for the coefficients reported in this table is the “No Code” language type. Marginal effects represent the average change in probability to be in the selected language category, relative to the reference category for that regressor. Huber-White standard errors are reported.

*p < .05; **p < .01; ***p < .001.

There is little evidence that randomly assigned language labels prime respondents’ own language use. Although there are no apparent language label effects in Table 4, there is an effect for professional language (RRR = 1.32, AME = 0.03, p < .05) when we exclude the 70 respondents that use multiple language types (Table D2, Appendix D). We interpret this as weak evidence for a mimicry effect, with the caveat that this finding is sensitive to model specification. By contrast, stronger effects are present for the credentials. Relative to no language type, the voluntary job training credential strongly increases the probability for a respondent to use person language (RRR = 1.51, AME = 0.05, p < .01). The involuntary job training credential also increases the probability that a respondent uses person (RRR = 1.37, AME = 0.03, p < .05, though this effect is also sensitive to model specification) or professional language (RRR = 1.40, AME = 0.04, p < .05).

The Relationships between Language Primes and Perceptions

Next, we test whether randomized language labels influence perceptions about the subject of the vignette. Table 5 reports the results from an OLS model predicting respondent perceptions of the job applicant.

Table 5. OLS models Predicting specific perceptions of the applicant (Study 2, n = 2775).

	Coefficient	SE
Credentials
No credential (reference category)	—	—
Involuntary job training	0.17***	0.05
Voluntary job training	0.36***	0.05
Label language
Person (reference category)	—	—
Person with felony conviction	−0.16***	0.05
Professional	−0.06	0.05
Model statistics
Intercept	−0.10 (0.04)
Root MSE	0.99
R-squared	0.03

MSE: mean squared error; SE: standard error.

Notes: Huber-White standard errors are reported.

*p < .05; **p < .01; ***p < .001.

Coefficients can be interpreted as increases (more positive appraisals) or decreases (more negative appraisals) in latent impressions of the subject of the vignette. We find that, net of the language labels used to refer to the subject, both involuntary (0.17, p < .001) and voluntary (0.36, p < .001) job training credentials increase positive perceptions of the subject as compared to the no credential condition. By contrast, language labels do not appear to have a positive impact on perceptions. While the professional language label is comparable to the person label, the person-first language label actually decreases positive perceptions of the subject in about the same magnitude as the involuntary job training credential increases them, essentially canceling out the positive credential effect for that combined condition. Figure 1 provides a visual depiction of the group means across all potential combinations of credential and language label conditions.

Figure 1. Predicted specific perceptions by randomized positive credential and language label conditions (n = 2775). FC: felony conviction.

We observe that appraisals are negative for the no credential condition regardless of the language used to describe the subject and tend to improve to varying degrees when respondents see an involuntary or voluntary job training credential. Notably, the involuntary credential only improves perceptions when used in combination with person language, and professional or person-first language negates the effect of this positive credential. Relative to the no credential condition, the voluntary job training credential improves positive appraisals across all language labels, but the effect is largest for the professional language condition. However, the person-first language condition, even when coupled with the voluntary job training credential, improves average perceptions to be just the sample average for the factor score.

Discussion

There is significant evidence that noun labels, such as “criminal,” can exacerbate stigma (Denver et al., 2017; Imhoff & Jahnke, 2015; Kelly & Westerhoff, 2010). Conversely, alternative language, which includes person-first terminology, can mitigate stigma. New research suggests that language that creates further separation between a person and their criminal-legal involvement, such as “returning citizen” instead of formerly incarcerated person, may be even less likely to trigger stigmatizing attitudes and more likely to elicit positive change (Jackl, 2023). Given the widespread and common use of criminal labels (Elderbroom et al., 2021), adjusting everyday language could potentially be a low-cost stigma mitigation strategy, and some researchers even suggest language priming could “lead to language change” at a population level by exposing people to, or reminding people of, different linguistic options (Pickering & Garrod, 2017, p. 173).

Building off prior labeling theory and research we explore how people use labels in the employment context, including testing whether language priming influences how people describe hypothetical job applicants with a conviction record and whether language priming has different effects when paired with positive credentials. There are two key takeaways in the current study: 1) positive credentials are more effective at influencing alternative language use than simply exposing people to that language, although 2) evoking other master-statuses through alternative labels (such as “applicant”) might further reduce stigma when compared to person-first language.

First, in Study 1, we coded how respondents described the vignette subject in open-ended responses into five categories: no code (a subject is never referenced), deviant or crime-first (e.g. “felon”), person (e.g. “this person”), person-first (e.g. “person with a conviction”), and professional (e.g. “applicant”). When viewing an applicant who presented a strong positive credential, respondents were more likely to proactively elect to use professional language to describe the applicant, although it is possible that this was the result of language priming. This finding was reiterated in Study 2, where respondents were more likely to use person-centered or professional language when viewing an applicant who had earned a positive credential. Importantly, Study 2’s design allowed us to test whether this effect was driven by the credential itself or language priming. The results suggest that positive credentials emphasizing non-criminal statuses or traits may lessen employers’ inclination to label a subject. More research is necessary to determine whether these credentials cause employers to also refrain from the essentialist thinking associated with the act of labeling, although initial research does suggest that positive credentials can influence perceptions of individuals with criminal records, including among employers (Denver & DeWitt, 2023).

Second, our results suggest that not all non-criminal language options are equal. In Study 2, where respondents were equally primed with person, professional, and person-first language options, those who were exposed to person-first language had less positive perceptions of the subject in the vignette. In addition, the beneficial impact of positive credentials disappeared when person-first language is used rather than other person-oriented language. Both results are in accordance with Jackl’s (2023) finding that the use of person-first language, which by definition includes a reference to the criminal record, may be more stigmatizing than neutral and potentially more ambiguous terms (e.g. “returning citizen”). These findings support the idea that once it is initially made clear—either in language or in context—that a person has a criminal history, it may be beneficial to refer to them using language that centers other master-statuses thereafter (e.g. applicant, student, parent) rather than continuing to reemphasize a criminal identity using person-first language.

There are several study limitations. To ensure sufficient statistical power in our analysis, we examined a limited number of language options and credential combinations. We also held the subject’s race, sex, age, and criminal history constant to focus on our key experimental conditions of interest. Future research could extend this initial effort to examine whether the findings vary by demographic characteristics, criminal record type, and the time since the criminal conviction occurred. For example, we were unable to test whether the language interventions used here may have different effects when applied to women or white men. We anticipated that testing alternative positive labels on a particularly stigmatized group (Black men with felony records) might uncover subtle differences in public perceptions by highlighting different identities or statuses (e.g. applicant) that could override criminal record stigma, but future research should explore whether alternative labeling is more successful with demographic groups that carry less stigma. In addition, using “identity-first” language (e.g. an autistic person) rather than person-first language (a person with autism) is an example of a broader movement to embrace the unique perspective of the labeled individual (Bury et al., 2023). Applied to the current context, terms like “justice-involved individual” or “convicted person” could achieve these same goals. Future research could expand upon the analyses presented here to examine the impact of identity-first compared to person-first language in addition to different label variations. Research in different contexts would also be valuable, as we focus strictly on employment and cannot speak to other social settings where criminal labels affect perceptions.

Additionally, critics of changing criminal labeling language express concerns that symbolic reform could take the place of more impactful system change. Adjusting words or phrases requires less investment than altering correctional culture, strengthening reentry and reintegration initiatives, or changing broader norms and institutional barriers related to criminal record stigma (e.g. Ortiz et al., 2022).⁸ In that sense, simply renaming a label can disguise harmful and problematic systemic behaviors and overshadow potentially more meaningful policy change (Cox, 2020). It is also possible that relabeling processes are ineffective; in the current study we find weak and inconsistent evidence of a mimicry effect, and there may be confusion about which language to use given the multitude of available label alternatives and lack of consensus regarding which is most appropriate. Convict criminologists have pointed out that those labeled—the very group purported to be helped by such policies—have been largely excluded from language policy conversations (e.g. Ortiz et al., 2022). Those impacted by the criminal legal system should have a leading voice in this discussion, and our goal is to provide empirical evidence to contribute to these policy conversations.

Ultimately, knowing a person has a criminal record might provide important information to decision makers and the public, and fully erasing any reference to it may not be realistic or preferrable in some situations. In those cases, the evidence overwhelmingly supports using person-first language instead of crime-first terms. However, avoiding repeated reminders that a person has a criminal record may also help ensure that one piece of information does not overpower other personal characteristics or life experiences. Using terms such as applicant, employee, relative, student, or person in subsequent references can shift the focus from the criminal record to other master statuses, and by extension, to other perceived characteristics. Yet the ability to share positive credentials appears to have a more powerful stigma mitigation effect than language alone. Therefore, pairing broader language policy reform efforts with tangible micro-level positive credentialing opportunities can help to redefine who the person is by displaying what the person has accomplished post-conviction.

Acknowledgments

We would like to thank our editor, Marvin Krohn, and three anonymous reviewers for their helpful feedback and suggestions.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 “Returning citizens” is formally used by city officials in Massachusetts (City of Boston, 2023); as another example, state officials have adopted the term “reentrants” as a replacement in Pennsylvania (Reentrant Handbook, 2019; see also PA DOC, 2023). Kearley et al. (2023) find evidence of juvenile justice stakeholders in Pennsylvania using “contradictory” terminology when discussing gangs, depending on the framing strategy the group is employing, providing additional evidence that alternative descriptors tended to be adopted purposefully.

2 It is also worth noting that language discussions in the criminal legal system extend to other groups, such as those working in corrections (e.g. “correctional officers” rather than “guards”), and recent work has shown that civilians were less likely to assign moral blame to police officers when a police killing is described using obfuscatory language (Moreno-Medina et al., 2022).

3 Specifically, the voluntary job training program referenced the “applicant” and “people,” and the involuntary job training program included the words “applicant” and “individuals.” The occupational license included “individual,” “barber,” and “individual with a felony conviction” (x2), while the reference letter discussed the “applicant” (x5) and “good worker.”

4 Although our initial study occurred prior to the “MTurk [data quality] crisis” (Chmielewski & Kucker, 2020; see also Kennedy et al., 2020), we fielded a pilot for Study 2 (n = 47) on MTurk in April 2022, which revealed extensive data quality issues and confirmed concerns reported in the academic community, leading us to switch platforms. See Appendix B for methodological decision details.

5 We opted to standardize race in the current study and selected a Black male applicant because Black males are disproportionately represented in violent state prison sentences (Carson, 2020). We selected the name Darnell Jackson, one of the five Black male names used in our original study, because it overlaps with names used in previous research (Agan & Starr, 2017; Gaddis, 2017). According to Agan and Starr’s (2017) vital records data, 93.4% of babies with the first name Darnell and 76.3% of babies with the last name Jackson are Black.

6 The exact item wording was as follows: I would be worried about this [label randomized] committing a workplace-related crime. I think this [label randomized] is likely to commit a crime in the future. I think this [label randomized] would be a trustworthy employee. I think this [label randomized] would regularly show up for his shifts and arrive on time. I think this [label randomized] has good moral character. The response options for the first four items were: “Strongly agree,” “Agree,” “Neither agree nor disagree,” “Disagree,” and “Strongly disagree.” Due to a survey design error, the response options for the good moral character question were slightly different (Strongly agree, Somewhat agree, Neither agree nor disagree, Somewhat disagree, and Strongly disagree).

7 While the key finding—the reference letter increases the use of professional language—is consistent in the employer subsample (b = 2.31, p < .001), the subanalysis also points to a new finding: the occupational license increases the use of deviant language (b = 2.28, p < .05). It is unclear why this would be the case for employers; it is possible the type of license we used (barber) is correlated with an underlying sentiment or assumption that we were unable to capture in this analysis, or given the various analyses we ran, it is possible this finding was due to chance. In either case, it is not a consistent result (for employer subsample results, see Appendix E).

8 For example, as Ortiz et al. (2022, p. 10) point out, “If academics were truly concerned about the impact of the word convict, instead of policing the language of convicts, academics would be demanding structural change in their individual institutions that would lead to reductions in the stigma surrounding the word convict.”

9 Some evidence indicates attention checks (also known as screeners or instructional manipulation checks) can have negative repercussions, including interference with the flow of survey and problematic respondent reactions. Removing respondents who fail attention checks can also be problematic because they are not randomly selected; paying attention is correlated with other attributes. At least among “reputable” or high quality MTurk workers, avoiding attention checks can be more beneficial (Paolacci et al., 2010; Peer et al., 2014).

10 These numbers are lower for the new experiment because the IPHub tool was integrated into the survey code, preventing most non-US respondents and VPS users from starting the survey; in the original survey, this process was completed after data collection was finished.

11 We asked, “In the past 5 years, have you taken a survey on Amazon MTurk that looks very similar to this one (includes a resume and asks you to make a hiring decision)?” Response options included (1) Yes, (2) No, and (3) Unsure. We kept respondents who were “Unsure” if they had taken a similar survey on MTurk within the past five years. Sensitivity analyses excluding these respondents indicate no difference in our main findings (results available upon request).

Appendix A:

Survey Language in Study 1

Survey Page 1—Instructions

For the first section, you will be asked to put yourself in the position of an employer tasked with deciding if an applicant should be called for an interview. The applicant is responding to a job advertisement for an entry-level full-time cashier position in a supermarket. This position requires no more than a high school degree and no prior work experience. On the next page, details will be provided about the applicant and application materials, and you will then be asked a series of questions about your perceptions of the applicant and the likelihood that you would select them for an interview. Please press the next page button below when you are ready to begin.

Survey Page 2—Applicant Details

Application Details

Applicant’s Name: [Randomization #1—10 possible conditions] Brent Baker/Todd McCarthy/Greg O’Brien/Brad Walsh/Brett Walsh/Darnell Jackson/Rasheed Jackson/Jamal Jones/Jermaine Jones/Tyrone Robinson

Age: 23-years old

Work History: 2.5 years of entry-level cashier experience at a supermarket and 1 year working in a barbershop

Highest Level of Education: High school diploma (obtained [3.5/5] years ago)

References: High school guidance counselor (known 5 years); Prior co-worker (known 18 months); Prior employer (known 1.5 years)*

All references have verified that the applicant’s work history is accurate.

*Prior employer reference is only shown if reference letter credential is randomly selected

Criminal Background Check: [Randomization #2—3 possible conditions] (1) No criminal record; Conviction for felony [(2) aggravated assault/ (3) possession (cocaine) with intent to distribute] for which they served 18 months in prison (occurred 3 years ago)

Additional Credentials**: [Randomization #3—5 possible conditions] (1) No additional credential; (2-3) Participation and completion of a job skills training program (completed post-conviction); (4) Occupational license (Barber) from the Department of State (obtained post-conviction); (5) Reference letter from a prior employer with verification of positive post-conviction work history ** Section not shown if the applicant has no criminal record

Survey Page 3—Additional Details and Definitions

Additional Details and Definitions***

Aggravated assault: Defined as an unlawful attack by one person upon another for the purpose of inflicting severe or aggravated bodily injury.

Drug possession with intent to distribute: Defined as the possession of an illegal narcotic or controlled substance with the intent to sell or otherwise distribute.

Completion of a job skills training program [Version 1—Voluntary]: Job skills training programs include training related to both hard and soft skills. Some examples of skills training include the ability to work under pressure, time management, conflict resolution, and self-motivation, among others. The specific vocational training the applicant received during the program was to prepare him to be a barber. It is well known locally that this program is voluntary—people choose to participate themselves.

Completion of a job skills training program [Version 2—Involuntary]: Job skills training programs include training related to both hard and soft skills. Some examples of skills training include the ability to work under pressure, time management, conflict resolution, and self-motivation, among others. The specific vocational training the applicant received during the program was to prepare him to be a barber. It is well known locally that this program is involuntary—local courts assign individuals to this program as a condition of their early release from prison.

Occupational License: An occupational license to be a barber is awarded by the Department of State after an individual has finished training under a licensed barber or an equivalent barber training program at a vocational school. Licenses are traditionally not available to individuals with felony convictions unless an individual with a felony conviction on their record can provide sufficient evidence to the Department of State that they currently possess good character.

Reference letter from prior employer: A prior employer who has known the applicant for 1.5 years wrote a reference letter for the applicant. The letter explains that the applicant was a good worker, showed up on time, and did not violate any company policies or workplace rules. Additionally, the employer states that they are aware the applicant has a felony conviction, and that they would rehire the applicant if they had the opportunity to do so. You have called the prior employer and have verified that they did write the letter.

***Provided in accord with experimental conditions. Details were not included for conditions the survey taker did not see.

Appendix B:

Survey Methodology Details

Study 1 and Study 2 both received IRB approval through the University of North Carolina at Charlotte (18-0310 and 22-1218, respectively). Participants viewed an electronic consent form before beginning each survey.

Respondent Eligibility Criteria

In both online surveys (Study 1 and Study 2), we specified that respondents needed a 95% approval rate on previous tasks, a minimum of 100 prior tasks completed (maximum of 1000), and United States residence.

Data Quality Checks

We used three quality check measures: screening respondents upfront based on approval ratings, checking IP addresses for suspected use of a virtual private server to mask respondents’ actual location, and examining “unusual comments” in open-ended questions (Chmielewski & Kucker, 2020, p. 466; Kennedy et al., 2020; Peer et al., 2014).⁹ After taking the three precautions, we did not detect data quality issues in our Study 1 MTurk sample.

Study 1 Sample Post-Data Quality Checks

Due to a bug in the survey programming code that presented a subset of respondents (n = 670) with an incorrect set of responses to the lead-in question for the qualitative responses included in this study, we remove those individuals. We conducted a series of independent samples t-tests on bugged/non-bugged responses and found no significant differences between these samples with respect to respondent demographics. Results are available upon request.

Then, starting with a sample of respondents who saw hypothetical job applicants with a criminal record ($N_1$ = 3864), we removed respondents identified as having low quality data. Specifically, we used an external tool (IPHub) and a user-written Stata function ‘checkipaddresses’ (Kennedy et al., 2020; Winter, 2019) to remove all respondents with duplicate IP addresses (n = 77), respondents from outside the United States (n = 124), and respondents flagged by IPHub as being suspected of using a virtual private server (VPS) to prevent identification of their true location (n = 55). Finally, we removed 47 low-quality responses. These included blank responses, single word responses (e.g. “Yes,” “Good,” “IDK”), and responses that consisted only of numbers.

Study 2 Sample Post-Data Quality Checks

Based on our respondent criteria (those who reported prior hiring or interviewing experience), close to 8000 people were active on the platform in the 90-day period prior to our survey launch. From a starting sample of individuals who began the survey ($N_2$ = 3367), we removed 88 observations with duplicate IP addresses, non-US respondents (n = 1), and all respondents flagged by the IPHub tool as using, or suspected of using, a VPS to access the survey (n = 9).¹⁰ We then exclude 273 respondents who responded “No” to having reviewed job applications and conducted interviews despite being flagged as employers by Prolific based on these same criteria. We also removed all respondents who indicated they had taken a similar survey on MTurk in the past five years (n = 142).¹¹ Finally, we removed nine observations that provided either no response (n = 3) to the open-ended question or an unintelligible/low-effort response (n = 6) to this question (e.g. “yes”, “Bad”, or “No comments”).

Data Outsourcing Concerns

Given recent concerns about data outsourcing (Enns, 2023), or the practice of survey companies recruiting respondents from a variety of panels, we asked Prolific if they outsource respondents from other survey companies. In their response to our query, the company confirmed that they only use their own participant pool. This avoids concerns about participants taking the survey through multiple platforms.

Appendix C:

Survey Language in Study 2

Survey Page 1—Instructions

For the first section of this survey, you will be asked to put yourself in the position of an employer tasked with making an interview decision about a job application. The application is in response to a job advertisement for an entry-level full-time cashier position in a supermarket. This position requires no more than a high school degree and no prior work experience.

On the next page, details will be provided about the application and you will then be asked a series of questions about your decision.

Please press the arrow button below when you are ready to begin.

Survey Page 2—Applicant Details

Application Details

Name: Darnell Jackson

Age: 23-years old

Work History: 2.5 years of entry-level cashier experience at a supermarket and 1 year working in a barbershop

Highest Level of Education: High school diploma (obtained 5 years ago)

References: High school guidance counselor (known 5 years); Prior co-worker (known 18 months) All references have verified that the work history information is accurate.

Criminal Background Check: Conviction for felony aggravated assault for which he served 18 months in prison (occurred 3 years ago). A felony assault is defined as an unlawful attack by one person upon another for the purpose of inflicting severe or aggravated bodily injury.

[Randomization: No credential/Voluntary Job Skills Training/Involuntary Job Skills Training. Note that this condition included labeling randomization: person/applicant/person with a felony conviction. Exact text below.]

No Credential

[No additional text]

Voluntary Job Skills Training

This [person/applicant/person with a felony conviction] also presented a certificate of completion for a VOLUNTARY JOB SKILLS TRAINING PROGRAM completed after their conviction. The program included training related to both hard and soft skills. Examples include the ability to work under pressure, time management, conflict resolution, and self-motivation, among other skills. The specific vocational training this [person/applicant/person with a felony conviction] received during the program was to prepare for a variety of entry-level positions. This [person/applicant/person with a felony conviction] chose to participate in this program after their early release from prison.

Involuntary Job Skills Training

This [person/applicant/person with a felony conviction] also presented a certificate of completion for an INVOLUNTARY JOB SKILLS TRAINING PROGRAM completed after their conviction. The program included training related to both hard and soft skills. Examples include the ability to work under pressure, time management, conflict resolution, and self-motivation, among other skills. The specific vocational training this [person/applicant/person with a felony conviction] received during the program was to prepare for a variety of entry-level positions. A local court assigned this [person/applicant/person with a felony conviction] to this program as a condition of their early release from prison.

Appendix D:

Excluding Respondents Using Multiple Language Types in Open-Ended Responses

Table D1. Multinomial logit predicting coded language type (Study 1, n = 2821).

Variables	Deviant			Person			Person-first			Professional
	RRR	SE	AME	RRR	SE	AME	RRR	SE	AME	RRR	SE	AME
Credentials
No credential (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Invol job training	1.06	0.31	0.00	0.90	0.16	−0.01	0.64	0.24	−0.01	0.99	0.16	0.00
Vol job training	0.69	0.23	−0.01	1.04	0.18	0.01	0.41*	0.18	−0.02	1.06	0.16	0.01
Occ license	1.31	0.37	0.01	1.31	0.22	0.03	0.89	0.32	−0.01	1.24	0.19	0.02
Reference letter	0.47	0.19	−0.02	1.29	0.22	0.01	0.82	0.31	−0.01	2.02***	0.30	0.12
Criminal record
Violent (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Drug	1.02	0.21	−0.00	1.22	0.14	0.02	0.88	0.23	−0.00	1.03	0.10	−0.00
Applicant race
White (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Black	1.22	0.25	0.01	1.09	0.12	0.01	0.78	0.20	−0.01	1.18	0.12	0.02

AME: average marginal effect; Invol: involuntary; Occ: occupational license; Ref: reference category; RRR: relative risk ratio; SE: standard error; Vol: voluntary.

Notes: The reference category for the coefficients reported in this table is the “No Code” language type. Marginal effects represent the change in probability to be in the selected language category, relative to the reference category for that regressor. Huber-White standard errors are reported.

*p < .05; **p < .01; ***p < .001.

Table D2. Multinomial logit predicting coded language type (Study 2, n = 2661).

Variables	Deviant			Person			Person-first			Professional
	RRR	SE	AME	RRR	SE	AME	RRR	SE	AME	RRR	SE	AME
Credentials
No credential (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Invol job training	1.05	0.25	−0.00	1.27	0.19	0.02	1.05	0.24	−0.00	1.35*	0.19	0.03
Vol job training	0.83	0.21	−0.01	1.52**	0.22	0.05	0.91	0.21	−0.01	1.10	0.16	0.01
Label language
Person (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Person w/FC	1.34	0.33	0.01	0.99	0.15	−0.00	1.00	0.23	0.00	0.93	0.13	−0.01
Professional	1.08	0.28	0.00	1.13	0.16	0.01	0.99	0.23	−0.00	1.32*	0.18	0.03

AME: average marginal effect; Invol: involuntary; Person w/FC: person with a felony conviction; Ref: reference category; RRR: relative risk ratio; SE: standard error; Vol: voluntary.

Notes: The reference category for the coefficients reported in this table is the “No Code” language type. Marginal effects represent the change in probability to be in the selected language category, relative to the reference category for that regressor. Huber-White standard errors are reported.

*p < .05; **p < .01; ***p < .001.

Table D3. OLS models predicting specific perceptions of the applicant (Study 2, n = 2661).

Variables	Coef.	SE
Credentials
No credential (Ref)	—	—
Invol job training	0.17***	0.05
Vol job training	0.35***	0.05
Label language
Person (Ref)	—	—
Person w/FC	−0.15**	0.05
Professional	−0.05	0.05
Model statistics
Intercept	−0.10
Root MSE	0.99
R-squared	0.02

Coef.: coefficient; Invol: involuntary; MSE: mean squared error; Ref: reference category; RRR: relative risk ratio; SE: standard error; Vol: voluntary.

Notes: Huber-White standard errors are reported.

*p < .05; **p < .01; ***p < .001

Appendix E:

Employer Subsample for Study 1

Table E1. Distribution of coded language types by randomized positive credential condition (Study 1).

Language	Full sample (n = 1408)	No credential (n = 350)	Positive credential (n = 1058)
No code	58.74%	61.14%	57.94%
Deviant	3.76%	3.43%	3.88%
Person	15.34%	16.29%	15.03%
Person-first	2.27%	2.29%	2.27%
Professional	19.89%	16.86%	20.89%

Table E2. Multinomial logit predicting coded language type (Study 1, n = 1408).

Variables	Deviant			Person			Person-first			Professional
	RRR	SE	AME	RRR	SE	AME	RRR	SE	AME	RRR	SE	AME
Credentials
No credential (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Invol job training	0.82	0.38	−.00	0.73	0.18	−.04	0.61	0.38	−.01	1.02	0.22	.01
Vol job training	0.96	0.44	−.00	0.79	0.19	−.03	0.96	0.53	−.00	1.18	0.25	.03
Occ license	2.28*	0.87	.04	1.21	0.27	.02	1.33	0.67	.00	1.03	0.23	−.01
reference Letter	0.62	0.37	−.02	1.27	0.30	.00	1.41	0.78	.00	2.31***	0.48	.14
Criminal record
Violent (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Drug	1.09	0.32	.00	1.09	0.17	.01	0.72	0.27	−.01	0.98	0.14	−.01
Applicant race
White (Ref)	—	—	—	—	—	—	—	—	—	—	—	—
Black	1.13	0.32	.00	1.35	0.21	.03	0.91	0.33	−.00	1.20	0.17	.02

AME: average marginal effect; Invol: involuntary; Occ: occupational license; Ref: reference category; RRR: relative risk ratio; SE: standard error; Vol: voluntary.

Notes: The reference category for the coefficients reported in this table is the “No Code” language type. Marginal effects represent the change in probability to be in the selected language category, relative to the reference category for that regressor. Huber-White standard errors are reported.

*p < .05; **p < .01; ***p < .001.