Principal investigator: Omar Barroso Khodr
Authors: Friedrich and Eckhard (2026)
Original title: Inequality in frontline communication: bureaucrats talk differently to men and women
Location of the Intervention: Germany
Sample Size: 20.000 statements made by 28 bureaucrats
Primary Variable of Interest: complexity and emotionality
Type of Intervention: Impact assessment in public administration
Methodology: Empirical statistics; Multiple linear regression; Boot-strapping
Summary
According to the authors, gender biases in public service delivery are widely documented, but there is still a lack of empirical evidence on the underlying behavioral micro-mechanisms. This article seeks to contribute to filling this persistent gap by investigating gender differences in the complexity and emotional charge of verbal communication between bureaucrats and citizens.
The authors use a dataset consisting of 154 recorded dialogues from different local public services in Germany. Combining rule-based classification methods and machine learning, they analyze the differences in verbal administrative communication in 20.000 utterances.
The results demonstrate that no association was found between the gender of bureaucrats and their communication. On the other hand, the gender of citizens generates a significant difference: employees communicate in a more complex and emotional way when interacting with men. No differences related to gender correspondence were observed between the parties.
Being the first study to systematically examine implicit (gender) biases in the communication of bureaucrats, the article expands the current understanding of the micromechanisms of administrative inequality: the results contradict the expectations of gender socialization theory, confirm expectations associated with gender stereotypes, and challenge the idea that interactions within one's own group reduce stereotypical biases at the communication level.
- Policy Problem
The study identifies four central problems for public policies related to the persistence of gender inequalities in bureaucratic services. First, there is a gap between diagnosing inequality and understanding its mechanisms. Although gender differences in service delivery are widely documented, knowledge about the actual behaviors of bureaucrats that sustain them is limited. Existing literature describes attitudes and perceptions or tests biases through experiments, but lacks direct observations of real interactions between bureaucrats and citizens, hindering precise interventions.
Secondly, stereotypical biases persist in communication, even without discriminatory intent. The results show that bureaucrats, regardless of their own gender, communicate in a more complex and emotional way with male citizens than with women. This pattern, consistent with stereotypes of male competence and female dependence, operates subtly and unconsciously. The consequences are concrete: female citizens receive less information, understand their rights less, and feel less welcome – effects that are as detrimental to equity and institutional trust as an explicit denial of service.
Third, representational policies are insufficient to mitigate biases. Gender matching between bureaucrat and citizen does not reduce observed differences in communication. This challenges the premise of representative bureaucracy that demographic diversity among civil servants would lead to more balanced treatment. Investing exclusively in representative hiring does not eliminate disparities, as the mere presence of civil servants of the same gender does not alter communication patterns rooted in deep social stereotypes.
Fourth, communicative inequality generates systemic consequences. Biases in communication can produce: (a) direct discrimination in service outcomes, (b) unequal burden in access to services, and (c) erosion of public legitimacy, undermining trust in the political system. Communicative inequality affects not only the effectiveness of service delivery, but the very relationship between the state and society.
In summary, gender equity policies in public administration face implicit mechanisms that escape traditional control instruments. The absence of explicit discrimination does not equate to gender neutrality, and demographic representativeness does not guarantee equitable treatment. Managers need to develop refined monitoring of communication as an indicator of biases and implement interventions aimed at raising awareness and standardizing practices, rather than relying solely on staff composition or generic codes of conduct.
- Policy Implementation Context
This study is based on the premise that face-to-face interactions between bureaucrats and citizens are the primary channel for the emergence of gender biases in the provision of public services. These encounters influence judgments based on stereotypes, affecting service quality, citizen access, and institutional trust.
To investigate this phenomenon, the research analyzes verbal communication between the parties, focusing on two dimensions of the taxonomy of administrative speech acts: complexity (clarity and intelligibility of information) and emotionality (Expression of appreciation and engagement with the citizen). Both are linguistically measurable and reflect implicit biases of bureaucrats. Based on the literature, the study tests three sets of competing hypotheses:
Firstly, the gender of the bureaucrat – while some theories indicate that women communicate in a less complex and more emotional way (H1a), others suggest that institutional roles negate these differences (H1b).
Secondly, the citizen's gender – female citizens are expected to receive less complex communication because they are stereotyped as more dependent (H2a). Regarding emotionality, it may be higher for women (perceived as relational) or for men (seen as having neglected affective needs) – H2b.
Finally, the gender correspondence between bureaucrat and citizen – interaction between people of the same gender tends to reduce stereotypical biases (H3a), but this effect can be negated by the institutional role of the bureaucrat (H3b). In this way, the empirical design seeks to clarify contradictions in the literature and advance the understanding of the micromechanisms of administrative inequality.
- Evaluation Details
The study is based on a corpus Anonymized text composed of 154 conversations between bureaucrats and citizens, recorded between 2021 and 2023 in three types of local public services in Germany: two employment centers (JobcentersThe data encompasses 20.000 statements made by 28 bureaucrats, in addition to a questionnaire answered by 20 of them, thus covering different administrative contexts – from assisting the unemployed and providing guidance on public daycare centers to issuing official certificates and records.
According to the authors, the data collection process was extensive and challenging. The research team contacted numerous local administrations throughout Germany that provide daily services to citizens. Interested agencies received detailed information about the project; however, some withdrew due to concerns about resource shortages, employee protection, or data privacy.
In the institutions that remained in the study, informational meetings were held to recruit individual staff members. Employees who agreed to participate first completed an online survey and then received portable recording devices with which they recorded their interactions with citizens over one to three months, without a researcher present on-site – a measure adopted to minimize reactivity bias.
Citizen participation was voluntary and based on informed consent, documented in writing or in the audio itself. The devices were configured to prevent the recordings from being listened to, altered, or deleted after collection, ensuring the integrity of the records. At the end of the period, the files were sent to a professional transcription service and subsequently anonymized, with the removal of all names and references to physical locations. corpus The resulting text is completely anonymous.
Despite precautions taken to avoid behavioral biases, the authors acknowledge important limitations: it is not possible to determine the citizen participation rate in relation to the total number of services provided, nor the proportion of employees who participated in the study. Therefore, this is a convenience sample.
However, the researchers argue that the measures analyzed operate at a highly disaggregated linguistic level, which participants could hardly consciously manipulate – making observations of administrative behavior comparable to studies using body camera recordings of police officers, allowing for more robust inferences about the actual communication between bureaucrats and citizens.
- Method
The study operationalizes the dependent variables – complexity and emotionality – through natural language processing techniques applied to the utterances of bureaucrats. Complexity is measured by a dictionary-based score, derived from guidelines for plain language in German, which considers lexical, morphological, and syntactic features to assess ease of comprehension by the citizen.
Emotionality, in turn, combines nine machine learning-based classifiers (BERT GoEmotions model) and seven linguistic dictionaries, covering different emotions, with the aim of capturing the degree of commitment and appreciation conveyed in the bureaucrat's speech.
The measures are applied at the utterance level – defined as each turn of speech between changes of speaker. For each utterance, an additive index is calculated from the standardized values (z-score) of the individual characteristics. Most variables show a sparse distribution (between 50% and 90% of the values are zero), resulting in negative or near-zero scores for most utterances, with scattered positive values.
Illustrative examples show that highly complex statements exhibit convoluted syntactic structures and technical terminology, while high emotionality is marked by expressions of optimism, sadness, politeness, and agreement. The independent variables – gender of the bureaucrat, gender of the citizen, and gender correspondence – were extracted from questionnaires administered to employees and from linguistic markers (formal addresses and suffixes) present in the conversations.
Approximately 70% of the statements are from female bureaucrats. About 20% of the statements are directed to male clients with a gender match, 40% to female clients with a match, 10% to female clients without a match, and 30% to male clients without a match.
For statistical analysis, the study employs multiple linear regression models with individual utterances as the unit of analysis. Although the data have a hierarchical structure (utterances nested within conversations and bureaucrats), the low variance explained at these levels (2% to 8%) justifies treating the utterances as independent observations, with the inclusion of robust controls. Due to the asymmetrical and leptokurtic distribution of the dependent variables, bootstrapping is used as a non-parametric alternative for calculating confidence intervals.
The models include controls at multiple levels: at the utterance level, the timing of the conversation (relative position of the utterance) and the average complexity/emotionality of the citizens are controlled; at the conversation level, the total number of utterances is included; at the bureaucrat level, demographic characteristics (age, education, migratory experience, length of service), work attitudes (commitment to the public interest, job satisfaction), perceptions of stress and overload, self-efficacy, emotional exhaustion, aversion and empathic concern towards clients are considered; and finally, the type of public service (employment center, registry office or social assistance) is controlled.
Therefore, to address the lack of research data for some bureaucrats, two models are estimated for each hypothesis: one with the entire speech sample (without research variables) and another including the bureaucrat's variables, confirming the hypotheses only when there is statistical significance in both models.
- Main results
According to the authors, the results regarding the first hypothesis (the bureaucrat's gender) do not provide sufficient evidence that employee communication varies according to their gender. Regarding complexity, the coefficient for male bureaucrats was close to zero and not significant in the first model, becoming significant only after the inclusion of the bureaucrat's variables – which does not allow for robust confirmation.
In this context, regarding emotionality, a significant negative relationship was observed in the first model, but the effect disappeared in the second, becoming close to zero. Taken together, these findings corroborate hypothesis H1b, that there are no systematic differences in communication between male and female bureaucrats.
Regarding the second hypothesis (citizen gender), the results are consistent and statistically significant in both models tested. Male citizens receive moderately more complex administrative communication than female citizens (average difference of 0,25 to 0,26 points).
Similarly, male customers are the target of more emotional administrative language, with effects ranging from 0,25 to 0,60 points. Considering the sparse distribution of the variables, these effects are classified as moderate – noticeable and substantive, although not very broad. The findings, therefore, confirm hypothesis H2b, according to which female citizens receive moderately lower levels of complexity and emotionality in communication.
Next, the authors demonstrate that robustness tests reinforce these results: analyses restricted to conversations with mixed-gender participation, models with fixed effects per bureaucrat, and alternative specifications all point in the same direction. Furthermore, male clients experience significantly higher levels of both positive and negative emotionality, indicating that the effect is not limited to a specific tone.
Finally, the third hypothesis (gender matching) finds no empirical support. Subgroup analysis – separating male and female clients – shows that gender matching between bureaucrat and citizen is not associated with reductions in the observed biases. Although some coefficients proved significant in isolated models, none remained statistically significant in both specifications. Thus, the results corroborate H3b: for the levels of complexity and emotionality experienced by citizens, it makes no difference whether or not the bureaucrat shares the same gender as the client.
- Lessons in Public Policy
The study's results offer important lessons for the design and implementation of public policies, especially regarding equity in the service provided by frontline bureaucrats.
Firstly, the finding that the gender of the bureaucrat does not systematically influence communication with citizens is encouraging evidence for public administration. This suggests that recruitment, training, and professional socialization processes may be fulfilling their role in standardizing conduct and reducing the influence of personal characteristics in service delivery. For public managers, this indicates that investments in institutional training and human resource management practices have the potential to neutralize individual biases, promoting more uniform treatment among civil servants.
Secondly, the most relevant – and also the most worrying – finding is that the citizen's gender affects the quality of communication received. Male citizens receive more complex and emotionally charged service than female citizens. This pattern, consistent with gender stereotypes that associate men with competence and independence, and women with dependence and a lower need for technical engagement, reveals a structural bias that operates subtly and probably unconsciously among bureaucrats.
In this way, the implications are significant: citizens may be receiving less detailed information and less emotionally engaging service, which can compromise their understanding of rights, procedures, and benefits, as well as affect their perception of justice and their trust in institutions. For public policies, this points to the need for mechanisms to monitor the quality of communication that consider the profile of the citizen served, and not just the average performance of the service.
Thirdly, the absence of gender matching effects between bureaucrat and citizen challenges a premise dear to bureaucratic representation theory – that interactions between people of the same social group tend to be more empathetic and less biased. The findings suggest that, at the level of verbal communication, simply sharing a gender is not sufficient to mitigate stereotypical biases.
This indicates that diversity policies in the composition of the workforce, while relevant for other reasons, may not, in themselves, be an effective solution to eliminate disparities in service. Instead, managers should consider more direct interventions on communicative behavior, such as the adoption of standardized scripts, informational clarity checklists, and specific training to raise awareness about implicit biases.
Finally, the study demonstrates the value of computational linguistics-based methods for evaluating public policies. The ability to systematically analyze thousands of utterances and identify communication patterns that the actors themselves cannot perceive or report opens new possibilities for the fine diagnosis of inequalities in service delivery. Transparency and accountability policies could incorporate similar tools to audit the quality of communication in public services, allowing for more precise and evidence-based corrections.
