The use of AI in recruitment processes perpetuates gender inequalities when drafting CVs and assessing applications

Artificial intelligence (AI) is used in recruitment processes by both candidates and employers. Research based on audit experiments, published in PNAS, has found that large language models (LLMs) reinforce gender biases that accumulate throughout the various stages, from CV writing to shortlisting. These biases influence assessments and salary recommendations, reflecting structural inequalities in the labour market.

Expert reactions

Carlos Carrasco Farré - sesgo género IA

Carlos Carrasco Farré

Profesor de IA en la Toulouse Business School (Francia), miembro del equipo editorial de PLoS ONE (Ciencias Sociales) y doctor en Management Sciences (ESADE Business School)

Science Media Centre España

El principal resultado del estudio es que la IA funciona como un espejo del mercado laboral: penaliza a las mujeres en las profesiones donde ya son minoría y a los hombres en aquellas donde lo son ellos. Dicho de otro modo, la IA no es un árbitro neutral, porque ha aprendido de un mercado laboral que tampoco lo es. Además, me ha parecido interesante que esto explique por qué los estudios anteriores eran contradictorios: ambos efectos se compensan y el sesgo medio es casi nulo. En otras palabras, si solo miramos la media, el sesgo desaparece y eso es precisamente lo más interesante: una IA puede parecer justa en promedio y no serlo en ninguna profesión concreta. 

Sin embargo, los resultados deben interpretarse con cautela. Las diferencias son pequeñas: alrededor de un 2,7 % en la puntuación en las ocupaciones más segregadas y un 1,6 % de media en el salario recomendado (unos 79 dólares al mes sobre 5.000), muy por debajo de la brecha salarial real. Aun así, en un proceso de selección una diferencia pequeña puede decidir quién pasa el corte y quién se queda fuera: en las ocupaciones con menos de un 10 % de mujeres, la probabilidad de que una candidata supere el primer filtro es 4,8 puntos porcentuales menor. Esto es importante, no tanto por la magnitud del efecto, sino por el potencial masivo en el mercado laboral: una diferencia pequeña que la IA aplica de forma automática a miles de candidaturas (o potencialmente millones) se puede acumular generando un efecto muy grande. 

En cuanto a las implicaciones, me ha parecido especialmente interesante que pedirle a la IA que sea neutral no sirve de casi nada y pedirle que favorezca a las mujeres provoca una sobrecorrección enorme (más de un 20 % en la puntuación). Esto es importante porque, basándome en mi experiencia, muchas empresas confían en este tipo de instrucciones para corregir sesgos. Por el lado político/legislativo, en Europa, el Reglamento de IA clasifica los sistemas de selección de personal como de alto riesgo y las auditorías deberían hacerse profesión por profesión, porque una auditoría basada en promedios daría por bueno un sistema que no lo es.

The author has declared they have no conflicts of interest
ES

Josep Curto - sesgo género IA

Josep Curto

Academic Director of the Master's Degree in Business Intelligence and Big Data at the Open University of Catalonia (UOC) and Adjunct Professor at IE Business School

Science Media Centre Spain

Is it of good quality? Are the conclusions supported by sound data?

“As regards its quality, we can consider it to be of high scientific quality. It is no surprise that it has passed the rigorous peer-review process of PNAS (Proceedings of the National Academy of Sciences) and features a pre-registered experimental design (AEARCTR-0013538) inspired by classic correspondence studies in labour economics, now adapted to language models (LLMs). The results are supported by a very robust and well-structured dataset: 35,280 CVs generated across 98 real-world occupations (US SOC classification) and evaluated on six models, both proprietary and open-source (GPT 5.4, Claude Opus 4.6, DeepSeek 3.2, Mistral Large 3, Grok 4.1 and Llama 3.3 70B). From a data science perspective, it is particularly commendable that the authors do not use another LLM as a judge (thus avoiding circular biases), but rather expert-selected dictionaries, penalised logistic regression (Ridge) and cosine similarity against real job postings from Glassdoor, as well as publishing all the code and data on Zenodo to ensure replicability.”

How does this work fit in with existing evidence?

“This work provides a more holistic view. Until now, the scientific literature on generative AI had documented gender and racial biases in isolated tasks (for example, writing letters of recommendation or evaluating CVs in isolation). The focus here is on systemic risks.

This study goes a step further by demonstrating the cumulative or compound effect (adapting the theoretical framework of systemic discrimination proposed by Bohren et al.). In the real world, both candidates and HR departments use the same models: some to draft their CVs and others to screen them. The paper demonstrates that these two processes are not independent, but rather feed into one another.”

What impact does AI assessment bias have, according to this research?

“There are several points to note:

  • The study refutes the idea that AI penalises women uniformly. In reality, it acts as a structural mirror of the labour market: in male-dominated professions, female candidates receive lower employability scores (reducing their likelihood of passing the initial screening), whilst in female-dominated sectors they receive an equivalent bonus. An aggregated average analysis would conceal this problem as both extremes cancel each other out. This highlights the need to carry out an intersectional equity analysis.
  • In salary recommendations, AI assigns women lower salaries. This is mainly due to evaluation bias (rating the same profile more poorly solely on the basis of gender), exacerbated by generation bias (the model drafts CVs for women with more traits associated with non-promotable or support roles, and fewer traits geared towards high-visibility positions).
  • When the same AI ecosystem generates the CV and then evaluates it, the initial distortions are magnified rather than corrected. In other words, we observe one of the consequences of the phenomenon known as model collapse (where patterns are reinforced and the quality of the model deteriorates).”

Are there any significant limitations that need to be taken into account?

“Although the article is of high quality, five key limitations should be noted (the authors themselves acknowledge some of these):

  • It is a controlled audit experiment, not a field experiment. It abstracts from real human interaction: in a company, a recruiter may oversee the tool and a candidate usually edits the draft provided by the AI.
  • The study is conducted exclusively in English and uses occupational data from the United States. In languages with marked gender inflection, such as Spanish, the propagation of bias from generation to evaluation might behave differently (and even more intensely).
  • Salary outcomes are sensitive to how gender is indicated (explicit vs. proper name) and to the specific model (for example, GPT 5.4 exhibited atypical behaviour compared to the other five models when using proper names). Furthermore, the most recent versions of the models do not show a reduction in bias; in some cases, it is even greater.
  • In the baseline design, the models generate CVs without any real prior experience or education. Although the authors demonstrate in sensitivity tests that evaluation bias persists when real credentials from LiveCareer are included, generation bias decreases when the user provides detailed information from the outset.
  • It provides a snapshot in time, given the speed at which new models emerge. That does not mean it is irrelevant, since new models cannot generally be understood as an update to the previous one, but rather as a new artefact due to their internal structural changes. In this sense, they do not always lead to improvements, and it is often the case that they introduce new problems.”

What are the implications?

“From my perspective, the conclusions of this study send clear messages to organisations, developers and regulators:

  • Automating recruitment processes with generative AI does not guarantee objectivity; without specific audits, companies run the risk of automating and scaling up existing inequality under the guise of technological efficiency.
  • Instructing the model to ‘be gender-neutral’ is practically ineffective. Conversely, explicitly instructing it to favour the recruitment of women leads to massive overcorrection across all sectors, without distinguishing where there is actual under-representation.
  • Given that models are constantly updated and global averages are misleading, companies must implement granular audits by role and sector, make statistical adjustments during the screening phases, and incorporate various governance mechanisms, aligning with regulatory frameworks such as the European AI Act, which classifies recruitment systems as high-risk.
  • It is necessary to implement genuine governance safeguards to avoid token compliance.”
The author has declared they have no conflicts of interest
EN

Patricia Sánchez-Holgado - sesgo género IA

Patricia Sánchez-Holgado

Lecturer in the Department of Sociology and Communication at the University of Salamanca
Science Media Centre Spain

This study is highly relevant in the context of the use of artificial intelligence (AI) in the human resources and recruitment sector, particularly when examining gender inequalities in the labour market.

What is most interesting is that it presents an innovative approach, as it carries out an empirical analysis of the interaction between supply and demand – that is, job applications (AI-generated CVs) and the assessment of those applications using the same AI tools. It stands out because it does so on a large scale, as it evaluates, in a controlled environment (not in a real-world setting), six of the most advanced models (GPT 5.4, Claude Opus 4.6, DeepSeek 3.2, Mistral Large 3, Grok 4.1 and Llama 3.3 70B) across 98 real occupational categories from the US census, generating and evaluating over 35,000 CVs.

The key finding is that, as the same AI systems are used both to draft CVs and to assess them and recommend salaries, an algorithmic amplification occurs: the biases generated during CV drafting are subsequently reinforced when the AI assesses that same profile. In other words, it demonstrates that AI not only absorbs the stereotypes from its training data, but also amplifies them when used both to draft applications and to evaluate them. Along the same lines, with regard to gender, it is observed that AI does not penalise women uniformly, but rather replicates the segregation found in the real labour market. If only the applicant’s gender is changed, the models penalise women in male-dominated sectors and favour them in female-dominated sectors, including in relation to pay, meaning that gender pay gaps will persist.

The practical implications of this research are direct for companies, regulators and jobseekers. The automation of human resources does not guarantee objectivity, but rather carries a measurable risk of algorithmic amplification if the same systems filter the texts they previously helped to create. Statistical audits disaggregated by sector are necessary, as overall averages conceal the real inequalities.

The author has not responded to our request to declare conflicts of interest
EN

Mikel Galar - sesgo género IA

Mikel Galar

Lecturer in the Department of Computer Science and Artificial Intelligence at the Public University of Navarra

Science Media Centre Spain

This methodologically sound study examines gender biases in artificial intelligence models applied to recruitment processes. Its main contribution is to analyse how these biases can accumulate at different stages: first, when AI helps to generate a CV, and subsequently, when it is used to assess it. The results show that the models reproduce existing inequalities in the labour market, penalising women in male-dominated professions and favouring them in female-dominated ones (by around 3 per cent in each direction). As these two effects cancel each other out, on average the bias appears non-existent, demonstrating that relying on averages alone is insufficient when auditing these systems.

It is important to emphasise that the study demonstrates that the models reproduce gender biases and that these can accumulate, but not that their use increases discrimination compared with current recruitment processes. In fact, the pay gaps in the models’ recommendations (around 1.6 per cent) are smaller than those existing in the labour market, although the two figures are not directly comparable. Nor are there any simple solutions to avoid these biases: asking the model to be neutral changes nothing, and asking it to favour women results in overcorrection. Rather than proving that AI exacerbates workplace discrimination, the study identifies risks. The outstanding question is whether these systems discriminate more or less than the alternatives they replace.

The author has declared they have no conflicts of interest
EN

Pablo Haya Coll - sesgo género IA

Pablo Haya Coll

Researcher at the Computer Linguistics Laboratory of the Autonomous University of Madrid (UAM) and director of Business & Language Analytics (BLA) of the Institute of Knowledge Engineering (IIC)

Science Media Centre Spain

The article contributes through a series of experiments that identify gender biases both in the generation of CVs (generation biases) and in the assessment of candidates and salary recommendations (assessment bias) carried out by language models (LLMs). The results show that the LLMs evaluated reproduce existing gender stereotypes in the labour market.

These findings are relevant for candidates using AI, who need to refine their prompts so that LLMs do not reproduce gender-influenced labour market stereotypes. Companies must bear in mind that AI systems designed to assess candidates in recruitment processes are generally classified as high-risk systems under the RIA, which means they must be particularly careful when assessing biases in their candidate evaluation systems.

Finally, in my view, these audit studies should be implemented as evaluation testing platforms that enable the regular monitoring of AI systems. One-off assessments are insufficient given the constant changes in technology. The adoption of reproducible and comparable benchmarks would help to improve transparency, detect potential discriminatory effects and facilitate compliance with regulatory obligations.

The author has not responded to our request to declare conflicts of interest
EN
Publications
Journal
PNAS
Publication date
Authors

Elass et al.

Study types:
  • Research article
  • Peer reviewed
The 5Ws +1
Publish it
FAQ
Contact