Job vacancies have long been predominantly published digitally on various websites. The dataset of the Swedish Public Employment Service, which has been collecting all job advertisements since 2006, currently contains 6.9 million ads. [1] The length of each ad can range from a few dozen words to nearly 1000. These descriptions are designed to both attract applicants and provide an accurate overview of the workplace, organisation, and position. Additionally, the ads usually include a list of requirements and desired qualifications and competencies that the applicant should have. The collected job advertisements are considered a unique source for gaining insights into trends and needs in the labour market. Interest in this type of labour market data has hardly diminished with the recent advancements in AI development. There is a strong demand for data-driven solutions/answers regarding skills supply, job matching, and labour market statistics. Historical job advertisements have received attention from researchers, statisticians, and job matching professionals alike. The fundamental question is how job advertisement data can enhance the functioning of the labour market. Equally important is how to ensure that this data does not negatively impact the labour market's functionality.

Today, there is significant knowledge about the linguistic content of job advertisements, such as which skills and abilities are relevant and associated with the advertised position. Using NLP (Natural Language Processing), machine learning, and AI, these components have been successfully extracted. The method relies on training data originally coded by employment officers and then further trained using machine learning.

The Undesirable Effects of Language

There is now a significant knowledge of how the design and linguistic content (wording) of job advertisements can have undesirable effects on the candidate pool. Studies have shown that the language used in an ad affects how likely or unlikely women are to apply for a job. [2] For instance, women are less inclined than men to apply if they perceive the job requirements to be to too high. Additionally, certain words can discourage women from applying. Studies have also investigated how an advertisement's depiction of an employer's commitment to inclusion and fairness influences the willingness of various minority groups to apply for the position. [3] Requirements for flexibility also result in fewer women applying for a position, despite their equal chances of securing the jobs if they apply. In other words, the wording of the advertisement impacts the result. In the study "Evidence that gendered wording in job advertisements exists and sustains gender inequality" (2011), the authors argue that masculine-coded words and phrases create an implicit bias or discriminatory effect, as opposed to explicit discriminatory statements (4).

Currently, there are seven grounds for discrimination in Sweden. Job advertisements intended for publication on the Swedish Job Board (Platsbanken) undergo check-ups to ensure they do not discriminate against certain groups. However, the screening process overlooks elements that may lead to indirect discrimination. At an aggregated level, this means that the inequality existing in the labour market persists, resulting in both economic consequences for different societal groups and contradicting with the Swedish Public Employment Service's directive (in Swedish: Förordning (2022:811) med instruktion för Arbetsförmedlingen) to "promote diversity and equality within its scope of operation”.

Risk of Discrimination

A study, analysing discrimination in German job revealed that while nearly none of the examined ads were explicitly discriminatory, one-fifth of them carried a risk of discrimination (5). By "risk of discrimination," referred to expressions or elements in the ads that could potentially exclude certain groups. Today, there are several available tools that automatically detect words and phrases that may lead to indirect discrimination and provide suggestions for neutral alternatives. However, the ads in the Historical Ads dataset have not been scrutinized with such a tool. Whether this scrutiny is desirable, it is not self-evident, but the data consumers must be aware that implicit bias exists and its potential impact on usage.