Introduction
Looking for work is the next step in the life of a recent graduate. And so I too started looking for a job some time ago. During this period, going into the relevant platforms on a daily basis, I noticed a significant number of ads that set out rather controversial terms. On the one hand, someone could argue that each employer is entitled to set their own conditions for the staff who will make up their business. That is perfectly reasonable. The problem starts when there are conditions that our staff must have certain innate characteristics, ones we are born with and (for the most part) cannot change. Sex, ethnicity, age, or even “beauty” genes are some of the relevant requirements we will read in quite a few ads. Obviously the above constitute serious forms of discrimination that still prevail even today.
This initial impression is also supported by the statistical data. Discrimination in the labour market does not concern only isolated ads, but is reflected in structural inequalities that are systematically recorded at the European level. According to Eurostat, in Greece three population groups consistently face less favourable terms in the labour market: women, people with disabilities, and foreigners. Let’s examine each one separately.
The gender gap in employment
Let’s start with the differences between the sexes, which are perhaps the most widespread form of labour inequality. We will look at the unemployment rates by sex and will find a persistent, and still ongoing, higher unemployment of women compared with men. Since 1998, unemployment among women has never come down to the same levels as that of men. There is, however, a positive trend: the difference between the two sexes has gradually decreased, from about 10 percentage points in 1998 to about 4 percentage points today.
It is worth pausing on what happened during the crisis. Between 2010 and 2013, unemployment shot up for both sexes, but the interesting thing is that the gap between men and women narrowed. This did not happen because women’s position improved, but because male unemployment rose sharply, particularly in sectors such as construction and manufacturing that are hit first in times of recession. After 2013, as the economy recovered, the gap started to widen slightly again, a reminder that the convergence during the crisis years was partly illusory.
Beyond the unemployment rates, Greece exhibits one of the largest gender gaps in employment across the whole EU. In 2024, the difference in the employment rate between men and women was 18.8 percentage points, the second largest in the Union, right after Italy. By contrast, in countries such as Finland, Lithuania and Estonia, the difference is below 3 percentage points.
The closing of the gap that we see in the left-hand chart is a positive development, but if we compare it with the EU average (right), it becomes clear that Greece started from a much worse position. In the EU, the difference between male and female unemployment has consistently stayed below 1 percentage point in recent years, while in Greece it remains close to 4. In other words, even if the trend is right, the pace of convergence is slow.
But how does Greece stand relative to the rest of the EU countries? The chart below shows the difference in unemployment between women and men (in percentage points) for each member state. Positive values mean that women’s unemployment exceeds that of men, while negative values indicate the reverse.
In the chart we observe that Greece (in red) is consistently among the top positions, together with Spain and Italy. It is no coincidence that these are Mediterranean countries with similar cultural characteristics: a strong role of the family, limited public childcare, and traditional patterns of the division of labour within the household. On the other side, Baltic and Scandinavian countries show a zero or even negative difference (i.e. lower female unemployment), something that is linked to extensive childcare networks, parental leave for both parents, and equality policies applied for decades.
Another point worth commenting on: in several Eastern European countries (Latvia, Estonia, Lithuania), female unemployment is lower than male. This is partly explained by the structure of their economies, where sectors with high male employment (industry, construction) are more vulnerable to cyclical fluctuations.
People with disabilities: out of the labour market
Another population group that is in a particularly disadvantaged position in the labour market is people with disabilities. Eurostat data for 2024 reveal a bleak picture: in Greece, roughly three quarters (75.7%) of people with disabilities aged 15 to 64 are outside the labour force. This share is almost 1.7 times higher than the EU average (44.5%), placing the country in the three worst positions Europe-wide, together with Bulgaria and Romania.
The problem does not stop there. Even for those people with disabilities who actively seek work, the likelihood of being trapped in long-term unemployment is disproportionately high. In Greece, 66.2% of unemployed people with disabilities have been looking for work for over a year. The corresponding European average is 40.6%. At the European level, the employment gap between people with and without disabilities amounts to 24 percentage points.
Behind these figures lie multiple barriers. Beyond employers’ prejudices, the Greek labour market largely lacks accessibility infrastructure: physical (buildings, transport), digital (websites, software), but also institutional (flexible hours, teleworking, supported employment). It is telling that Greece is ranked in the last positions on teleworking rates in the EU too, a form of work that, according to recent studies, noticeably improved the employability of people with disabilities after the pandemic in countries that adopted it widely.
| Indicator | Greece | EU average |
|---|---|---|
| Outside the labour force (15-64) | 75.7% | 44.5% |
| Long-term unemployment (of unemployed) | 66.2% | 40.6% |
| Unemployment gap (disabled vs not) | n/a | 3.6 pp |
In the chart, Greece’s distance (red) is evident, both from the European average (blue) and from the Scandinavian countries, where fewer than one third of people with disabilities remain outside the labour market. Finland, for example, achieves a rate of 28% thanks to a comprehensive system of supported employment, whereas in Greece such structures remain sporadic.
Bulgaria, which holds first place, is an extreme case: almost 9 out of 10 people with disabilities do not participate at all in the labour market. This reflects broader problems in the disability system of several Eastern European countries, where disability benefits often act as a disincentive to work rather than as a tool for inclusion.
Natives and foreigners: inequalities in access
The third dimension worth examining concerns the differences between native and foreign workers. The data show that Greece holds a particularly unfavourable position in the EU: in 2023, it recorded the lowest employment rate of foreign-born people in the Union. Correspondingly, unemployment among non-EU citizens is systematically higher in almost all member states.
It is worth noting a peculiarity: Greece was the only country in the EU where, in 2024, the highest unemployment rates concerned not third-country nationals but citizens of other member states. This can be explained in part by the structure of the migration flow towards Greece: quite a few European citizens (mainly from Bulgaria and Romania) work in seasonal or informal sectors, while a part of third-country nationals is employed in the agricultural economy or in shipping, sectors with relatively stable demand. In any case, the findings suggest structural difficulties of integration into the Greek labour market even for citizens with free access to it.
| Origin | Greece | EU average |
|---|---|---|
| Native (2 native parents) | n/a | 76.5% |
| Native (1 foreign parent) | 52.1% | 76.9% |
| Native (2 foreign parents) | n/a | 74.0% |
| Foreign-born | Lowest in EU | 69.9% |
The table reveals a striking imbalance: in Greece, natives with one foreign parent show an employment rate of just 52.1%, while the corresponding European average exceeds 76%. This means that even the “second generation” of migrants faces significant integration barriers in our country, something that cannot be explained solely by linguistic or cultural obstacles, since these are people who grew up in Greece.
Age discrimination: the invisible “expiry date”
A fourth dimension of inequality that runs through the Greek labour market (and is particularly visible in the ads themselves) concerns age. The problem manifests itself at the two ends of the age scale: in the young who cannot get in, and in the older who cannot get back in.
At one end, youth unemployment (15–24) in Greece remains among the highest in the EU. In December 2024, unemployment in that particular age group was 21%, significantly reduced compared with the 33.2% of 2019, but Greece was nonetheless ranked in the 4th-worst position in the EU. It is worth noting that the difference between the youth unemployment rate and the youth unemployment ratio exceeds 10 percentage points in Greece, which indicates that many young people simply do not participate in the labour force at all, either because they are studying, or because they have given up the search.
At the other end, workers aged 55–64 face a different but equally serious problem. Greece belongs to the EU countries with an employment rate of older workers below 55%, whereas in countries such as the Czech Republic, Denmark, Germany and Sweden the corresponding rates exceed 70%. In 2023, the employment rate in the 55–64 age group in Greece was just 54.1%. At the same time, Greece exhibits the highest rate of long-term unemployment in the EU (5.4%), a phenomenon that disproportionately hits older unemployed people: if someone loses their job after 50, the likelihood of finding a new position drops dramatically.
Behind these figures lies a mix of factors. On the demand side, employers often treat older workers as “expensive” or “hard to adapt”, preferring younger candidates who can be paid less and trained “from scratch”. On the supply side, Greece lags significantly in lifelong-learning and vocational-retraining programmes, tools that in northern countries allow workers to remain competitive at older ages.
Age discrimination, unlike sex discrimination, is rarely recognised as such. Ads asking for a “person up to 35 years old” or a “young employee” are common practice, without raising the reactions that a corresponding reference to sex or ethnicity would provoke. However, setting age limits without objective justification constitutes a form of discrimination both under European law (Directive 2000/78/EC) and under Greek law. The difference is that, in practice, enforcement remains minimal.
In the chart we can see that youth unemployment (15–24) always follows a much more extreme course: it shot up to 58% during the crisis and, despite the significant reduction, it remains consistently triple that of the other age groups. The 55–74 group shows a lower unemployment rate, but this is partly misleading: many older workers who lose their job are not recorded as unemployed but withdraw entirely from the labour force, not appearing in the unemployment statistics.
Age discrimination, then, works as a double filter: it closes the doors on the young who do not yet have “experience” and, at the same time, on the older who are considered to have “passed” a vague expiry date.
| Indicator | Greece | EU average |
|---|---|---|
| Youth unemployment 15-24 (2024) | 21.0% | 14.9% |
| Employment 55-64 (2023) | 54.1% | 65.2% |
| Long-term unemployment (2024) | 5.4% | 1.9% |
The overall picture
The above data compose a picture that leaves little room for doubt. The Greek labour market continues to operate with significant structural inequalities. Women consistently face higher unemployment than men, with a gap four times the European average. People with disabilities remain, by an overwhelming majority, outside the labour market. And foreigners, even of the second generation, find closed doors to a degree we do not encounter anywhere else in the EU. And age works as a double filter of exclusion: the young cannot enter the labour market with unemployment rates triple the average, while older workers, if they lose their position, are trapped in long-term unemployment or withdraw entirely from the labour force.
These differences are not merely statistical anomalies. They translate into real exclusions and become visible in the ads themselves: ads asking for a “young girl with a presentable appearance”, ads that set age limits without any justification, or that silently exclude those who do not match a narrow profile.
On this basis, let us now turn our gaze to the ads themselves and to what they reveal about hiring practices in Greece.
Characteristics of the ads
Let us now examine what the ads themselves reveal. Making use of data collected through scraping from the popular job-search platform xe.gr, we can form a picture of the Greek labour market, at least as it is reflected in the ads. The sample includes 5,093 ads and we will examine three basic characteristics: the working hours, the contract type, and the work mode (physical presence or teleworking).
Working hours
Starting with the hours, the picture is unambiguous: 88.1% of ads concern full-time positions, just 8.1% offer part-time work, and a small 3.9% do not specify the hours at all.
The near-absence of part-time work is not innocent. In countries with higher rates of female participation in the labour market (e.g. the Netherlands, Denmark, Germany), part-time work acts as a bridge to integration, especially for parents, students or people with disabilities. In Greece, the limited supply of such positions may reinforce the exclusion of precisely those groups we examined in the introduction. When the market offers almost exclusively full-time work, those who cannot meet this model are left out.
Contract type
On the contract type, the finding that stands out is not the distribution between permanent and fixed-term contracts, but the lack of information. 77.1% of ads do not specify the contract type at all. Of the rest, 21.1% mention a permanent contract and just 1.9% a fixed-term one.
This opacity is not merely a technical gap. For a candidate seeking job stability, the absence of information acts as a deterrent or, worse, leads to unpleasant surprises after hiring. It is worth wondering whether this practice is accidental or whether it reflects a conscious strategy: ambiguity over the terms of work favours the employer, leaving the candidate in a position of bargaining weakness.
Work mode
The most striking finding perhaps concerns the work mode. 98.3% of ads require physical presence. Teleworking represents just 0.6% and the hybrid model 1.1%. In other words, in a sample of over 5,000 ads, fewer than 90 offer some form of flexibility in the place of work.
This finding is directly linked to what was mentioned in the introduction about people with disabilities. International studies demonstrated that teleworking, as it developed during the pandemic, noticeably improved the employability of people with mobility or other difficulties in countries that adopted it widely. In Greece, this option essentially does not exist. The almost universal requirement of physical presence acts as a form of indirect discrimination: it does not explicitly exclude anyone, but in practice it closes the door on those who face difficulties with movement or accessibility.
What the characteristics tell us overall
If we combine these three findings, the picture that emerges is of a fairly rigid labour market: full hours, physical presence, and opacity over the terms. This rigidity does not hit everyone equally. It disproportionately hits women who seek flexibility because of family obligations, people with disabilities who need alternative forms of work, and young people or foreigners who could start out through part-time or remote work.
With these data in hand, let us now turn our gaze to the methodology of the experiment, examining how we can systematically detect discrimination within the content of the ads.
Experiment methodology
In the previous sections we saw that the Greek labour market exhibits significant structural inequalities across four axes: sex, disability, ethnicity and age. These data come from macroeconomic indicators and labour-force surveys. The question now posed is a different one: can we detect these forms of discrimination within the ads themselves? And if so, what is their extent?
The trigger and the research question
The trigger for this analysis was a simple observation: browsing job-search platforms, I encountered, on a daily basis, ads with controversial terms: age limits without justification, references to sex or appearance, requirements unrelated to the nature of the position. These observations led to the central research question: to what extent are practices that constitute (directly or indirectly) discrimination expressed in the text of job ads?
Unlike the study of individual axes, the analysis attempts to cover several forms of discrimination simultaneously: age preference, sex, appearance. At the same time, it attempts to distinguish whether the age references reflect genuine employer preferences or are the result of state policies (e.g. subsidised DYPA programmes that target specific age groups).
Why scraping and not a questionnaire
We could study hiring practices by distributing questionnaires to employers. This approach, however, suffers from a fundamental problem: no employer would openly declare that they set age criteria or that they prefer “presentable girls”. Social desirability bias makes questionnaires unsuitable for measuring behaviours that the respondent themselves recognises as problematic.
The alternative (and the one adopted) is scraping data directly from ads. The logic is simple: ads are revealed preferences. The employer freely writes what they are looking for, without filtering their answers as they would in a survey. If an ad states “age up to 35”, this is not a value judgement of a researcher; it is the employer’s own intention, recorded publicly.
This methodological approach has parallels in the international literature. The correspondence study technique, used, among others, by Bertrand & Mullainathan (2004), also exploits real ads as a field for measuring discrimination, although with a different mechanism (sending fake CVs). Our approach focuses on the content of the ads rather than the responses to them.
Data collection
The data were collected from the platform xe.gr (Chrysi Efkairia), one of the largest job-search platforms in Greece. The scraping was carried out on 2 February 2025 using the R language and the rvest and httr packages. The process lasted about 4 hours and covered 255 pages of results (20 ads per page).
The final sample consists of 5,093 ads, with a posting range of 15 days (16 January – 2 February 2025). For each ad the following details were collected:
| Variable | Description |
|---|---|
| Job category (spc) | The job title/category |
| Work type (type) | Hours, work location, contract type |
| Wage (wage) | Monthly wage (where stated) |
| Region (loc) | Geographic region |
| Experience (exp) | Minimum required experience |
| Description (descr) | Free-text ad description |
| Posting date (p_date) | Posting date |
The data are freely available via GitHub Releases for reproducibility.
Text preprocessing
The analysis of free text in Greek presents particular technical challenges. The Greek language is characterised by cases, accentuation and a variety of endings, which means that the same word can appear in multiple forms (e.g. “νέος”, “νέα”, “νέων”, “νέο”, “νέους”). If this is not addressed, keyword detection will significantly underestimate the extent of the phenomenon.
The preprocessing was carried out in two phases: first character normalisation and then linguistic analysis through lemmatization.
Character normalisation
As a first step, each description text was converted to lowercase (str_to_lower()) and the accents were removed through Unicode normalization (stri_trans_general("NFD; [:Nonspacing Mark:] Remove; NFC")). This ensures that words such as “Νέος”, “νέος” and “νεος” are treated as the same. At the same time, the type column was decomposed into three separate variables (hours, presence, contract type) and the dates were converted from Greek format into a standard Date.
Lemmatization
Removing accents is not enough to address the problem of cases. The word “νέος” appears in an ad as “νέα”, “νέο”, “νέων”, “νέους”, “νέες”, six different forms that must be recognised as one and the same lemma. For this purpose, lemmatization was applied using the udpipe package with the pre-trained greek-gdt model, which is based on the Greek Dependency Treebank.
After lemmatization, the keyword search is done on the descr_lemma column, where each word has been reduced to its lemma form. Thus, instead of searching with regex for multiple inflections of a word (e.g. νε[οα]ς?|νεων|νεους|νεες), it suffices to search for the lemma νεος.
Lemmatization in action
The sentence «Ζητούνται νέοι υπάλληλοι με εμπειρία σε νεανικό περιβάλλον» (“Young employees wanted, with experience in a youthful environment”) is transformed into «ζητώ νέος υπάλληλος με εμπειρία σε νεανικός περιβάλλον», allowing the detection of both νεος (age reference) and νεανικος (indirect age reference) with a simple string match.
However, the lemmatizer is not perfect, particularly in informal ad texts it can fail on spelling mistakes or unusual inflections. For this reason, the keyword lists also include some additional inflected forms as a safety net.
Method A: Keyword detection
The core of the first method is based on detecting keywords within the lemmatized text (descr_lemma) of each ad. For each ad, five Boolean variables were created, covering four axes of possible discrimination and one institutional factor:
| Variable | What it detects | Indicative keys (lemmas + fallback) |
|---|---|---|
isAgeism | Age discrimination | νεος, νεαρος, νεανικος, ετων, ηλικια, φρεσκος αποφοιτος |
isGenderBias | Gender discrimination | κοπελα, κοριτσι, κυριος, κυρια, αρρεν, θηλυ, αντρας, γυναικα |
isAppearanceBias | Appearance criterion | ευπαρουσιαστος, εμφανισιμος, παρουσιαστικο, καλη εικονα |
isPolicy | State programmes | οαεδ, δυπα, ανεργος, επιδοτηση, κοινωφελης, 21-29 |
isRacialBias | Ethnicity / origin | υπηκοοτητα, καταγωγη, αλλοδαπος, μονο ελλην |
| Axis | Ads (n) | Share (%) |
|---|---|---|
| Age | 760 | 14.9 |
| Gender | 355 | 7 |
| Appearance | 38 | 0.7 |
| Policy | 51 | 1 |
| Ethnicity | 31 | 0.6 |
The keyword method is fully reproducible, transparent, and does not depend on external models. However, it is subject to two significant limitations. First, it produces false positives: the word “νέα” may refer to a “new position” rather than a “young person”. Second, it is unable to detect indirect or paraphrased discrimination, e.g. “dynamic profile” as a substitute for age preference, or the use of only the feminine gender in the job title without an explicit reference to sex.
Method B: LLM classification
To overcome the limitations of the keyword method, each ad was also classified by a large language model (LLM). The classification was carried out through the R package ellmer, using OpenAI’s gpt-4.1-nano model via structured output.
Prompt design
The LLM receives as input the full text of each ad inside <αγγελία>...</αγγελία> tags and produces a structured output with five Boolean fields (one for each axis of discrimination) and a short justification. The system prompt includes clear rules of interpretation:
- A reference to sex through grammatical gender (e.g. “saleswoman wanted”) counts as gender discrimination, unless the position naturally requires a specific sex.
- Expressions such as “dynamic environment” or “energetic person” do not constitute age discrimination, unless combined with a clear age reference.
- The requirement of knowledge of the Greek language is not racial discrimination; the requirement of a specific ethnicity or nationality is.
- References to DYPA/OAED are noted as policy-related even if they coexist with age discrimination.
Execution
The classification was run in batches of 500 ads, with intermediate progress saved so that recovery is possible in case of interruption. In total, all 5,093 ads were classified. The results are freely available via GitHub.
Advantages and limitations of the LLM
The LLM method is qualitatively superior: it understands context, detects indirect discrimination, and avoids false positives like the ones the word “νέα” produces in the keyword method. However, it is not fully deterministic; the same ad could be classified slightly differently on a second run. Moreover, the classification depends on the wording of the prompt and on the “decisions” of an opaque model.
What we measure and what we don’t
Both methods detect explicit or semi-explicit references to discrimination criteria. They constitute a lower bound of the real phenomenon:
- Silent discrimination: An employer may prefer younger candidates without writing it; the exclusion happens at the CV or the interview.
- Platform coverage: The data come only from xe.gr. Ads on LinkedIn, in newspapers or by word of mouth are not covered.
- Temporal coverage: This is a 15-day snapshot; it does not capture seasonal fluctuations.
The use of two independent methods (rules + LLM) works as an internal triangulation: the points of agreement strengthen confidence in the findings, while the points of disagreement highlight the limits of each approach.
Results
In this section the results of the analysis are presented at three levels. First, we compare the two detection methods (keywords and LLM) in order to assess their reliability. Then, we examine each axis of discrimination separately, highlighting the sectoral and geographic patterns that emerge. Finally, we compose an overall picture, examining how many ads exhibit multiple forms of discrimination simultaneously.
Throughout the section, we use the LLM method as the main reference (since it provides more accurate categorisation thanks to its understanding of context) and the keyword method as a point of comparison.
Comparison of detection methods
Before proceeding to the analysis of each axis, it is worth examining how far the two methods agree with each other. The table below presents, for each axis of discrimination, the share of ads each method detects, the share of agreement between them, as well as the number of cases detected by only one method.
| Axis | Keywords (%) | LLM (%) | Agreement (%) | LLM only (n) | KW only (n) |
|---|---|---|---|---|---|
| Age | 14.9 | 8 | 87.7 | 137 | 490 |
| Gender | 7 | 20.7 | 78.2 | 903 | 206 |
| Appearance | 0.7 | 0 | 99.2 | 1 | 38 |
| Policy | 1 | 9.2 | 90.9 | 440 | 22 |
| Ethnicity | 0.6 | 0.5 | 99 | 21 | 28 |
From the table and the chart, certain interesting patterns emerge. On the axes of appearance and ethnicity, the two methods show high agreement, something expected, since the corresponding expressions (e.g. “presentable”, “of Greek nationality”) are relatively unambiguous. By contrast, on the age axis the divergence is greater: the keyword method tends to overestimate, mainly due to false positives from the polysemy of words such as “νέος” (which can mean “recent” rather than “young”), while the LLM detects indirect references (e.g. “first work experience”) that the keyword method misses. On the sex axis, the LLM detects more cases, probably because it recognises the use of grammatical gender as indirect discrimination, a point that keywords do not fully cover.
The divergence between the methods is not a weakness; on the contrary, it highlights the blind spots of each approach and strengthens confidence in the points where they agree.
Age discrimination
The age reference is the most frequent form of explicit discrimination in the sample. However, before assessing its extent, we must answer a crucial question: do the age references reflect genuine employer preferences or are they the result of institutional incentives?
The role of state programmes
The Greek state, through DYPA (formerly OAED), subsidises employment programmes that target specific age groups (e.g. 21–29 years old). Ads that refer to such programmes do not necessarily express a prejudice of the employer but a response to institutional incentives. However, the two variables are not mutually exclusive; an ad may refer to a DYPA programme and at the same time ask for a “young profile”. This overlap may indicate that some employers use the state programmes as an institutional cover for age preferences that would exist anyway.
The existence of ads that belong to the “Both” category (that is, they refer to a state programme but at the same time express an age preference beyond the one the programme imposes) is a significant finding. It suggests that the institutional incentives, although designed to strengthen youth employment, may operate in parallel as a mechanism for legitimising age criteria in hiring.
Sectoral distribution
Age discrimination is not distributed uniformly. Certain job categories show disproportionately high rates of age references. The chart below focuses on the categories with at least 20 ads, so as to avoid distortions from very small samples.
The distribution reveals an interesting pattern: the sectors with the highest frequency of age references tend to be those that do not require specialised qualifications, positions where an employer “preference” for younger candidates cannot be justified by technical requirements. This is consistent with the view that age discrimination is primarily a cultural phenomenon (a reflection of stereotypical notions of a “suitable age”) rather than a result of objective operational needs.
Gender discrimination
Gender discrimination is the second most frequent form of explicit discrimination in the sample. A significant part of these references is not necessarily hostile; many ads simply use grammatical gender (e.g. “saleswoman wanted” or “waiter wanted”). This, however, is not innocent: the exclusive use of the feminine gender for a secretary or the masculine for transport reflects (and reproduces) a silent belief that the position “belongs” to a specific sex. The central question is not only how many ads make a reference to sex, but in which sectors these references are concentrated.
The distribution is not random. The sectors that show the highest frequency of references to sex tend to be those with strong gender stereotypes: sales, hospitality, secretarial support and care. This confirms the view that occupational segregation is not merely a result of individual choices, but is actively reproduced through hiring practices themselves.
Geographic distribution of discrimination
Discrimination is not distributed uniformly across geographic space. Certain areas show higher rates of explicit references, which may reflect differences in the structure of the local labour market, in the size of businesses, or in awareness around equality issues.
The geographic distribution must be interpreted with caution. The differences between areas may be due to the composition of sectors (e.g. tourist areas with many hospitality positions), to the size of businesses (smaller businesses tend to use a more informal linguistic style), or to cultural factors. Given that our sample covers only 15 days, the geographic differences should be confirmed with larger time windows.
The overall picture: multiple forms of discrimination
Having examined each axis separately, it is worth turning our gaze to the overall picture. How many ads exhibit at least one form of discrimination? And how many combine multiple ones?
For this calculation we exclude the ads that refer only to state programmes (with no other indication of age preference), so as to focus on genuine employer practices.
In total, 26.9% of ads exhibit at least one indication of discrimination (including state programmes). If we exclude the ads concerning exclusively state programmes, the share works out at 25.9%. Even this “clean” share constitutes a lower bound; it reflects only the explicit references and does not include the silent discrimination that takes place in the later stages of the hiring process.
Ads with two or more axes of discrimination are of particular interest, since they reveal the intersectional nature of certain hiring practices. A candidate who is a woman, older in age, or without a “presentable appearance” is excluded simultaneously by multiple criteria, and the cost of that exclusion is cumulative.
Indicative ad examples
Statistics acquire meaning only when accompanied by concrete examples. The table below lists excerpts from real ads (anonymised as to the employer) together with the classification each method gave. The examples were selected so as to highlight three categories: cases of agreement between the methods, false positives of the keyword method, and cases detected only by the LLM.
| Category | Sector | Excerpt | KW: Age | LLM: Age | KW: Gender | LLM: Gender |
|---|---|---|---|---|---|---|
| LLM only | Logistics / warehouse | WAREHOUSE WORKER – driver wanted by the company Deco Idea, which operates in kitchen furniture, sanitary ware and tiles in Markopoulo, Attica; full-time position, five-day week, permanent contract. Category B driving licence... | — | ✓ | — | — |
| LLM only | Secretarial / office staff | Our company in Anavyssos is looking for an employee for the position of Car Rental Clerk, who will take on the following responsibilities. - Reception and customer service at the rental office. - Management of bookings and contracts... | — | ✓ | — | — |
| False positive (KW) | Salespeople | Why become a member of our team? Because it's not just a job in Retail, it's your story! At Public Group we believe in the power of our people in a sustainable way. That's why we focus not only on what you do, but on who you can... | ✓ | — | — | ✓ |
| LLM only | Manicurists | Nail technician (f), 23–35 years old, with experience and excellent knowledge of semi-permanent and artificial nails and prior work experience; serious inquiries only. Required qualifications: excellent customer service, communication skills, willingness to work. A satisfactory... | — | ✓ | — | — |
| LLM only | Salespeople | Saleswoman wanted for a mini market in Moschato, up to 50 years old. Satisfactory salary, full-time. Rotating shift. For more information you can contact us by phone. | — | ✓ | — | — |
| False positive (KW) | Labourers | A women's clothing company in Neo Irakleio (near the ISAP station) is looking for staff for internal 2-month work as assistants in production and quality control, with flexible hours. Candidates can send their CV via... | ✓ | — | — | — |
The examples highlight the advantages of each method. The keyword method flags every occurrence of words such as “νέος/νέα”, even if they refer to a “new position” or a “new store”, false positives that the LLM avoids thanks to its understanding of context. Conversely, the LLM detects indirect age references (e.g. “first work experience”, “fresh graduate”, “dynamic environment” combined with an age reference) that the keyword method does not cover by definition.