The headline result looked fine
A top-line check of Barcelona Activa's recruitment data would have found little to worry about. Women represented 51.5% of registered candidates and 49.8% of people recorded as hired. The difference was not statistically significant.
A new independent audit argues that this answer is technically neat and practically incomplete. It followed nearly 497,000 candidate-vacancy records through seven stages of a public employment pipeline, from an employer's vacancy to the final handoff of shortlisted people.
Once the researchers broke the data down by salary, contract type, age and gender identity, five gaps appeared that the aggregate result did not show.
The point is bigger than one platform. A hiring process is not a single score. It is a chain of data, software, search choices, human judgement and employer decisions. Fairness can change at every link.
The gaps appeared where the jobs changed
For low-salary roles, women and men were shortlisted at similar rates. In the €15,000 to €24,000 band, women were shortlisted 7.43% of the time, compared with 9.45% for men. The disparate-impact ratio was 0.786, below the 0.80 benchmark used by the study.
Full-time work showed a wider difference: 9.00% for women and 11.87% for men. Part-time roles did not show the same adverse impact. Within the 46–55 age group, the rate was 10.4% for women and 13.3% for men.
The study also found that women's profiles were matched to vacancies with an average minimum salary €1,147 lower than men's. The gap persisted in 15 of 20 sectors and was statistically significant in 12.
These are associations in operational data. They do not prove that one algorithm or one person caused the differences. They do show why a single final hiring ratio can miss what happened before the result.
Two findings need particular care
The dataset contained 285 candidates recorded as non-binary or other. Their shortlisting rate was 3.51%, compared with 11.89% for men, and none were recorded as forwarded to an employer during the period studied.
The difference was statistically significant, but the sample was tiny beside the two binary-gender groups, each with more than 240,000 records. The author labels the result indicative, not conclusive. The exact size of the disadvantage is uncertain.
People aged 55 and over were effectively absent from the pipeline even though they represented 15.6% of Barcelona's active labour force. The paper cannot tell whether that came from registration, search practices, employer requirements or a problem in how the data was cleaned.
That uncertainty is itself useful. A system cannot be audited well if the records do not show who disappeared, why a filter was chosen or what happened after a shortlist left the agency.
The agency could not see the whole system
Barcelona Activa's analysts translated employer requests into filters and keywords, searched the third-party TalentClue platform, reviewed the returned profiles and built shortlists. Employers made the final decisions.
According to the paper, the agency and audit team did not have access to TalentClue's matching and ranking logic, its signal weights or internal fairness evaluations. The system also kept no useful record of analyst keywords, excluded candidate pools or reasons for manual inclusion and exclusion.
This leaves the audit unable to divide responsibility cleanly between candidate data, employer requirements, the vendor's software and human judgement. The study is observational. It also lacks qualifications, skills and work-experience variables; 24% of records have no gender information and 14.5% have no origin information.
The lesson is not that software alone discriminated. It is that the organisation using the software lacked enough visibility to know where the outcome came from.
What is confirmed, found and still open
Confirmed: the paper was submitted on 13 August 2026 by Gemma Galdón-Clavell of Eticas.ai. It analyses Barcelona Activa data from September 2017 to September 2022 and describes cooperation from the agency. Barcelona Activa publicly says its talent service helps employers define roles and pre-select candidates.
The research findings: aggregate gender parity coexisted with lower rates in several slices of the pipeline. The author also reports that the gender shortlisting gap narrowed from 6.5 percentage points in 2017 to 1.3 in 2022, while overall shortlisting fell for both groups.
The policy context: the European Commission lists AI used to analyse and filter job applications or evaluate candidates among the employment uses treated as high risk under the AI Act. Its current guidance says those rules apply from 2 December 2027 and require lifecycle monitoring and human oversight.
Still open: whether the patterns continued after 2022, how TalentClue's current system works, what interventions Barcelona Activa has made, and which parts of the old pipeline caused each disparity. Model Current contacted no party for this article, so there is no fresh response from the agency or vendor.
A fair-looking final number can be true. It can also be far too small an answer.
Sources
- Galdón-Clavell — Applied and Filtered paper recordPrimary research record submitted 13 August 2026. Source for authorship, scope, sample, headline findings and preprint status.
- Galdón-Clavell — full HTML manuscriptFull primary manuscript. Source for the seven-stage system description, detailed results, statistical boundaries, limitations, funding and recommendations.
- Barcelona Activa — Talent MarketplacesOfficial agency description of its role supporting employers with profile definition and candidate pre-selection. It does not address the audit findings.
- European Commission — Navigating the AI ActOfficial current guidance for high-risk employment uses, the 2 December 2027 application date and provider/deployer obligations including monitoring and human oversight.
