Algorithmic bias is a systematic distortion in the results of an automated system that favours or penalises certain groups of candidates. This page defines the phenomenon, looks back at documented cases such as the tool Amazon abandoned in 2018, summarises what the AI Act requires, explains how Seeqle's Match Agent limits these biases, and states clearly what no system can guarantee.
A recruitment AI can be GDPR compliant, but no system can guarantee the complete absence of algorithmic bias. Compliance rests on a clear rule: Article 22 of the GDPR in principle prohibits fully automated decisions on an application, so the tool prepares the decision and a human makes it. Against bias, Seeqle acts on three levers: distribution targeting that excludes discriminatory criteria, a Match Agent that scores on declared and predicted skills rather than on degree or previous employer, and an explained score that the recruiter can challenge. Human oversight and regular audits remain necessary.
It is a systematic, not random, error in the results of an automated screening or targeting system: with equal skills, some profiles are ranked higher or more exposed to job offers than others, for reasons unrelated to suitability for the role.
Bias can affect both stages of AI-assisted recruitment:
Algorithmic bias requires no discriminatory intent: it emerges from data and optimisation objectives. That is precisely what makes it hard to see without measurement.
Mainly from training data: a model trained on past hires learns past preferences, including the bad ones. Add to that correlated variables (proxies), which reintroduce a protected characteristic without naming it, and feedback loops that amplify an initial imbalance.
The three documented mechanisms:
The best-documented case is Amazon's: in 2018, Reuters revealed that its experimental CV rating tool, trained on ten years of mostly male applications, penalised CVs containing the word “women's” and downgraded graduates of certain women's colleges. Amazon abandoned the tool.
What this case establishes, according to the Reuters investigation (Jeffrey Dastin, October 2018):
Other academic and journalistic work has documented distortions in how job ads are delivered on advertising platforms. The common lesson: bias is measured on results, it cannot be guessed by reading the code.
The AI Act (Regulation (EU) 2024/1689) classifies recruitment AI systems as high risk (Annex III, point 4) and requires their providers to examine training data for possible bias, take measures to detect and correct it, and ensure effective human oversight.
The requirements most directly related to bias:
Timeline: these obligations for Annex III high-risk systems will apply on 2 December 2027, after the postponement decided by the Digital Omnibus, which entered into force in July 2026. The GDPR, which already prohibits in principle fully automated decisions on an application (Article 22), applies right now.
Sources: Regulation (EU) 2024/1689, Articles 10, 13, 14 and Annex III (EUR-Lex, 2024) · European Commission, AI Act implementation timeline (ai-act-service-desk.ec.europa.eu).
Through three architecture choices: scoring relies on skills criteria and role prerequisites, not on personal characteristics; the score is explainable, broken down criterion by criterion; and no decision is automatic, the recruiter always decides.
In detail:
No system, Seeqle included, eliminates all bias. Residual correlations can persist in any data processing, and a fair score can be used unfairly. Regular human auditing of decisions remains necessary, and saying so is part of using AI honestly.
What a tool cannot guarantee:
That is why Seeqle's position is this: the tool reduces the opportunities for bias (objective criteria, explainable score, human decision), and the employer remains responsible for monitoring the results. The Amazon case showed it: it is stopping measurement, not the tool alone, that creates the risk.
By measuring results at every stage, not by rereading the rules. Compare pass-through rates between groups when your data legally allows it, sample rejected applications for human review, and reread job description prerequisites for every campaign.
A practical checklist:
A systematic distortion in an automated system that favours or penalises certain groups of candidates for reasons unrelated to skills, most often inherited from training data or correlated variables.
Yes. In October 2018, Reuters documented that an experimental Amazon tool, trained on mostly male applications, penalised CVs mentioning the word “women's”. Amazon abandoned the project.
The question cannot be settled in the abstract. An algorithm can industrialise a bias, but it can also be measured, audited and corrected, which human judgement does not easily allow. The deciding factor is measuring results and keeping the decision human.
The Match Agent scores on role-related skills criteria, displays a score that is explainable criterion by criterion and never rejects a candidate: the decision rests with the recruiter, who can depart from the score.
No, and a promise of zero bias should raise a red flag. A serious tool reduces the opportunities for bias and makes its results auditable; regular human oversight of decisions remains essential, which the AI Act builds into its requirements.
Sources cited: Reuters, “Amazon scraps secret AI recruiting tool that showed bias against women”, Jeffrey Dastin, 10 October 2018 (reuters.com) · Regulation (EU) 2024/1689 (AI Act), Articles 10, 13, 14 and Annex III point 4 (eur-lex.europa.eu, 2024) · European Commission, AI Act implementation timeline (ai-act-service-desk.ec.europa.eu) · Regulation (EU) 2016/679 (GDPR), Article 22 (eur-lex.europa.eu, 2016).
A score explainable criterion by criterion, with the decision kept human: a 15-minute demo shows how the Match Agent makes screening auditable.
Go further: Talent Attraction Diagnostic · Attract Agent · Match Agent · Pricing · Where your job offers are distributed · HR Glossary