AI systems make decisions that directly affect people's lives: who is considered for a job, a loan, or a visa. When those decisions systematically disadvantage certain groups, the result is algorithmic bias. Bias is not an unavoidable technical residual: it is a risk that organisations are legally obliged to identify and manage. The EU AI Act (Regulation (EU) 2024/1689) sets out concrete requirements to that end, and the FAST principles (Fairness, Accountability, Safety, and Transparency) offer a practical framework for meeting those obligations.

What is algorithmic bias?

Bias in AI systems occurs when a model systematically favours or disadvantages certain groups. The most common cause is skewed training data: historical data mirrors existing societal inequalities and transfers them into the model. A model trained on years of job applications from a male-dominated sector learns that male candidates are the norm. Women are disadvantaged, not because the system was designed that way, but because the data taught it to be.

A well-known example is the CV-screening system Amazon tested internally between 2014 and 2017. The system was trained on historical hiring data, the vast majority of which came from male applicants. The algorithm learned to treat female-coded signals as negative indicators: CVs containing the word "women's" were ranked lower, as were degrees from women's colleges. Amazon discontinued the project once it became clear that the bias was systematic and not easily correctable.

Bias need not be direct. Indirect discrimination occurs when a model does not use a protected characteristic such as gender or ethnicity as an input, but relies on features that strongly correlate with it, so-called proxies. A selection criterion such as "uninterrupted years of service" statistically disadvantages women and people with disabilities, who more frequently have career breaks.

The FAST principles as an operational framework

The FAST principles provide a practical framework for organisations seeking to deploy AI responsibly.

Fairness means that AI systems treat all affected groups equitably and without discrimination. This requires representative training data, systematic bias testing, and concrete mitigation measures. Fairness is not a static property: models drift over time as the world changes and the data feeding them shifts accordingly.

Accountability refers to the requirement that it is clear, for every AI system, who bears responsibility for its outcomes, how oversight is organised, and how individuals and staff can challenge a decision that is incorrect or unlawful. Accountability goes beyond technical logging: it demands a governance structure that can be demonstrated to function.

Safety concerns the technical robustness of the system: reliable performance under varying conditions, resilience against manipulation and unintended use, and consistent functioning across all relevant subgroups.

Transparency means that the operation of an AI system is understandable and traceable for those who work with it and those affected by it. Transparency is a precondition for meaningful human oversight: someone who does not understand how a system reaches an outcome cannot meaningfully assess that outcome.

Legal basis in the EU AI Act

The FAST principles are not optional guidance. They are largely codified in the EU AI Act, which entered into force on 1 August 2024.

Recital 27 recalls the seven ethical principles for trustworthy AI, including "diversity, non-discrimination and fairness": AI systems shall be developed and used in a way that promotes equal access and prevents discriminatory effects prohibited under Union law.

Article 10 imposes concrete data governance obligations on providers of high-risk AI systems. They must systematically examine training data for potential bias that could adversely affect fundamental rights or lead to prohibited discrimination, and must take appropriate measures to detect, prevent, and mitigate such bias. Article 13 requires sufficient transparency so that deployers can interpret and appropriately use the system's output.

For deployers, Article 26(4) establishes the obligation to ensure that input data is relevant and sufficiently representative for the system's intended purpose.

What goes wrong in practice?

Dutch and European practice demonstrates that bias and opacity cause real harm. The Autoriteit Persoonsgegevens reported in 2025 that AI systems in the labour market carry significant discrimination risks: CV-filtering systems systematically exclude candidates from certain population groups, while employers are typically not transparent about this.

In August 2025, the Rechtbank Den Haag found that the Minister of Foreign Affairs had used an algorithm (Informatie Ondersteund Beslissen) to reject a visa application without being transparent about it. Neither the decision letter nor the Algoritmeregister made clear which criteria had been applied. The court ruled that the minister was required to provide more information about the selection process (ECLI:NL:RBDHA:2025:14544). The ruling illustrates that a lack of transparency about algorithmic decision-making can be incompatible with the principles of good governance.

The role of the Algoritmeregister

The Dutch government established the Algoritmeregister as a transparency instrument: public-sector organisations publish information about the algorithms they deploy. Registration of high-impact algorithms will in future become a statutory obligation. The visa case demonstrates, however, that presence in the register does not automatically mean the information is sufficiently clear and accessible. Transparency requires not just publication, but also comprehensibility and up-to-date content.

Practical steps for organisations

Organisations that deploy or develop AI systems can address bias and opacity in a systematic manner through a number of measures:

  • Conduct a bias audit on training data before a system is deployed, and repeat that analysis periodically.
  • Document the composition of training data and the design choices made, in accordance with Article 10 of the EU AI Act.
  • Establish a procedure for objection and correction, so that individuals disadvantaged by an AI decision can challenge it.
  • Communicate actively and in plain language to those affected when and how AI has been used in a decision.
  • Ensure meaningful human oversight: a staff member who does not understand a system's output cannot effectively evaluate it.

Ethical AI is not a philosophical ideal but a demonstrable operational practice, underpinned by a legal framework that makes compliance enforceable.