From Research to Regulation: How the science of AI Fairness shaped the EU AI Act
The EU AI Act did not appear from nowhere.
It is, in significant part, the legislative crystallization of a decade of academic work on algorithmic fairness — work that began in computer science and statistics, spilled into public discourse through a handful of scandals, and ultimately reached the desks of regulators in Brussels.
Understanding that lineage matters — not just historically, but practically.
The science came first.
In 2012, Dwork et al. introduced the concept of individual fairness in “Fairness Through Awareness” (ITCS 2012) — the idea that similar individuals should be treated similarly by an algorithm. Four years later, Hardt, Price & Srebro introduced equalized odds in “Equality of Opportunity in Supervised Learning” (NeurIPS 2016), formalizing what it means for a classifier to perform equally well across demographic groups.
These were not policy documents. They were mathematical proofs published in top-tier CS venues. Researchers were developing the formal vocabulary before regulators even knew they needed it. Then came the moment that changed everything.
In May 2016, ProPublica’s “Machine Bias” investigation showed that the COMPAS recidivism algorithm exhibited racial disparities in false-positive rates. Months later, Chouldechova’s formal analysis — first circulated as a preprint in October 2016 and published in Big Data in 2017 — supplied the theoretical explanation: you cannot simultaneously satisfy all common fairness definitions when base rates differ across groups. This is a hard result. Not a political opinion. A theorem.
The EU AI Act operationalizes exactly this debate.
Article 10 mandates data governance practices to address bias — requiring that training data be examined for biases that are likely to affect health and safety or negatively impact fundamental rights. Recital 67 explicitly acknowledges that bias can be inherent in underlying data sets — particularly historical data — or emerge once a system is deployed in real-world settings, including through feedback loops. The Act’s entire risk-based architecture for high-risk AI systems (Annex III) maps, category by category, onto the harm typologies that fairness researchers had been publishing for years: credit, employment, education, and law enforcement.
The Act does not resolve the Chouldechova impossibility. No regulation can. But it does require organizations to explicitly choose their fairness criteria and document that choice — precisely what the research community has been arguing for years is the intellectually honest approach.
What this means for organizations using AI today:
Compliance with the EU AI Act is, at its core, a structured engagement in a scientific debate that has been ongoing since at least 2012. If your team lacks access to that underlying science, you are navigating regulation without a map.
The key references worth reading:
Dwork et al. (2012), “Fairness Through Awareness,” https://arxiv.org/abs/1104.3913
Hardt, Price & Srebro (2016), “Equality of Opportunity in Supervised Learning,” https://arxiv.org/abs/1610.02413
Chouldechova (2017), “Fair Prediction with Disparate Impact,” https://arxiv.org/abs/1610.07524
Barocas, Hardt & Narayanan, Fairness and Machine Learning, fairmlbook.org
Mehrabi et al. (2021), “A Survey on Bias and Fairness in Machine Learning,” https://arxiv.org/abs/1908.09635
Which of these have you read? In the age of AI, the research informing the AI Act policymaking is almost classic literature. Do you know of any younger publications I should read?
Be sure to visit us and see what we can do for you: https://waveimpact.de/en