You can't be fair three ways — and the EU AI Act clock is ticking
A landmark result proves demographic parity, equalized odds, and calibration can't all hold when base rates differ. With high-risk obligations applying from August 2026, fairness is now an engineering pipeline, not a footnote.
A loan model can be 92% accurate overall and still approve one group at twice the rate of another with identical qualifications. “Accurate” is not “fair,” and in 2026 that gap is both a mathematical certainty and a legal liability. Two things every ML engineer should internalize: you provably can’t satisfy all the fairness definitions at once, and the regulatory clock is now running.
The impossibility is a theorem, not an opinion
You might hope to just optimize a single “fairness” number. You can’t. Kleinberg, Mullainathan, and Raghavan (2016) formalize three fairness conditions and prove that, except in highly constrained special cases, no method can satisfy all three simultaneously. Chouldechova (2017) shows the same tension concretely: calibration and error-rate balance cannot both hold when prevalence (base rates) differs across groups, producing disparate impact. The practical consequence is liberating once you accept it: fairness is a decision. A hiring tool might prioritize equal selection rates; a medical model might prioritize equal error rates. You choose which fairness your context demands — ideally with stakeholders, before you optimize — rather than chasing a number that can’t exist.
The EU AI Act made it an engineering requirement
Fairness stopped being optional. The EU AI Act’s implementation timeline sets 2 August 2026 as the date the high-risk system obligations start to apply, and providers of high-risk systems must establish lifecycle risk management, ensure representative and error-minimized data governance, prepare technical documentation, and enable human oversight. The shift for MLOps teams is from “write a governance report” to “the pipeline emits the evidence.” Notably, model cards substantially overlap with the Act’s required technical documentation and help satisfy the Article 11/Annex IV and Article 13 transparency obligations — so an auto-updated model card becomes living compliance, not a stale document. And the mitigation tooling is mature: Fairlearn is an open-source package that assesses group fairness and provides mitigation algorithms.
By the 2026 deadline, high-risk providers are expected to complete conformity assessments, finalize technical documentation, affix CE marking, register in the EU database, and maintain a quality management system. The teams that do well are treating all of it as structured metadata their pipeline emits — not a scramble before an audit.
Related lessons
- Fairness & bias in ML — the metrics and the impossibility result, with an interactive threshold demo
- Responsible-AI ops — operationalizing fairness and model cards
- Model registry & promotion — where the compliance gate lives