Part of the AI Regulation News hub.
The ACPR published a discussion paper on algorithmic fairness in the financial sector and put it out for public consultation
The supervisor says out loud that fairness by ignoring sensitive variables no longer works, and that picking a fairness metric is a business decision rather than a modelling detail.
Bottom line: Not binding. This is a discussion paper issued for public consultation. The ACPR states in the paper itself that it does not express an official ACPR position.
Who this affects: Model risk officers, credit underwriting and insurance pricing teams at French banks and insurers, compliance officers, and the finance-side accountants who sign off on model governance.
Issue date: The paper is dated July 2026 and the publication page carries an update date of 1 July 2026. The page does not state a closing date for the consultation.
What changed: A prudential supervisor put a structured fairness framework, and a questionnaire, in front of the banking and insurance industry it supervises.
Analysis: The paper says the three families of group fairness metrics cannot all be satisfied at once when risk levels genuinely differ between groups. That makes the metric choice an explicit trade-off a firm has to document, not a technical default it can inherit from a vendor.
Primary sources: ACPR publication page (FR) · Discussion paper PDF (FR)
- Instrument (EN)
- Algorithmic fairness in the financial sector, discussion paper
- Authority
- Autorite de controle prudentiel et de resolution (ACPR), Banque de France
- Jurisdiction
- France
- Status
- Published for public consultation, questionnaire annexed
- Bindingness
- Non-binding. The paper states it does not express an official ACPR position
- Issue date / next deadline
- July 2026; page updated 1 July 2026. No closing date stated on the publication page
- Authors
- Cyril Chhun, Olivier Fliche, Julien Uri, Directorate for Innovation, Data and Technology Risk
- Scope
- Mainly traditional predictive systems, with a shorter forward-looking part on generative AI
- Primary source
- https://acpr.banque-france.fr/fr/publications-et-statistiques/publications/lequite-algorithmique-dans-le-secteur-financier
What the ACPR actually published
On its publications page the ACPR posted a document de reflexion titled L'equite algorithmique dans le secteur financier, together with a consultation questionnaire in DOCX form. The page carries an update date of 1 July 2026 and the paper is dated July 2026.
The paper names its own authors and its own limits. It was written by the ACPR technology risk supervision service, drawing on a scientific literature review and on technical workshops the ACPR ran with volunteer financial institutions in spring and autumn 2025. In its introduction it says plainly that it is not meant to give an exhaustive view of algorithmic fairness, nor to express an official ACPR position, and that its purpose is to develop first analyses to be discussed with stakeholders in a public consultation.
That framing matters for how you read it. Nothing in the document creates an obligation. It is the supervisor thinking out loud, in public, before it decides whether to say anything more formal.
Fairness by ignorance is described as obsolete
The core problem the paper sets up is old: a lender or insurer has to differentiate people by risk level, because that is what makes the business model sustainable, while sensitive characteristics are frequently correlated with the risks actually observed.
The long-dominant answer was what the paper calls equity by ignorance, meaning you simply drop sensitive variables from the statistical treatment. The ACPR says that approach has always been debated and is now largely made obsolete by the development of AI models, which can reconstitute the information those variables carried. High dimensionality plus strong collinearity between variables means the algorithm can find proxies.
There is a second-order point the paper does not hide. AI models are more performant, which in theory can reduce unfair treatment; but the mechanisms that produce discrimination become more diffuse and harder to identify and interpret, because the systems are complex and opaque.
How the paper handles the legal frame
Part one walks the legal ground: the French constitutional and criminal-law non-discrimination rules, article 20 and article 21 of the EU Charter of Fundamental Rights, and article 14 of the European Convention on Human Rights. The paper notes that article 225-1 of the French penal code lists 26 protected criteria.
Regulation (EU) 2024/1689, the AI Act, is presented as completing that frame: it sets fairness requirements for so-called high-risk AI systems and reaffirms a non-discrimination principle for all AI systems deployed in the Union.
The ACPR then draws a distinction worth keeping. Sectoral customer protection rules in banking and insurance also carry fairness requirements, but often through a logic of protection by abstention, meaning by not granting or not selling. The AI Act, on the ACPR's account, puts more weight on the risk of exclusion. Those two logics can pull in opposite directions on the same file.
The impossibility result, and why it is the operational point
The paper separates three levels that get confused in public debate: disparity, bias and discrimination. A disparity is not necessarily a discrimination, since discrimination involves a normative and contextual judgement.
It then sets out three families of group fairness metrics: independence, separation and sufficiency. Each corresponds to a distinct normative conception, respectively parity of outcomes, parity of errors, and parity of the reliability of decisions, and each leads to different practical implications.
The literature, the paper says, shows that where risk levels differ between groups it is impossible to satisfy these requirements simultaneously. Choosing a fairness metric therefore necessarily involves an arbitrage between competing objectives, fairness, performance and inclusion, and the ACPR says that arbitrage should be made explicit.
The paper also prefers group fairness over individual fairness in practice. Individual fairness is theoretically attractive, including on performance grounds, but the paper judges its implementation conditions particularly demanding.
Governance, not just modelling
Part five is the one a compliance reader should look at first. The paper argues fairness cannot be reduced to a technical question owned by the modelling teams. It calls it a transversal issue involving strategic choices and trade-offs that fall under the overall responsibility of the financial institution.
Concretely, it asks for explicit objectives, documented choices, and control arrangements consistent with existing model risk management frameworks, applied across strategic, business and technical decision levels and across the whole system lifecycle.
A final part turns to generative AI. The ACPR observes it is developing fast in the financial sector but is not yet deployed at scale for use cases with strong fairness stakes, and that bias evaluation methods built for classic predictive models do not transpose directly. It points to emerging evaluation approaches combining a representational layer, a behavioural layer and an allocative layer.
What we did not verify
We opened the ACPR publication page and the first eight pages of the 59-page discussion paper PDF, both in French, and took every fact and the quotation from those.
We did not open the DOCX consultation questionnaire, and we did not read parts three through six of the paper in full; the descriptions of those parts here come from the paper's own summary and table of contents. We found no English version.
We do not claim a consultation closing date, because the publication page does not state one and we did not find one in the pages we read. We do not claim the ACPR will follow this paper with guidance, a position or a recommendation, and we make no claim about how the ACPR would treat any particular metric choice in supervision.
A supervisor has now said in writing that dropping sensitive variables does not deliver fairness once AI models can rebuild them from proxies. If your institution still relies on that as its defence, the defence is weaker than it was. The transferable move is to write down which fairness metric you chose, what you traded away to choose it, and who approved that trade, because the paper's impossibility point means there is no neutral default to fall back on.
Source File
Open the ACPR publication page and confirm the update date of 1 July 2026, the PDF and DOCX download links, and the paragraph stating the AI Act sets fairness requirements for high-risk systems. Then open the PDF and confirm the introduction's sentence that the document does not express an official ACPR position, and the summary's statement that the three metric families cannot be satisfied simultaneously.
Il n'a pas vocation a donner une vision exhaustive de l'ensemble des sujets lies a l'equite algorithmique, ni a exprimer une position officielle de l'ACPR. ACPR discussion paper, July 2026
FAQ
Does this document impose any obligation on French banks or insurers?
No. It is a discussion paper published for consultation, and the paper says in its introduction that it does not express an official ACPR position. The binding requirements it discusses come from existing non-discrimination law, sectoral customer protection rules and the AI Act, not from this document.
What does the ACPR mean by equity by ignorance?
The practice of excluding sensitive variables from statistical processing. The paper says this approach has always been debated and is now largely obsolete, because AI models can reconstitute the information those variables carried by finding substitute variables among highly collinear, high-dimensional data.
Why can a firm not satisfy all three fairness metrics at once?
Because independence, separation and sufficiency encode different normative goals, and the literature the paper relies on shows they cannot be met simultaneously where risk levels genuinely differ between groups. The paper's response is that the choice has to be an explicit, documented arbitrage.
Does the paper cover generative AI?
Only in a short, forward-looking sixth part. The paper says its focus is traditional predictive systems, which are what is currently deployed at scale in the sector with fairness stakes, and that bias evaluation methods for those models do not transpose directly to generative systems.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.