We are working on a variety of cases. Once an audit commission has gathered and examined the ethical issue, the original problem statement and corresponding advice will be published on this website.

We are open to new cases. Please reach out to us when you know a case for review.

Type of SIM card as a predictor variable to detect payment fraud

Summary advice

The audit commission advises against using type of SIM card as an input variable in algorithmic models that predict payment defaults and block afterpay services for specific customers. As it is likely that type of SIM card acts as a proxy-variable for sensitive demographic categories, the model would run an intolerable risk of disproportionally excluding vulnerable demographic groups from the payment service. Absent reliable data that demonstrates otherwise, the ethical risk of including the SIM card variable outweighs potential benefits. The commission advises to consider a variety of alternatives in dealing with payment defaults.

Source of case

Anonymized large multinational company with e-commerce platform.


Dowload the full report and problem statement here.

Audit commission

  • Fleur Jongepier, Assistant Professor Digital Ethics, Radboud University
  • Jan Overgoor, PhD Management Science and Engineering, Stanford University
  • Merel Noorman, Assistant Professor Law, Technology and Society, Tilburg University
  • Roel Dobbe, Assistant Professor Engineering Systems and Services, TU Delft
    Page: /

Ethical risk assessment tool

This project sheds light on the complex tension between fairness, non-discrimination and statistics in risk assessment tools. Which characteristics of citizens and customers should be considered as protected variables and proxies for protected variables? To what extent and under what circumstances should these (proxy)variables be included or excluded from algorithmic predictions?

Contested algorithm

In April 2020, the Rotterdam Court of Auditors ruled that an algorithmic risk assessment tool used by the municipality of Rotterdam was biased. The reason: the use of ‘illiteracy’ as a prediction variable to generate risk scores for welfare claimants. A certain percentage of the claimants deemed high-risk were invited for re-examination. The Court of Auditors considered the variable of ‘illiteracy’ to act as a proxy for non-Dutch origin, a protected characteristic. The risk assesment algorithm used by the municipality was ruled to be discriminatory as a consequence.

Ethical-statistical gridlock

From a statistical perspective, to qualify an algorithm as discriminatory, solely on the ground that it made use of a proxy variable, is flawed and deficient. It is statistically inevitable that most if not all variables used by an algorithmic model are to various degrees correlated with protected variables. All these variables would therefore, strictly speaking, need to be considered as proxy variables. If only variables are admitted that are proven to be statistically independent and uncorrelated to protected variables, this would mean in practice that nearly all variables in the algorithmic model would be excluded from use, rendering accurate prediction all but impossible. In this line of reasoning, the only ethical use of prediction algorithms would be to use highly inaccurate models or using no algorithms at all. At Algorithm Audit, we have coined a term for this impasse: an ethical-statistical gridlock.

Fair ML research

How to resolve this gridlock? Researchers in fair machine learning warn not to debate at what correlation rate variables act as predictor variables or not (see paper). Instead, researchers advise to examine whether including protected variables in the model hold predictive power. If so, researchers frequently advocate to include protected variables in the prediction model such that the role of proxies becomes moot. The deliberate use of protected variables might sound counter-intuitive, but it allows to check and correct the algorithmic model for actual discriminatory bias.


However, including protected variables – such as gender, race and age – in algorithmic models is not always desirable. Moreover, it is often not even practically achievable. What is to be done to escape this gridlock? Selecting legitimate variables on which fair decisions can be made needs careful inspection and deliberation. By composing independent, professional and diverse audit commissions, Algorithm Audit provides ethical reviews of the ethical-statistical gridlock in a holistic and context-sensitive way that is mindful of societal impact.


We are working with academic partners on an advice regarding the above-mentioned ethical issue that emerged at the Municipality of Rotterdam. For this project we are supported by the SIDN Fund.

Audit commission

The composition of the audit commission for this case is made publicly available once the final advice of the commission is published.

Supported by